diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aic/exp_gate_up_matmul.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aic/exp_gate_up_matmul.cpp
new file mode 100644
index 000000000..93eaef9ad
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aic/exp_gate_up_matmul.cpp
@@ -0,0 +1,402 @@
+// Kernel Function: exp_gate_up_matmul
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_gate_up_matmul(__gm__ int8_t* v1, __gm__ int8_t* v2, __gm__ int8_t* v3, __gm__ int32_t* v4, __gm__ int32_t* v5, int32_t v6, int32_t v7) {
+  unsigned v8 = 0;
+  const int32_t v9 = 128;
+  const int32_t v10 = 0;
+  const int32_t v11 = 512;
+  const int32_t v12 = 16777216;
+  const int32_t v13 = 256;
+  const int32_t v14 = 1;
+  const int32_t v15 = 4096;
+  const int32_t v16 = 16;
+  const int64_t v17 = 32768;
+  const int64_t v18 = 2048;
+  const int64_t v19 = 16384;
+  const int64_t v20 = 8192;
+  const int64_t v21 = 0;
+  using T = float;
+
+  #if defined(__DAV_CUBE__)
+  size_t v22 = (size_t) v13;
+  size_t v23 = (size_t) v11;
+  size_t v24 = (size_t) v10;
+  Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v11);
+  uint64_t v26 = (uint64_t) v21;
+  TASSIGN(v25, v26);
+  pto::Shape<1, 1, 1, 16, 512> v27 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v28 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v29 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v8 + v8 * (unsigned) v15 + v8 * (unsigned) v14), v27, v28);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID2);
+  TLOAD(v25, v29);
+  Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v11);
+  uint64_t v31 = (uint64_t) v21;
+  TASSIGN(v30, v31);
+  pipe_barrier(PIPE_MTE2);
+  TLOAD(v30, v29);
+  Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v32 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v33 = (uint64_t) v20;
+  TASSIGN(v32, v33);
+  pto::Shape<1, 1, 1, 512, 256> v34 = pto::Shape<1, 1, 1, 512, 256>();
+  pto::Stride<16777216, 16777216, 16777216, 1, 4096> v35 = pto::Stride<16777216, 16777216, 16777216, 1, 4096>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN> v36 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN>(v2 + ((v8 + (unsigned) v6 * (unsigned) v12) + v8 * (unsigned) v14 + (unsigned) v7 * (unsigned) v15), v34, v35);
+  TLOAD(v32, v36);
+  set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+  uint64_t v38 = (uint64_t) v19;
+  TASSIGN(v37, v38);
+  wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  for (size_t v39 = v24; v39 < v23; v39 += v22) {
+    int32_t v40 = (int32_t) v39;
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v42 = (uint64_t) v18;
+    TASSIGN(v41, v42);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    pipe_barrier(PIPE_MTE1);
+    TEXTRACT(v41, v30, v10, v39);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+    uint64_t v44 = (uint64_t) v21;
+    TASSIGN(v43, v44);
+    TEXTRACT(v43, v32, v39, v10);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v46 = (uint64_t) v21;
+    TASSIGN(v45, v46);
+    int32_t v47 = (int32_t) ((uint32_t) v40 + (uint32_t) v9);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    TEXTRACT(v45, v30, v10, v47);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v48 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+    uint64_t v49 = (uint64_t) v17;
+    TASSIGN(v48, v49);
+    TEXTRACT(v48, v32, v47, v10);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    if (v40 == v10) {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v51 = (uint64_t) v19;
+      TASSIGN(v50, v51);
+      pipe_barrier(PIPE_M);
+      TMATMUL(v50, v41, v43);
+    } else {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v53 = (uint64_t) v19;
+      TASSIGN(v52, v53);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v52, v52, v41, v43);
+    };
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+    uint64_t v55 = (uint64_t) v19;
+    TASSIGN(v54, v55);
+    pipe_barrier(PIPE_M);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    TMATMUL_ACC(v54, v54, v45, v48);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  }
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v11);
+  uint64_t v57 = (uint64_t) v21;
+  TASSIGN(v56, v57);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  pipe_barrier(PIPE_MTE2);
+  TLOAD(v56, v29);
+  Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v59 = (uint64_t) v20;
+  TASSIGN(v58, v59);
+  pto::Shape<1, 1, 1, 512, 256> v60 = pto::Shape<1, 1, 1, 512, 256>();
+  pto::Stride<16777216, 16777216, 16777216, 1, 4096> v61 = pto::Stride<16777216, 16777216, 16777216, 1, 4096>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN> v62 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN>(v3 + ((v8 + (unsigned) v6 * (unsigned) v12) + v8 * (unsigned) v14 + (unsigned) v7 * (unsigned) v15), v60, v61);
+  TLOAD(v58, v62);
+  set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+  Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v63 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+  uint64_t v64 = (uint64_t) v21;
+  TASSIGN(v63, v64);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  for (size_t v65 = v24; v65 < v23; v65 += v22) {
+    int32_t v66 = (int32_t) v65;
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v67 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v68 = (uint64_t) v18;
+    TASSIGN(v67, v68);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+    pipe_barrier(PIPE_MTE1);
+    TEXTRACT(v67, v56, v10, v65);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v69 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+    uint64_t v70 = (uint64_t) v21;
+    TASSIGN(v69, v70);
+    TEXTRACT(v69, v58, v65, v10);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v71 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v72 = (uint64_t) v21;
+    TASSIGN(v71, v72);
+    int32_t v73 = (int32_t) ((uint32_t) v66 + (uint32_t) v9);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+    TEXTRACT(v71, v56, v10, v73);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v74 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+    uint64_t v75 = (uint64_t) v17;
+    TASSIGN(v74, v75);
+    TEXTRACT(v74, v58, v73, v10);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+    if (v66 == v10) {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v76 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v77 = (uint64_t) v21;
+      TASSIGN(v76, v77);
+      pipe_barrier(PIPE_M);
+      TMATMUL(v76, v67, v69);
+    } else {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v78 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v79 = (uint64_t) v21;
+      TASSIGN(v78, v79);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v78, v78, v67, v69);
+    };
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+    Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v80 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+    uint64_t v81 = (uint64_t) v21;
+    TASSIGN(v80, v81);
+    pipe_barrier(PIPE_M);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+    TMATMUL_ACC(v80, v80, v71, v74);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  }
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID5);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID5);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+  for (size_t v82 = v23; v82 < ((size_t) v15); v82 += v23) {
+    int32_t v83 = (int32_t) v82;
+    Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v84 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v11);
+    uint64_t v85 = (uint64_t) v21;
+    TASSIGN(v84, v85);
+    pto::Shape<1, 1, 1, 16, 512> v86 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v87 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v88 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v8 + v8 * (unsigned) v15 + (unsigned) v83 * (unsigned) v14), v86, v87);
+    pipe_barrier(PIPE_MTE2);
+    wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID2);
+    TLOAD(v84, v88);
+    Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v89 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v11);
+    uint64_t v90 = (uint64_t) v21;
+    TASSIGN(v89, v90);
+    pipe_barrier(PIPE_MTE2);
+    TLOAD(v89, v88);
+    Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v91 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v92 = (uint64_t) v20;
+    TASSIGN(v91, v92);
+    pto::Shape<1, 1, 1, 512, 256> v93 = pto::Shape<1, 1, 1, 512, 256>();
+    pto::Stride<16777216, 16777216, 16777216, 1, 4096> v94 = pto::Stride<16777216, 16777216, 16777216, 1, 4096>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN> v95 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN>(v2 + ((v8 + (unsigned) v6 * (unsigned) v12) + (unsigned) v83 * (unsigned) v14 + (unsigned) v7 * (unsigned) v15), v93, v94);
+    TLOAD(v91, v95);
+    set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID2);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+    wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID2);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+    for (size_t v96 = v24; v96 < v23; v96 += v22) {
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v97 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+      uint64_t v98 = (uint64_t) v18;
+      TASSIGN(v97, v98);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+      TEXTRACT(v97, v89, v10, v96);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v99 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+      uint64_t v100 = (uint64_t) v21;
+      TASSIGN(v99, v100);
+      TEXTRACT(v99, v91, v96, v10);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID4);
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v101 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+      uint64_t v102 = (uint64_t) v21;
+      TASSIGN(v101, v102);
+      int32_t v103 = (int32_t) ((uint32_t) ((int32_t) v96) + (uint32_t) v9);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+      TEXTRACT(v101, v89, v10, v103);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v104 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+      uint64_t v105 = (uint64_t) v17;
+      TASSIGN(v104, v105);
+      TEXTRACT(v104, v91, v103, v10);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID5);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v106 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v107 = (uint64_t) v19;
+      TASSIGN(v106, v107);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID4);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v106, v106, v97, v99);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v108 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v109 = (uint64_t) v19;
+      TASSIGN(v108, v109);
+      pipe_barrier(PIPE_M);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID5);
+      TMATMUL_ACC(v108, v108, v101, v104);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    };
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID3);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v110 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v11);
+    uint64_t v111 = (uint64_t) v21;
+    TASSIGN(v110, v111);
+    wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID3);
+    pipe_barrier(PIPE_MTE2);
+    TLOAD(v110, v88);
+    Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v112 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v113 = (uint64_t) v20;
+    TASSIGN(v112, v113);
+    pto::Shape<1, 1, 1, 512, 256> v114 = pto::Shape<1, 1, 1, 512, 256>();
+    pto::Stride<16777216, 16777216, 16777216, 1, 4096> v115 = pto::Stride<16777216, 16777216, 16777216, 1, 4096>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN> v116 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN>(v3 + ((v8 + (unsigned) v6 * (unsigned) v12) + (unsigned) v83 * (unsigned) v14 + (unsigned) v7 * (unsigned) v15), v114, v115);
+    TLOAD(v112, v116);
+    set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID3);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID3);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    for (size_t v117 = v24; v117 < v23; v117 += v22) {
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v118 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+      uint64_t v119 = (uint64_t) v18;
+      TASSIGN(v118, v119);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+      TEXTRACT(v118, v110, v10, v117);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v120 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+      uint64_t v121 = (uint64_t) v21;
+      TASSIGN(v120, v121);
+      TEXTRACT(v120, v112, v117, v10);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID6);
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v122 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+      uint64_t v123 = (uint64_t) v21;
+      TASSIGN(v122, v123);
+      int32_t v124 = (int32_t) ((uint32_t) ((int32_t) v117) + (uint32_t) v9);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+      TEXTRACT(v122, v110, v10, v124);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v125 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v13);
+      uint64_t v126 = (uint64_t) v17;
+      TASSIGN(v125, v126);
+      TEXTRACT(v125, v112, v124, v10);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID7);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v127 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v128 = (uint64_t) v21;
+      TASSIGN(v127, v128);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID6);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v127, v127, v118, v120);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v129 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v16, v13);
+      uint64_t v130 = (uint64_t) v21;
+      TASSIGN(v129, v130);
+      pipe_barrier(PIPE_M);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID7);
+      TMATMUL_ACC(v129, v129, v122, v125);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    };
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID2);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+  }
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+  set_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 256> v131 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v132 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v133 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v4 + ((v8 + v8 * (unsigned) v15) + v8 * (unsigned) v13 + v8 * (unsigned) v14), v131, v132);
+  wait_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  TSTORE(v133, v37);
+  pto::Shape<1, 1, 1, 16, 256> v134 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v135 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v136 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v5 + ((v8 + v8 * (unsigned) v15) + v8 * (unsigned) v13 + v8 * (unsigned) v14), v134, v135);
+  TSTORE(v136, v63);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID2);
+  #endif // __DAV_CUBE__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: recv_x_tile_i8_inline75__rv_v2
+    __gm__ Tensor* recv_x_tile_i8_inline75__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int8_t* recv_x_tile_i8_inline75__rv_v2 = reinterpret_cast<__gm__ int8_t*>(recv_x_tile_i8_inline75__rv_v2_tensor->buffer.addr) + recv_x_tile_i8_inline75__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: expert_w1__ssa_v0
+    __gm__ Tensor* expert_w1__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* expert_w1__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(expert_w1__ssa_v0_tensor->buffer.addr) + expert_w1__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: expert_w3__ssa_v0
+    __gm__ Tensor* expert_w3__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ int8_t* expert_w3__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(expert_w3__ssa_v0_tensor->buffer.addr) + expert_w3__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ int32_t* ret0__out = reinterpret_cast<__gm__ int32_t*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack tensor: ret1__out
+    __gm__ Tensor* ret1__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ int32_t* ret1__out = reinterpret_cast<__gm__ int32_t*>(ret1__out_tensor->buffer.addr) + ret1__out_tensor->start_offset;
+
+    // Unpack scalar: local_i_inline67__idx_v0
+    union { uint64_t u64; int64_t val; } local_i_inline67__idx_v0_conv;
+    local_i_inline67__idx_v0_conv.u64 = args[5];
+    int64_t local_i_inline67__idx_v0 = local_i_inline67__idx_v0_conv.val;
+
+    // Unpack scalar: n0_inline72__idx_v0
+    union { uint64_t u64; int64_t val; } n0_inline72__idx_v0_conv;
+    n0_inline72__idx_v0_conv.u64 = args[6];
+    int64_t n0_inline72__idx_v0 = n0_inline72__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_gate_up_matmul(recv_x_tile_i8_inline75__rv_v2, expert_w1__ssa_v0, expert_w3__ssa_v0, ret0__out, ret1__out, local_i_inline67__idx_v0, n0_inline72__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aic/exp_w2_matmul.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aic/exp_w2_matmul.cpp
new file mode 100644
index 000000000..a62a84424
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aic/exp_w2_matmul.cpp
@@ -0,0 +1,245 @@
+// Kernel Function: exp_w2_matmul
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_w2_matmul(__gm__ int8_t* v1, __gm__ int8_t* v2, __gm__ int32_t* v3, int32_t v4, int32_t v5) {
+  unsigned v6 = 0;
+  const int32_t v7 = 64;
+  const int32_t v8 = 128;
+  const int32_t v9 = 0;
+  const int32_t v10 = 16777216;
+  const int32_t v11 = 512;
+  const int32_t v12 = 1;
+  const int32_t v13 = 4096;
+  const int32_t v14 = 16;
+  const int64_t v15 = 32768;
+  const int64_t v16 = 1024;
+  const int64_t v17 = 8192;
+  const int64_t v18 = 0;
+  const int32_t v19 = 8192;
+  using T = float;
+
+  #if defined(__DAV_CUBE__)
+  size_t v20 = (size_t) v11;
+  size_t v21 = (size_t) v9;
+  size_t v22 = (size_t) v8;
+  Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v23 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+  uint64_t v24 = (uint64_t) v18;
+  TASSIGN(v23, v24);
+  pto::Shape<1, 1, 1, 16, 512> v25 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v26 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v27 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v13 + v6 * (unsigned) v12), v25, v26);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  TLOAD(v23, v27);
+  Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v11, v11);
+  uint64_t v29 = (uint64_t) v17;
+  TASSIGN(v28, v29);
+  pto::Shape<1, 1, 1, 512, 512> v30 = pto::Shape<1, 1, 1, 512, 512>();
+  pto::Stride<16777216, 16777216, 16777216, 1, 4096> v31 = pto::Stride<16777216, 16777216, 16777216, 1, 4096>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN> v32 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN>(v2 + ((v6 + (unsigned) v4 * (unsigned) v10) + v6 * (unsigned) v12 + (unsigned) v5 * (unsigned) v13), v30, v31);
+  TLOAD(v28, v32);
+  set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+  uint64_t v34 = (uint64_t) v18;
+  TASSIGN(v33, v34);
+  wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  for (size_t v35 = v21; v35 < v20; v35 += v22) {
+    int32_t v36 = (int32_t) v35;
+    Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v7);
+    uint64_t v38 = (uint64_t) v18;
+    TASSIGN(v37, v38);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    pipe_barrier(PIPE_MTE1);
+    TEXTRACT(v37, v23, v9, v35);
+    Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v7, v11);
+    uint64_t v40 = (uint64_t) v18;
+    TASSIGN(v39, v40);
+    TEXTRACT(v39, v28, v35, v9);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v7);
+    uint64_t v42 = (uint64_t) v16;
+    TASSIGN(v41, v42);
+    int32_t v43 = (int32_t) ((uint32_t) v36 + (uint32_t) v7);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    TEXTRACT(v41, v23, v9, v43);
+    Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v44 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v7, v11);
+    uint64_t v45 = (uint64_t) v15;
+    TASSIGN(v44, v45);
+    TEXTRACT(v44, v28, v43, v9);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    if (v36 == v9) {
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v46 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v47 = (uint64_t) v18;
+      TASSIGN(v46, v47);
+      pipe_barrier(PIPE_M);
+      TMATMUL(v46, v37, v39);
+    } else {
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v48 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v49 = (uint64_t) v18;
+      TASSIGN(v48, v49);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v48, v48, v37, v39);
+    };
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v51 = (uint64_t) v18;
+    TASSIGN(v50, v51);
+    pipe_barrier(PIPE_M);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    TMATMUL_ACC(v50, v50, v41, v44);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  }
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  for (size_t v52 = v20; v52 < ((size_t) v13); v52 += v20) {
+    int32_t v53 = (int32_t) v52;
+    Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v55 = (uint64_t) v18;
+    TASSIGN(v54, v55);
+    pto::Shape<1, 1, 1, 16, 512> v56 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v57 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v58 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v13 + (unsigned) v53 * (unsigned) v12), v56, v57);
+    pipe_barrier(PIPE_MTE2);
+    wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+    TLOAD(v54, v58);
+    Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v59 = Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v11, v11);
+    uint64_t v60 = (uint64_t) v17;
+    TASSIGN(v59, v60);
+    pto::Shape<1, 1, 1, 512, 512> v61 = pto::Shape<1, 1, 1, 512, 512>();
+    pto::Stride<16777216, 16777216, 16777216, 1, 4096> v62 = pto::Stride<16777216, 16777216, 16777216, 1, 4096>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN> v63 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<16777216, 16777216, 16777216, 1, 4096>, pto::Layout::DN>(v2 + ((v6 + (unsigned) v4 * (unsigned) v10) + (unsigned) v53 * (unsigned) v12 + (unsigned) v5 * (unsigned) v13), v61, v62);
+    TLOAD(v59, v63);
+    set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+    wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+    for (size_t v64 = v21; v64 < v20; v64 += v22) {
+      Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v65 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v7);
+      uint64_t v66 = (uint64_t) v18;
+      TASSIGN(v65, v66);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+      TEXTRACT(v65, v54, v9, v64);
+      Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v67 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v7, v11);
+      uint64_t v68 = (uint64_t) v18;
+      TASSIGN(v67, v68);
+      TEXTRACT(v67, v59, v64, v9);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+      Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v69 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v7);
+      uint64_t v70 = (uint64_t) v16;
+      TASSIGN(v69, v70);
+      int32_t v71 = (int32_t) ((uint32_t) ((int32_t) v64) + (uint32_t) v7);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+      TEXTRACT(v69, v54, v9, v71);
+      Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v72 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v7, v11);
+      uint64_t v73 = (uint64_t) v15;
+      TASSIGN(v72, v73);
+      TEXTRACT(v72, v59, v71, v9);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v74 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v75 = (uint64_t) v18;
+      TASSIGN(v74, v75);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v74, v74, v65, v67);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v76 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v77 = (uint64_t) v18;
+      TASSIGN(v76, v77);
+      pipe_barrier(PIPE_M);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+      TMATMUL_ACC(v76, v76, v69, v72);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+    };
+    set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  }
+  set_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v78 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v79 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v80 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v3 + ((v6 + v6 * (unsigned) v19) + v6 * (unsigned) v11 + v6 * (unsigned) v12), v78, v79);
+  wait_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  TSTORE(v80, v33);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  #endif // __DAV_CUBE__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: h_tile_i8_inline92__rv_v2
+    __gm__ Tensor* h_tile_i8_inline92__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int8_t* h_tile_i8_inline92__rv_v2 = reinterpret_cast<__gm__ int8_t*>(h_tile_i8_inline92__rv_v2_tensor->buffer.addr) + h_tile_i8_inline92__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: expert_w2__ssa_v0
+    __gm__ Tensor* expert_w2__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* expert_w2__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(expert_w2__ssa_v0_tensor->buffer.addr) + expert_w2__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ int32_t* ret0__out = reinterpret_cast<__gm__ int32_t*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: local_i_inline67__idx_v0
+    union { uint64_t u64; int64_t val; } local_i_inline67__idx_v0_conv;
+    local_i_inline67__idx_v0_conv.u64 = args[3];
+    int64_t local_i_inline67__idx_v0 = local_i_inline67__idx_v0_conv.val;
+
+    // Unpack scalar: d0_inline49__idx_v0
+    union { uint64_t u64; int64_t val; } d0_inline49__idx_v0_conv;
+    d0_inline49__idx_v0_conv.u64 = args[4];
+    int64_t d0_inline49__idx_v0 = d0_inline49__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_w2_matmul(h_tile_i8_inline92__rv_v2, expert_w2__ssa_v0, ret0__out, local_i_inline67__idx_v0, d0_inline49__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aic/sh_gate_up_matmul.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aic/sh_gate_up_matmul.cpp
new file mode 100644
index 000000000..5bbc279d0
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aic/sh_gate_up_matmul.cpp
@@ -0,0 +1,373 @@
+// Kernel Function: sh_gate_up_matmul
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_gate_up_matmul(__gm__ int8_t* v1, __gm__ int8_t* v2, __gm__ int8_t* v3, __gm__ int32_t* v4, __gm__ int32_t* v5, int32_t v6) {
+  unsigned v7 = 0;
+  const int32_t v8 = 128;
+  const int32_t v9 = 0;
+  const int32_t v10 = 512;
+  const int32_t v11 = 256;
+  const int32_t v12 = 1;
+  const int32_t v13 = 4096;
+  const int32_t v14 = 16;
+  const int64_t v15 = 32768;
+  const int64_t v16 = 2048;
+  const int64_t v17 = 16384;
+  const int64_t v18 = 139264;
+  const int64_t v19 = 8192;
+  const int64_t v20 = 0;
+  using T = float;
+
+  #if defined(__DAV_CUBE__)
+  size_t v21 = (size_t) v11;
+  size_t v22 = (size_t) v10;
+  size_t v23 = (size_t) v9;
+  Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v24 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v10);
+  uint64_t v25 = (uint64_t) v20;
+  TASSIGN(v24, v25);
+  pto::Shape<1, 1, 1, 16, 512> v26 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v27 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v28 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v7 + v7 * (unsigned) v13 + v7 * (unsigned) v12), v26, v27);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  TLOAD(v24, v28);
+  Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v29 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v10, v11);
+  uint64_t v30 = (uint64_t) v19;
+  TASSIGN(v29, v30);
+  pto::Shape<1, 1, 1, 512, 256> v31 = pto::Shape<1, 1, 1, 512, 256>();
+  pto::Stride<512, 512, 512, 1, 4096> v32 = pto::Stride<512, 512, 512, 1, 4096>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN> v33 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN>(v2 + (v7 + v7 * (unsigned) v12 + (unsigned) v6 * (unsigned) v13), v31, v32);
+  TLOAD(v29, v33);
+  set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v10, v11);
+  uint64_t v35 = (uint64_t) v18;
+  TASSIGN(v34, v35);
+  pto::Shape<1, 1, 1, 512, 256> v36 = pto::Shape<1, 1, 1, 512, 256>();
+  pto::Stride<512, 512, 512, 1, 4096> v37 = pto::Stride<512, 512, 512, 1, 4096>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN> v38 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN>(v3 + (v7 + v7 * (unsigned) v12 + (unsigned) v6 * (unsigned) v13), v36, v37);
+  TLOAD(v34, v38);
+  set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+  Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+  uint64_t v40 = (uint64_t) v17;
+  TASSIGN(v39, v40);
+  wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  for (size_t v41 = v23; v41 < v22; v41 += v21) {
+    int32_t v42 = (int32_t) v41;
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+    uint64_t v44 = (uint64_t) v16;
+    TASSIGN(v43, v44);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    pipe_barrier(PIPE_MTE1);
+    TEXTRACT(v43, v24, v9, v41);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+    uint64_t v46 = (uint64_t) v20;
+    TASSIGN(v45, v46);
+    TEXTRACT(v45, v29, v41, v9);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+    uint64_t v48 = (uint64_t) v20;
+    TASSIGN(v47, v48);
+    int32_t v49 = (int32_t) ((uint32_t) v42 + (uint32_t) v8);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    TEXTRACT(v47, v24, v9, v49);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+    uint64_t v51 = (uint64_t) v15;
+    TASSIGN(v50, v51);
+    TEXTRACT(v50, v29, v49, v9);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    if (v42 == v9) {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v53 = (uint64_t) v17;
+      TASSIGN(v52, v53);
+      pipe_barrier(PIPE_M);
+      TMATMUL(v52, v43, v45);
+    } else {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v55 = (uint64_t) v17;
+      TASSIGN(v54, v55);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v54, v54, v43, v45);
+    };
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v57 = (uint64_t) v17;
+    TASSIGN(v56, v57);
+    pipe_barrier(PIPE_M);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    TMATMUL_ACC(v56, v56, v47, v50);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  }
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+  uint64_t v59 = (uint64_t) v20;
+  TASSIGN(v58, v59);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  for (size_t v60 = v23; v60 < v22; v60 += v21) {
+    int32_t v61 = (int32_t) v60;
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+    uint64_t v63 = (uint64_t) v16;
+    TASSIGN(v62, v63);
+    pipe_barrier(PIPE_MTE1);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+    TEXTRACT(v62, v24, v9, v60);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+    uint64_t v65 = (uint64_t) v20;
+    TASSIGN(v64, v65);
+    TEXTRACT(v64, v34, v60, v9);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+    Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+    uint64_t v67 = (uint64_t) v20;
+    TASSIGN(v66, v67);
+    int32_t v68 = (int32_t) ((uint32_t) v61 + (uint32_t) v8);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+    TEXTRACT(v66, v24, v9, v68);
+    Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v69 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+    uint64_t v70 = (uint64_t) v15;
+    TASSIGN(v69, v70);
+    TEXTRACT(v69, v34, v68, v9);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+    if (v61 == v9) {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v71 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v72 = (uint64_t) v20;
+      TASSIGN(v71, v72);
+      pipe_barrier(PIPE_M);
+      TMATMUL(v71, v62, v64);
+    } else {
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v73 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v74 = (uint64_t) v20;
+      TASSIGN(v73, v74);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v73, v73, v62, v64);
+    };
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+    Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v75 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v76 = (uint64_t) v20;
+    TASSIGN(v75, v76);
+    pipe_barrier(PIPE_M);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+    TMATMUL_ACC(v75, v75, v66, v69);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  }
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID5);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID5);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+  for (size_t v77 = v22; v77 < ((size_t) v13); v77 += v22) {
+    int32_t v78 = (int32_t) v77;
+    Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v79 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v10);
+    uint64_t v80 = (uint64_t) v20;
+    TASSIGN(v79, v80);
+    pto::Shape<1, 1, 1, 16, 512> v81 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v82 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v83 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v7 + v7 * (unsigned) v13 + (unsigned) v78 * (unsigned) v12), v81, v82);
+    pipe_barrier(PIPE_MTE2);
+    wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+    TLOAD(v79, v83);
+    Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v84 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v10, v11);
+    uint64_t v85 = (uint64_t) v19;
+    TASSIGN(v84, v85);
+    pto::Shape<1, 1, 1, 512, 256> v86 = pto::Shape<1, 1, 1, 512, 256>();
+    pto::Stride<512, 512, 512, 1, 4096> v87 = pto::Stride<512, 512, 512, 1, 4096>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN> v88 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN>(v2 + (v7 + (unsigned) v78 * (unsigned) v12 + (unsigned) v6 * (unsigned) v13), v86, v87);
+    TLOAD(v84, v88);
+    set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID2);
+    Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v89 = Tile<TileType::Mat, int8_t, 512, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v10, v11);
+    uint64_t v90 = (uint64_t) v18;
+    TASSIGN(v89, v90);
+    pto::Shape<1, 1, 1, 512, 256> v91 = pto::Shape<1, 1, 1, 512, 256>();
+    pto::Stride<512, 512, 512, 1, 4096> v92 = pto::Stride<512, 512, 512, 1, 4096>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN> v93 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 256>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN>(v3 + (v7 + (unsigned) v78 * (unsigned) v12 + (unsigned) v6 * (unsigned) v13), v91, v92);
+    TLOAD(v89, v93);
+    set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID3);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+    wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID2);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+    for (size_t v94 = v23; v94 < v22; v94 += v21) {
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v95 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+      uint64_t v96 = (uint64_t) v16;
+      TASSIGN(v95, v96);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+      TEXTRACT(v95, v79, v9, v94);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v97 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+      uint64_t v98 = (uint64_t) v20;
+      TASSIGN(v97, v98);
+      TEXTRACT(v97, v84, v94, v9);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID4);
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v99 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+      uint64_t v100 = (uint64_t) v20;
+      TASSIGN(v99, v100);
+      int32_t v101 = (int32_t) ((uint32_t) ((int32_t) v94) + (uint32_t) v8);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+      TEXTRACT(v99, v79, v9, v101);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v102 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+      uint64_t v103 = (uint64_t) v15;
+      TASSIGN(v102, v103);
+      TEXTRACT(v102, v84, v101, v9);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID5);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v104 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v105 = (uint64_t) v17;
+      TASSIGN(v104, v105);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID4);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v104, v104, v95, v97);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v106 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v107 = (uint64_t) v17;
+      TASSIGN(v106, v107);
+      pipe_barrier(PIPE_M);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID5);
+      TMATMUL_ACC(v106, v106, v99, v102);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    };
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID7);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID3);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    for (size_t v108 = v23; v108 < v22; v108 += v21) {
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v109 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+      uint64_t v110 = (uint64_t) v16;
+      TASSIGN(v109, v110);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+      TEXTRACT(v109, v79, v9, v108);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v111 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+      uint64_t v112 = (uint64_t) v20;
+      TASSIGN(v111, v112);
+      TEXTRACT(v111, v89, v108, v9);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID6);
+      Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v113 = Tile<TileType::Left, int8_t, 16, 128, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v14, v8);
+      uint64_t v114 = (uint64_t) v20;
+      TASSIGN(v113, v114);
+      int32_t v115 = (int32_t) ((uint32_t) ((int32_t) v108) + (uint32_t) v8);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+      TEXTRACT(v113, v79, v9, v115);
+      Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v116 = Tile<TileType::Right, int8_t, 128, 256, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v8, v11);
+      uint64_t v117 = (uint64_t) v15;
+      TASSIGN(v116, v117);
+      TEXTRACT(v116, v89, v115, v9);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID7);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v118 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v119 = (uint64_t) v20;
+      TASSIGN(v118, v119);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID6);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v118, v118, v109, v111);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+      Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v120 = Tile<TileType::Acc, int32_t, 16, 256, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v14, v11);
+      uint64_t v121 = (uint64_t) v20;
+      TASSIGN(v120, v121);
+      pipe_barrier(PIPE_M);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID7);
+      TMATMUL_ACC(v120, v120, v113, v116);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    };
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+  }
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID6);
+  set_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 256> v122 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v123 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v124 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v4 + (v7 + v7 * (unsigned) v11 + v7 * (unsigned) v12), v122, v123);
+  wait_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  TSTORE(v124, v39);
+  pto::Shape<1, 1, 1, 16, 256> v125 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v126 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v127 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v5 + (v7 + v7 * (unsigned) v11 + v7 * (unsigned) v12), v125, v126);
+  TSTORE(v127, v58);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  #endif // __DAV_CUBE__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_local_i8_inline43__rv_v2
+    __gm__ Tensor* x_local_i8_inline43__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int8_t* x_local_i8_inline43__rv_v2 = reinterpret_cast<__gm__ int8_t*>(x_local_i8_inline43__rv_v2_tensor->buffer.addr) + x_local_i8_inline43__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: shared_w1__ssa_v0
+    __gm__ Tensor* shared_w1__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* shared_w1__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(shared_w1__ssa_v0_tensor->buffer.addr) + shared_w1__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: shared_w3__ssa_v0
+    __gm__ Tensor* shared_w3__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ int8_t* shared_w3__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(shared_w3__ssa_v0_tensor->buffer.addr) + shared_w3__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ int32_t* ret0__out = reinterpret_cast<__gm__ int32_t*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack tensor: ret1__out
+    __gm__ Tensor* ret1__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ int32_t* ret1__out = reinterpret_cast<__gm__ int32_t*>(ret1__out_tensor->buffer.addr) + ret1__out_tensor->start_offset;
+
+    // Unpack scalar: n0_inline113__idx_v0
+    union { uint64_t u64; int64_t val; } n0_inline113__idx_v0_conv;
+    n0_inline113__idx_v0_conv.u64 = args[5];
+    int64_t n0_inline113__idx_v0 = n0_inline113__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    sh_gate_up_matmul(x_local_i8_inline43__rv_v2, shared_w1__ssa_v0, shared_w3__ssa_v0, ret0__out, ret1__out, n0_inline113__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aic/sh_w2_matmul.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aic/sh_w2_matmul.cpp
new file mode 100644
index 000000000..3c3299c6e
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aic/sh_w2_matmul.cpp
@@ -0,0 +1,238 @@
+// Kernel Function: sh_w2_matmul
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_w2_matmul(__gm__ int8_t* v1, __gm__ int8_t* v2, __gm__ int32_t* v3, int32_t v4) {
+  unsigned v5 = 0;
+  const int32_t v6 = 64;
+  const int32_t v7 = 128;
+  const int32_t v8 = 0;
+  const int32_t v9 = 512;
+  const int32_t v10 = 1;
+  const int32_t v11 = 4096;
+  const int32_t v12 = 16;
+  const int64_t v13 = 32768;
+  const int64_t v14 = 1024;
+  const int64_t v15 = 8192;
+  const int64_t v16 = 0;
+  using T = float;
+
+  #if defined(__DAV_CUBE__)
+  size_t v17 = (size_t) v9;
+  size_t v18 = (size_t) v8;
+  size_t v19 = (size_t) v7;
+  Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v20 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v12, v9);
+  uint64_t v21 = (uint64_t) v16;
+  TASSIGN(v20, v21);
+  pto::Shape<1, 1, 1, 16, 512> v22 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v23 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v24 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v11 + v5 * (unsigned) v10), v22, v23);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  TLOAD(v20, v24);
+  Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+  uint64_t v26 = (uint64_t) v15;
+  TASSIGN(v25, v26);
+  pto::Shape<1, 1, 1, 512, 512> v27 = pto::Shape<1, 1, 1, 512, 512>();
+  pto::Stride<512, 512, 512, 1, 4096> v28 = pto::Stride<512, 512, 512, 1, 4096>();
+  GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN> v29 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN>(v2 + (v5 + v5 * (unsigned) v10 + (unsigned) v4 * (unsigned) v11), v27, v28);
+  TLOAD(v25, v29);
+  set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v12, v9);
+  uint64_t v31 = (uint64_t) v16;
+  TASSIGN(v30, v31);
+  wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID0);
+  for (size_t v32 = v18; v32 < v17; v32 += v19) {
+    int32_t v33 = (int32_t) v32;
+    Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v12, v6);
+    uint64_t v35 = (uint64_t) v16;
+    TASSIGN(v34, v35);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    pipe_barrier(PIPE_MTE1);
+    TEXTRACT(v34, v20, v8, v32);
+    Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v6, v9);
+    uint64_t v37 = (uint64_t) v16;
+    TASSIGN(v36, v37);
+    TEXTRACT(v36, v25, v32, v8);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v12, v6);
+    uint64_t v39 = (uint64_t) v14;
+    TASSIGN(v38, v39);
+    int32_t v40 = (int32_t) ((uint32_t) v33 + (uint32_t) v6);
+    wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+    TEXTRACT(v38, v20, v8, v40);
+    Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v6, v9);
+    uint64_t v42 = (uint64_t) v13;
+    TASSIGN(v41, v42);
+    TEXTRACT(v41, v25, v40, v8);
+    set_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID0);
+    if (v33 == v8) {
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v12, v9);
+      uint64_t v44 = (uint64_t) v16;
+      TASSIGN(v43, v44);
+      pipe_barrier(PIPE_M);
+      TMATMUL(v43, v34, v36);
+    } else {
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v12, v9);
+      uint64_t v46 = (uint64_t) v16;
+      TASSIGN(v45, v46);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v45, v45, v34, v36);
+    };
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+    Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v12, v9);
+    uint64_t v48 = (uint64_t) v16;
+    TASSIGN(v47, v48);
+    pipe_barrier(PIPE_M);
+    wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID1);
+    TMATMUL_ACC(v47, v47, v38, v41);
+    set_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  }
+  set_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID2);
+  for (size_t v49 = v17; v49 < ((size_t) v11); v49 += v17) {
+    int32_t v50 = (int32_t) v49;
+    Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v51 = Tile<TileType::Mat, int8_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v12, v9);
+    uint64_t v52 = (uint64_t) v16;
+    TASSIGN(v51, v52);
+    pto::Shape<1, 1, 1, 16, 512> v53 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v54 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v55 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v11 + (unsigned) v50 * (unsigned) v10), v53, v54);
+    pipe_barrier(PIPE_MTE2);
+    wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+    TLOAD(v51, v55);
+    Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Mat, int8_t, 512, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+    uint64_t v57 = (uint64_t) v15;
+    TASSIGN(v56, v57);
+    pto::Shape<1, 1, 1, 512, 512> v58 = pto::Shape<1, 1, 1, 512, 512>();
+    pto::Stride<512, 512, 512, 1, 4096> v59 = pto::Stride<512, 512, 512, 1, 4096>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN> v60 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 512, 512>, pto::Stride<512, 512, 512, 1, 4096>, pto::Layout::DN>(v2 + (v5 + (unsigned) v50 * (unsigned) v10 + (unsigned) v4 * (unsigned) v11), v58, v59);
+    TLOAD(v56, v60);
+    set_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+    wait_flag(PIPE_MTE2, PIPE_MTE1, EVENT_ID1);
+    for (size_t v61 = v18; v61 < v17; v61 += v19) {
+      Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v12, v6);
+      uint64_t v63 = (uint64_t) v16;
+      TASSIGN(v62, v63);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+      TEXTRACT(v62, v51, v8, v61);
+      Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v6, v9);
+      uint64_t v65 = (uint64_t) v16;
+      TASSIGN(v64, v65);
+      TEXTRACT(v64, v56, v61, v8);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+      Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Left, int8_t, 16, 64, BLayout::RowMajor, -1, -1, SLayout::RowMajor, 512, PadValue::Null, CompactMode::Null>(v12, v6);
+      uint64_t v67 = (uint64_t) v14;
+      TASSIGN(v66, v67);
+      int32_t v68 = (int32_t) ((uint32_t) ((int32_t) v61) + (uint32_t) v6);
+      wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+      TEXTRACT(v66, v51, v8, v68);
+      Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null> v69 = Tile<TileType::Right, int8_t, 64, 512, BLayout::RowMajor, -1, -1, SLayout::ColMajor, 512, PadValue::Null, CompactMode::Null>(v6, v9);
+      uint64_t v70 = (uint64_t) v13;
+      TASSIGN(v69, v70);
+      TEXTRACT(v69, v56, v68, v8);
+      set_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v71 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v12, v9);
+      uint64_t v72 = (uint64_t) v16;
+      TASSIGN(v71, v72);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID2);
+      pipe_barrier(PIPE_M);
+      TMATMUL_ACC(v71, v71, v62, v64);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+      Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null> v73 = Tile<TileType::Acc, int32_t, 16, 512, BLayout::ColMajor, -1, -1, SLayout::RowMajor, 1024, PadValue::Null, CompactMode::Null>(v12, v9);
+      uint64_t v74 = (uint64_t) v16;
+      TASSIGN(v73, v74);
+      pipe_barrier(PIPE_M);
+      wait_flag(PIPE_MTE1, PIPE_M, EVENT_ID3);
+      TMATMUL_ACC(v73, v73, v66, v69);
+      set_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+    };
+    set_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  }
+  set_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v75 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v76 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v77 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v3 + (v5 + v5 * (unsigned) v9 + v5 * (unsigned) v10), v75, v76);
+  wait_flag(PIPE_M, PIPE_FIX, EVENT_ID0);
+  TSTORE(v77, v30);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID0);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID1);
+  wait_flag(PIPE_MTE1, PIPE_MTE2, EVENT_ID1);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID3);
+  wait_flag(PIPE_M, PIPE_MTE1, EVENT_ID4);
+  #endif // __DAV_CUBE__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: sh_tile_i8_inline126__rv_v2
+    __gm__ Tensor* sh_tile_i8_inline126__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int8_t* sh_tile_i8_inline126__rv_v2 = reinterpret_cast<__gm__ int8_t*>(sh_tile_i8_inline126__rv_v2_tensor->buffer.addr) + sh_tile_i8_inline126__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: shared_w2__ssa_v0
+    __gm__ Tensor* shared_w2__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* shared_w2__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(shared_w2__ssa_v0_tensor->buffer.addr) + shared_w2__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ int32_t* ret0__out = reinterpret_cast<__gm__ int32_t*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: d0_inline64__idx_v0
+    union { uint64_t u64; int64_t val; } d0_inline64__idx_v0_conv;
+    d0_inline64__idx_v0_conv.u64 = args[3];
+    int64_t d0_inline64__idx_v0 = d0_inline64__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    sh_w2_matmul(sh_tile_i8_inline126__rv_v2, shared_w2__ssa_v0, ret0__out, d0_inline64__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/cast_x.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/cast_x.cpp
new file mode 100644
index 000000000..66adc9169
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/cast_x.cpp
@@ -0,0 +1,103 @@
+// Kernel Function: cast_x
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void cast_x(__gm__ bfloat16_t* v1, __gm__ float* v2, int32_t v3) {
+  RoundMode v4 = RoundMode::CAST_ROUND;
+  unsigned v5 = 0;
+  const int32_t v6 = 512;
+  const int32_t v7 = 1;
+  const int32_t v8 = 16384;
+  const int32_t v9 = 16;
+  const int64_t v10 = 16384;
+  const int64_t v11 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v12 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v6);
+  uint64_t v13 = (uint64_t) v11;
+  TASSIGN(v12, v13);
+  pto::Shape<1, 1, 1, 16, 512> v14 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<262144, 262144, 262144, 16384, 1> v15 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+  GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v16 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v8 + (unsigned) v3 * (unsigned) v7), v14, v15);
+  TLOAD(v12, v16);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v17 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v6);
+  uint64_t v18 = (uint64_t) v10;
+  TASSIGN(v17, v18);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v17, v12, v4);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v19 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<262144, 262144, 262144, 16384, 1> v20 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v21 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v2 + (v5 + v5 * (unsigned) v8 + (unsigned) v3 * (unsigned) v7), v19, v20);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v21, v17);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_flat_inline36__ssa_v0
+    __gm__ Tensor* x_flat_inline36__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ bfloat16_t* x_flat_inline36__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(x_flat_inline36__ssa_v0_tensor->buffer.addr) + x_flat_inline36__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: x_flat_fp32_inline49__iter_v1
+    __gm__ Tensor* x_flat_fp32_inline49__iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* x_flat_fp32_inline49__iter_v1 = reinterpret_cast<__gm__ float*>(x_flat_fp32_inline49__iter_v1_tensor->buffer.addr) + x_flat_fp32_inline49__iter_v1_tensor->start_offset;
+
+    // Unpack scalar: k0_inline59__ssa_v0
+    union { uint64_t u64; int64_t val; } k0_inline59__ssa_v0_conv;
+    k0_inline59__ssa_v0_conv.u64 = args[2];
+    int64_t k0_inline59__ssa_v0 = k0_inline59__ssa_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    cast_x(x_flat_inline36__ssa_v0, x_flat_fp32_inline49__iter_v1, k0_inline59__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/comb_sinkhorn.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/comb_sinkhorn.cpp
new file mode 100644
index 000000000..5279a3ac9
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/comb_sinkhorn.cpp
@@ -0,0 +1,3326 @@
+// Kernel Function: comb_sinkhorn
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void comb_sinkhorn(__gm__ float* v1, __gm__ float* v2) {
+  unsigned v3 = 12;
+  unsigned v4 = 8;
+  unsigned v5 = 4;
+  unsigned v6 = 0;
+  const int32_t v7 = 255;
+  const int32_t v8 = 251;
+  const int32_t v9 = 247;
+  const int32_t v10 = 243;
+  const int32_t v11 = 254;
+  const int32_t v12 = 250;
+  const int32_t v13 = 246;
+  const int32_t v14 = 242;
+  const int32_t v15 = 253;
+  const int32_t v16 = 249;
+  const int32_t v17 = 245;
+  const int32_t v18 = 241;
+  const int32_t v19 = 252;
+  const int32_t v20 = 248;
+  const int32_t v21 = 244;
+  const int32_t v22 = 240;
+  const int32_t v23 = 239;
+  const int32_t v24 = 235;
+  const int32_t v25 = 231;
+  const int32_t v26 = 227;
+  const int32_t v27 = 238;
+  const int32_t v28 = 234;
+  const int32_t v29 = 230;
+  const int32_t v30 = 226;
+  const int32_t v31 = 237;
+  const int32_t v32 = 233;
+  const int32_t v33 = 229;
+  const int32_t v34 = 225;
+  const int32_t v35 = 236;
+  const int32_t v36 = 232;
+  const int32_t v37 = 228;
+  const int32_t v38 = 224;
+  const int32_t v39 = 223;
+  const int32_t v40 = 219;
+  const int32_t v41 = 215;
+  const int32_t v42 = 211;
+  const int32_t v43 = 222;
+  const int32_t v44 = 218;
+  const int32_t v45 = 214;
+  const int32_t v46 = 210;
+  const int32_t v47 = 221;
+  const int32_t v48 = 217;
+  const int32_t v49 = 213;
+  const int32_t v50 = 209;
+  const int32_t v51 = 220;
+  const int32_t v52 = 216;
+  const int32_t v53 = 212;
+  const int32_t v54 = 208;
+  const int32_t v55 = 207;
+  const int32_t v56 = 203;
+  const int32_t v57 = 199;
+  const int32_t v58 = 195;
+  const int32_t v59 = 206;
+  const int32_t v60 = 202;
+  const int32_t v61 = 198;
+  const int32_t v62 = 194;
+  const int32_t v63 = 205;
+  const int32_t v64 = 201;
+  const int32_t v65 = 197;
+  const int32_t v66 = 193;
+  const int32_t v67 = 204;
+  const int32_t v68 = 200;
+  const int32_t v69 = 196;
+  const int32_t v70 = 192;
+  const int32_t v71 = 191;
+  const int32_t v72 = 187;
+  const int32_t v73 = 183;
+  const int32_t v74 = 179;
+  const int32_t v75 = 190;
+  const int32_t v76 = 186;
+  const int32_t v77 = 182;
+  const int32_t v78 = 178;
+  const int32_t v79 = 189;
+  const int32_t v80 = 185;
+  const int32_t v81 = 181;
+  const int32_t v82 = 177;
+  const int32_t v83 = 188;
+  const int32_t v84 = 184;
+  const int32_t v85 = 180;
+  const int32_t v86 = 176;
+  const int32_t v87 = 175;
+  const int32_t v88 = 171;
+  const int32_t v89 = 167;
+  const int32_t v90 = 163;
+  const int32_t v91 = 174;
+  const int32_t v92 = 170;
+  const int32_t v93 = 166;
+  const int32_t v94 = 162;
+  const int32_t v95 = 173;
+  const int32_t v96 = 169;
+  const int32_t v97 = 165;
+  const int32_t v98 = 161;
+  const int32_t v99 = 172;
+  const int32_t v100 = 168;
+  const int32_t v101 = 164;
+  const int32_t v102 = 160;
+  const int32_t v103 = 159;
+  const int32_t v104 = 155;
+  const int32_t v105 = 151;
+  const int32_t v106 = 147;
+  const int32_t v107 = 158;
+  const int32_t v108 = 154;
+  const int32_t v109 = 150;
+  const int32_t v110 = 146;
+  const int32_t v111 = 157;
+  const int32_t v112 = 153;
+  const int32_t v113 = 149;
+  const int32_t v114 = 145;
+  const int32_t v115 = 156;
+  const int32_t v116 = 152;
+  const int32_t v117 = 148;
+  const int32_t v118 = 144;
+  const int32_t v119 = 143;
+  const int32_t v120 = 139;
+  const int32_t v121 = 135;
+  const int32_t v122 = 131;
+  const int32_t v123 = 142;
+  const int32_t v124 = 138;
+  const int32_t v125 = 134;
+  const int32_t v126 = 130;
+  const int32_t v127 = 141;
+  const int32_t v128 = 137;
+  const int32_t v129 = 133;
+  const int32_t v130 = 129;
+  const int32_t v131 = 140;
+  const int32_t v132 = 136;
+  const int32_t v133 = 132;
+  const int32_t v134 = 128;
+  const int32_t v135 = 127;
+  const int32_t v136 = 123;
+  const int32_t v137 = 119;
+  const int32_t v138 = 115;
+  const int32_t v139 = 126;
+  const int32_t v140 = 122;
+  const int32_t v141 = 118;
+  const int32_t v142 = 114;
+  const int32_t v143 = 125;
+  const int32_t v144 = 121;
+  const int32_t v145 = 117;
+  const int32_t v146 = 113;
+  const int32_t v147 = 124;
+  const int32_t v148 = 120;
+  const int32_t v149 = 116;
+  const int32_t v150 = 112;
+  const int32_t v151 = 111;
+  const int32_t v152 = 107;
+  const int32_t v153 = 103;
+  const int32_t v154 = 99;
+  const int32_t v155 = 110;
+  const int32_t v156 = 106;
+  const int32_t v157 = 102;
+  const int32_t v158 = 98;
+  const int32_t v159 = 109;
+  const int32_t v160 = 105;
+  const int32_t v161 = 101;
+  const int32_t v162 = 97;
+  const int32_t v163 = 108;
+  const int32_t v164 = 104;
+  const int32_t v165 = 100;
+  const int32_t v166 = 96;
+  const int32_t v167 = 95;
+  const int32_t v168 = 91;
+  const int32_t v169 = 87;
+  const int32_t v170 = 83;
+  const int32_t v171 = 94;
+  const int32_t v172 = 90;
+  const int32_t v173 = 86;
+  const int32_t v174 = 82;
+  const int32_t v175 = 93;
+  const int32_t v176 = 89;
+  const int32_t v177 = 85;
+  const int32_t v178 = 81;
+  const int32_t v179 = 92;
+  const int32_t v180 = 88;
+  const int32_t v181 = 84;
+  const int32_t v182 = 80;
+  const int32_t v183 = 79;
+  const int32_t v184 = 75;
+  const int32_t v185 = 71;
+  const int32_t v186 = 67;
+  const int32_t v187 = 78;
+  const int32_t v188 = 74;
+  const int32_t v189 = 70;
+  const int32_t v190 = 66;
+  const int32_t v191 = 77;
+  const int32_t v192 = 73;
+  const int32_t v193 = 69;
+  const int32_t v194 = 65;
+  const int32_t v195 = 76;
+  const int32_t v196 = 72;
+  const int32_t v197 = 68;
+  const int32_t v198 = 64;
+  const int32_t v199 = 63;
+  const int32_t v200 = 59;
+  const int32_t v201 = 55;
+  const int32_t v202 = 51;
+  const int32_t v203 = 62;
+  const int32_t v204 = 58;
+  const int32_t v205 = 54;
+  const int32_t v206 = 50;
+  const int32_t v207 = 61;
+  const int32_t v208 = 57;
+  const int32_t v209 = 53;
+  const int32_t v210 = 49;
+  const int32_t v211 = 60;
+  const int32_t v212 = 56;
+  const int32_t v213 = 52;
+  const int32_t v214 = 48;
+  const int32_t v215 = 47;
+  const int32_t v216 = 43;
+  const int32_t v217 = 39;
+  const int32_t v218 = 35;
+  const int32_t v219 = 46;
+  const int32_t v220 = 42;
+  const int32_t v221 = 38;
+  const int32_t v222 = 34;
+  const int32_t v223 = 45;
+  const int32_t v224 = 41;
+  const int32_t v225 = 37;
+  const int32_t v226 = 33;
+  const int32_t v227 = 44;
+  const int32_t v228 = 40;
+  const int32_t v229 = 36;
+  const int32_t v230 = 32;
+  const int32_t v231 = 31;
+  const int32_t v232 = 27;
+  const int32_t v233 = 23;
+  const int32_t v234 = 19;
+  const int32_t v235 = 30;
+  const int32_t v236 = 26;
+  const int32_t v237 = 22;
+  const int32_t v238 = 18;
+  const int32_t v239 = 29;
+  const int32_t v240 = 25;
+  const int32_t v241 = 21;
+  const int32_t v242 = 17;
+  const int32_t v243 = 28;
+  const int32_t v244 = 24;
+  const int32_t v245 = 20;
+  const int32_t v246 = 15;
+  const int32_t v247 = 11;
+  const int32_t v248 = 7;
+  const int32_t v249 = 3;
+  const int32_t v250 = 14;
+  const int32_t v251 = 10;
+  const int32_t v252 = 6;
+  const int32_t v253 = 2;
+  const int32_t v254 = 13;
+  const int32_t v255 = 9;
+  const int32_t v256 = 5;
+  const float v257 = 9.99999997E-7f;
+  const int32_t v258 = 12;
+  const int32_t v259 = 8;
+  const int32_t v260 = 0;
+  const int32_t v261 = 4;
+  const int32_t v262 = 1;
+  const int32_t v263 = 16;
+  const int64_t v264 = 4800;
+  const int64_t v265 = 3776;
+  const int64_t v266 = 3264;
+  const int64_t v267 = 2752;
+  const int64_t v268 = 2240;
+  const int64_t v269 = 1728;
+  const int64_t v270 = 1216;
+  const int64_t v271 = 128;
+  const int64_t v272 = 64;
+  const int64_t v273 = 0;
+  const int64_t v274 = 5376;
+  const int64_t v275 = 4864;
+  const int64_t v276 = 4288;
+  const int64_t v277 = 704;
+  const int64_t v278 = 192;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v279 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v261);
+  uint64_t v280 = (uint64_t) v278;
+  TASSIGN(v279, v280);
+  pto::Shape<1, 1, 1, 16, 4> v281 = pto::Shape<1, 1, 1, 16, 4>();
+  pto::Stride<256, 256, 256, 16, 1> v282 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v283 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v263 + v6 * (unsigned) v262), v281, v282);
+  TLOAD(v279, v283);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v284 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v285 = (uint64_t) v277;
+  TASSIGN(v284, v285);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TFILLPAD(v284, v279);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v286 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v261);
+  uint64_t v287 = (uint64_t) v278;
+  TASSIGN(v286, v287);
+  pto::Shape<1, 1, 1, 16, 4> v288 = pto::Shape<1, 1, 1, 16, 4>();
+  pto::Stride<256, 256, 256, 16, 1> v289 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v290 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v263 + v5 * (unsigned) v262), v288, v289);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v286, v290);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v291 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v292 = (uint64_t) v276;
+  TASSIGN(v291, v292);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TFILLPAD(v291, v286);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v293 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v261);
+  uint64_t v294 = (uint64_t) v278;
+  TASSIGN(v293, v294);
+  pto::Shape<1, 1, 1, 16, 4> v295 = pto::Shape<1, 1, 1, 16, 4>();
+  pto::Stride<256, 256, 256, 16, 1> v296 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v297 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v263 + v4 * (unsigned) v262), v295, v296);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  TLOAD(v293, v297);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v298 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v299 = (uint64_t) v275;
+  TASSIGN(v298, v299);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  TFILLPAD(v298, v293);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v300 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v261);
+  uint64_t v301 = (uint64_t) v278;
+  TASSIGN(v300, v301);
+  pto::Shape<1, 1, 1, 16, 4> v302 = pto::Shape<1, 1, 1, 16, 4>();
+  pto::Stride<256, 256, 256, 16, 1> v303 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v304 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 4>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v263 + v3 * (unsigned) v262), v302, v303);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  TLOAD(v300, v304);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v305 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v306 = (uint64_t) v274;
+  TASSIGN(v305, v306);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  TFILLPAD(v305, v300);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v307 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v308 = (uint64_t) v273;
+  TASSIGN(v307, v308);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v309 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v310 = (uint64_t) v272;
+  TASSIGN(v309, v310);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v311 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v312 = (uint64_t) v271;
+  TASSIGN(v311, v312);
+  TROWMAX(v311, v284, v307);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v313 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v314 = (uint64_t) v277;
+  TASSIGN(v313, v314);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDSUB(v313, v284, v311);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v315 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v316 = (uint64_t) v277;
+  TASSIGN(v315, v316);
+  pipe_barrier(PIPE_V);
+  TEXP(v315, v313);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v317 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v318 = (uint64_t) v271;
+  TASSIGN(v317, v318);
+  TROWMAX(v317, v291, v307);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v319 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v320 = (uint64_t) v276;
+  TASSIGN(v319, v320);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDSUB(v319, v291, v317);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v321 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v322 = (uint64_t) v276;
+  TASSIGN(v321, v322);
+  pipe_barrier(PIPE_V);
+  TEXP(v321, v319);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v323 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v324 = (uint64_t) v271;
+  TASSIGN(v323, v324);
+  TROWMAX(v323, v298, v307);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v325 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v326 = (uint64_t) v275;
+  TASSIGN(v325, v326);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDSUB(v325, v298, v323);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v327 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v328 = (uint64_t) v275;
+  TASSIGN(v327, v328);
+  pipe_barrier(PIPE_V);
+  TEXP(v327, v325);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v329 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v330 = (uint64_t) v273;
+  TASSIGN(v329, v330);
+  TROWMAX(v329, v305, v307);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v331 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v332 = (uint64_t) v274;
+  TASSIGN(v331, v332);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDSUB(v331, v305, v329);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v333 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v334 = (uint64_t) v274;
+  TASSIGN(v333, v334);
+  pipe_barrier(PIPE_V);
+  TEXP(v333, v331);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v335 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v336 = (uint64_t) v273;
+  TASSIGN(v335, v336);
+  TROWSUM(v335, v315, v309);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v337 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v338 = (uint64_t) v277;
+  TASSIGN(v337, v338);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v337, v315, v335);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v339 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v340 = (uint64_t) v277;
+  TASSIGN(v339, v340);
+  pipe_barrier(PIPE_V);
+  TADDS(v339, v337, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v341 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v342 = (uint64_t) v273;
+  TASSIGN(v341, v342);
+  TROWSUM(v341, v321, v309);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v343 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v344 = (uint64_t) v276;
+  TASSIGN(v343, v344);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v343, v321, v341);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v345 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v346 = (uint64_t) v276;
+  TASSIGN(v345, v346);
+  pipe_barrier(PIPE_V);
+  TADDS(v345, v343, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v347 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v348 = (uint64_t) v273;
+  TASSIGN(v347, v348);
+  TROWSUM(v347, v327, v309);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v349 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v350 = (uint64_t) v275;
+  TASSIGN(v349, v350);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v349, v327, v347);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v351 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v352 = (uint64_t) v275;
+  TASSIGN(v351, v352);
+  pipe_barrier(PIPE_V);
+  TADDS(v351, v349, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v353 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v354 = (uint64_t) v273;
+  TASSIGN(v353, v354);
+  TROWSUM(v353, v333, v309);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v355 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v356 = (uint64_t) v274;
+  TASSIGN(v355, v356);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v355, v333, v353);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null> v357 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Min, CompactMode::Null>(v263, v259);
+  uint64_t v358 = (uint64_t) v274;
+  TASSIGN(v357, v358);
+  pipe_barrier(PIPE_V);
+  TADDS(v357, v355, v257);
+  v339.SetValidShape(v263, v261);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v359 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v360 = (uint64_t) v270;
+  TASSIGN(v359, v360);
+  TFILLPAD(v359, v339);
+  v345.SetValidShape(v263, v261);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v361 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v362 = (uint64_t) v269;
+  TASSIGN(v361, v362);
+  TFILLPAD(v361, v345);
+  v351.SetValidShape(v263, v261);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v363 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v364 = (uint64_t) v268;
+  TASSIGN(v363, v364);
+  TFILLPAD(v363, v351);
+  v357.SetValidShape(v263, v261);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v365 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v366 = (uint64_t) v267;
+  TASSIGN(v365, v366);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v365, v357);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v367 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v368 = (uint64_t) v273;
+  TASSIGN(v367, v368);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v369 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v370 = (uint64_t) v266;
+  TASSIGN(v369, v370);
+  TADD(v369, v359, v361);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v371 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v372 = (uint64_t) v265;
+  TASSIGN(v371, v372);
+  pipe_barrier(PIPE_V);
+  TADD(v371, v363, v365);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v373 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v374 = (uint64_t) v266;
+  TASSIGN(v373, v374);
+  pipe_barrier(PIPE_V);
+  TADD(v373, v369, v371);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v375 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v376 = (uint64_t) v266;
+  TASSIGN(v375, v376);
+  pipe_barrier(PIPE_V);
+  TADDS(v375, v373, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v377 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v378 = (uint64_t) v270;
+  TASSIGN(v377, v378);
+  pipe_barrier(PIPE_V);
+  TDIV(v377, v359, v375);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v379 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v380 = (uint64_t) v269;
+  TASSIGN(v379, v380);
+  TDIV(v379, v361, v375);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v381 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v382 = (uint64_t) v268;
+  TASSIGN(v381, v382);
+  TDIV(v381, v363, v375);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v383 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v384 = (uint64_t) v267;
+  TASSIGN(v383, v384);
+  TDIV(v383, v365, v375);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v385 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v386 = (uint64_t) v272;
+  TASSIGN(v385, v386);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v385, v377, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v387 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v388 = (uint64_t) v272;
+  TASSIGN(v387, v388);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v389 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v390 = (uint64_t) v264;
+  TASSIGN(v389, v390);
+  pipe_barrier(PIPE_V);
+  TADDS(v389, v387, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v391 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v392 = (uint64_t) v264;
+  TASSIGN(v391, v392);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v393 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v394 = (uint64_t) v270;
+  TASSIGN(v393, v394);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v393, v377, v391);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v395 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v396 = (uint64_t) v272;
+  TASSIGN(v395, v396);
+  TROWSUM(v395, v379, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v397 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v398 = (uint64_t) v272;
+  TASSIGN(v397, v398);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v399 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v400 = (uint64_t) v264;
+  TASSIGN(v399, v400);
+  pipe_barrier(PIPE_V);
+  TADDS(v399, v397, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v401 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v402 = (uint64_t) v264;
+  TASSIGN(v401, v402);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v403 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v404 = (uint64_t) v269;
+  TASSIGN(v403, v404);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v403, v379, v401);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v405 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v406 = (uint64_t) v272;
+  TASSIGN(v405, v406);
+  TROWSUM(v405, v381, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v407 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v408 = (uint64_t) v272;
+  TASSIGN(v407, v408);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v409 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v410 = (uint64_t) v264;
+  TASSIGN(v409, v410);
+  pipe_barrier(PIPE_V);
+  TADDS(v409, v407, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v411 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v412 = (uint64_t) v264;
+  TASSIGN(v411, v412);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v413 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v414 = (uint64_t) v268;
+  TASSIGN(v413, v414);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v413, v381, v411);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v415 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v416 = (uint64_t) v272;
+  TASSIGN(v415, v416);
+  TROWSUM(v415, v383, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v417 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v418 = (uint64_t) v272;
+  TASSIGN(v417, v418);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v419 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v420 = (uint64_t) v264;
+  TASSIGN(v419, v420);
+  pipe_barrier(PIPE_V);
+  TADDS(v419, v417, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v421 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v422 = (uint64_t) v264;
+  TASSIGN(v421, v422);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v423 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v424 = (uint64_t) v267;
+  TASSIGN(v423, v424);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v423, v383, v421);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v425 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v426 = (uint64_t) v266;
+  TASSIGN(v425, v426);
+  TADD(v425, v393, v403);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v427 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v428 = (uint64_t) v265;
+  TASSIGN(v427, v428);
+  pipe_barrier(PIPE_V);
+  TADD(v427, v413, v423);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v429 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v430 = (uint64_t) v266;
+  TASSIGN(v429, v430);
+  pipe_barrier(PIPE_V);
+  TADD(v429, v425, v427);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v431 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v432 = (uint64_t) v266;
+  TASSIGN(v431, v432);
+  pipe_barrier(PIPE_V);
+  TADDS(v431, v429, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v433 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v434 = (uint64_t) v270;
+  TASSIGN(v433, v434);
+  pipe_barrier(PIPE_V);
+  TDIV(v433, v393, v431);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v435 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v436 = (uint64_t) v269;
+  TASSIGN(v435, v436);
+  TDIV(v435, v403, v431);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v437 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v438 = (uint64_t) v268;
+  TASSIGN(v437, v438);
+  TDIV(v437, v413, v431);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v439 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v440 = (uint64_t) v267;
+  TASSIGN(v439, v440);
+  TDIV(v439, v423, v431);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v441 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v442 = (uint64_t) v272;
+  TASSIGN(v441, v442);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v441, v433, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v443 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v444 = (uint64_t) v272;
+  TASSIGN(v443, v444);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v445 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v446 = (uint64_t) v264;
+  TASSIGN(v445, v446);
+  pipe_barrier(PIPE_V);
+  TADDS(v445, v443, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v447 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v448 = (uint64_t) v264;
+  TASSIGN(v447, v448);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v449 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v450 = (uint64_t) v270;
+  TASSIGN(v449, v450);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v449, v433, v447);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v451 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v452 = (uint64_t) v272;
+  TASSIGN(v451, v452);
+  TROWSUM(v451, v435, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v453 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v454 = (uint64_t) v272;
+  TASSIGN(v453, v454);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v455 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v456 = (uint64_t) v264;
+  TASSIGN(v455, v456);
+  pipe_barrier(PIPE_V);
+  TADDS(v455, v453, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v457 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v458 = (uint64_t) v264;
+  TASSIGN(v457, v458);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v459 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v460 = (uint64_t) v269;
+  TASSIGN(v459, v460);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v459, v435, v457);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v461 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v462 = (uint64_t) v272;
+  TASSIGN(v461, v462);
+  TROWSUM(v461, v437, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v463 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v464 = (uint64_t) v272;
+  TASSIGN(v463, v464);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v465 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v466 = (uint64_t) v264;
+  TASSIGN(v465, v466);
+  pipe_barrier(PIPE_V);
+  TADDS(v465, v463, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v467 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v468 = (uint64_t) v264;
+  TASSIGN(v467, v468);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v469 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v470 = (uint64_t) v268;
+  TASSIGN(v469, v470);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v469, v437, v467);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v471 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v472 = (uint64_t) v272;
+  TASSIGN(v471, v472);
+  TROWSUM(v471, v439, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v473 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v474 = (uint64_t) v272;
+  TASSIGN(v473, v474);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v475 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v476 = (uint64_t) v264;
+  TASSIGN(v475, v476);
+  pipe_barrier(PIPE_V);
+  TADDS(v475, v473, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v477 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v478 = (uint64_t) v264;
+  TASSIGN(v477, v478);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v479 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v480 = (uint64_t) v267;
+  TASSIGN(v479, v480);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v479, v439, v477);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v481 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v482 = (uint64_t) v266;
+  TASSIGN(v481, v482);
+  TADD(v481, v449, v459);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v483 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v484 = (uint64_t) v265;
+  TASSIGN(v483, v484);
+  pipe_barrier(PIPE_V);
+  TADD(v483, v469, v479);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v485 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v486 = (uint64_t) v266;
+  TASSIGN(v485, v486);
+  pipe_barrier(PIPE_V);
+  TADD(v485, v481, v483);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v487 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v488 = (uint64_t) v266;
+  TASSIGN(v487, v488);
+  pipe_barrier(PIPE_V);
+  TADDS(v487, v485, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v489 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v490 = (uint64_t) v270;
+  TASSIGN(v489, v490);
+  pipe_barrier(PIPE_V);
+  TDIV(v489, v449, v487);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v491 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v492 = (uint64_t) v269;
+  TASSIGN(v491, v492);
+  TDIV(v491, v459, v487);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v493 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v494 = (uint64_t) v268;
+  TASSIGN(v493, v494);
+  TDIV(v493, v469, v487);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v495 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v496 = (uint64_t) v267;
+  TASSIGN(v495, v496);
+  TDIV(v495, v479, v487);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v497 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v498 = (uint64_t) v272;
+  TASSIGN(v497, v498);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v497, v489, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v499 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v500 = (uint64_t) v272;
+  TASSIGN(v499, v500);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v501 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v502 = (uint64_t) v264;
+  TASSIGN(v501, v502);
+  pipe_barrier(PIPE_V);
+  TADDS(v501, v499, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v503 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v504 = (uint64_t) v264;
+  TASSIGN(v503, v504);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v505 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v506 = (uint64_t) v270;
+  TASSIGN(v505, v506);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v505, v489, v503);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v507 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v508 = (uint64_t) v272;
+  TASSIGN(v507, v508);
+  TROWSUM(v507, v491, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v509 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v510 = (uint64_t) v272;
+  TASSIGN(v509, v510);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v511 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v512 = (uint64_t) v264;
+  TASSIGN(v511, v512);
+  pipe_barrier(PIPE_V);
+  TADDS(v511, v509, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v513 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v514 = (uint64_t) v264;
+  TASSIGN(v513, v514);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v515 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v516 = (uint64_t) v269;
+  TASSIGN(v515, v516);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v515, v491, v513);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v517 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v518 = (uint64_t) v272;
+  TASSIGN(v517, v518);
+  TROWSUM(v517, v493, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v519 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v520 = (uint64_t) v272;
+  TASSIGN(v519, v520);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v521 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v522 = (uint64_t) v264;
+  TASSIGN(v521, v522);
+  pipe_barrier(PIPE_V);
+  TADDS(v521, v519, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v523 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v524 = (uint64_t) v264;
+  TASSIGN(v523, v524);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v525 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v526 = (uint64_t) v268;
+  TASSIGN(v525, v526);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v525, v493, v523);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v527 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v528 = (uint64_t) v272;
+  TASSIGN(v527, v528);
+  TROWSUM(v527, v495, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v529 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v530 = (uint64_t) v272;
+  TASSIGN(v529, v530);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v531 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v532 = (uint64_t) v264;
+  TASSIGN(v531, v532);
+  pipe_barrier(PIPE_V);
+  TADDS(v531, v529, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v533 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v534 = (uint64_t) v264;
+  TASSIGN(v533, v534);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v535 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v536 = (uint64_t) v267;
+  TASSIGN(v535, v536);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v535, v495, v533);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v537 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v538 = (uint64_t) v266;
+  TASSIGN(v537, v538);
+  TADD(v537, v505, v515);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v539 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v540 = (uint64_t) v265;
+  TASSIGN(v539, v540);
+  pipe_barrier(PIPE_V);
+  TADD(v539, v525, v535);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v541 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v542 = (uint64_t) v266;
+  TASSIGN(v541, v542);
+  pipe_barrier(PIPE_V);
+  TADD(v541, v537, v539);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v543 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v544 = (uint64_t) v266;
+  TASSIGN(v543, v544);
+  pipe_barrier(PIPE_V);
+  TADDS(v543, v541, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v545 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v546 = (uint64_t) v270;
+  TASSIGN(v545, v546);
+  pipe_barrier(PIPE_V);
+  TDIV(v545, v505, v543);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v547 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v548 = (uint64_t) v269;
+  TASSIGN(v547, v548);
+  TDIV(v547, v515, v543);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v549 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v550 = (uint64_t) v268;
+  TASSIGN(v549, v550);
+  TDIV(v549, v525, v543);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v551 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v552 = (uint64_t) v267;
+  TASSIGN(v551, v552);
+  TDIV(v551, v535, v543);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v553 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v554 = (uint64_t) v272;
+  TASSIGN(v553, v554);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v553, v545, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v555 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v556 = (uint64_t) v272;
+  TASSIGN(v555, v556);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v557 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v558 = (uint64_t) v264;
+  TASSIGN(v557, v558);
+  pipe_barrier(PIPE_V);
+  TADDS(v557, v555, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v559 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v560 = (uint64_t) v264;
+  TASSIGN(v559, v560);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v561 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v562 = (uint64_t) v270;
+  TASSIGN(v561, v562);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v561, v545, v559);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v563 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v564 = (uint64_t) v272;
+  TASSIGN(v563, v564);
+  TROWSUM(v563, v547, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v565 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v566 = (uint64_t) v272;
+  TASSIGN(v565, v566);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v567 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v568 = (uint64_t) v264;
+  TASSIGN(v567, v568);
+  pipe_barrier(PIPE_V);
+  TADDS(v567, v565, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v569 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v570 = (uint64_t) v264;
+  TASSIGN(v569, v570);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v571 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v572 = (uint64_t) v269;
+  TASSIGN(v571, v572);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v571, v547, v569);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v573 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v574 = (uint64_t) v272;
+  TASSIGN(v573, v574);
+  TROWSUM(v573, v549, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v575 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v576 = (uint64_t) v272;
+  TASSIGN(v575, v576);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v577 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v578 = (uint64_t) v264;
+  TASSIGN(v577, v578);
+  pipe_barrier(PIPE_V);
+  TADDS(v577, v575, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v579 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v580 = (uint64_t) v264;
+  TASSIGN(v579, v580);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v581 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v582 = (uint64_t) v268;
+  TASSIGN(v581, v582);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v581, v549, v579);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v583 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v584 = (uint64_t) v272;
+  TASSIGN(v583, v584);
+  TROWSUM(v583, v551, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v585 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v586 = (uint64_t) v272;
+  TASSIGN(v585, v586);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v587 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v588 = (uint64_t) v264;
+  TASSIGN(v587, v588);
+  pipe_barrier(PIPE_V);
+  TADDS(v587, v585, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v589 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v590 = (uint64_t) v264;
+  TASSIGN(v589, v590);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v591 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v592 = (uint64_t) v267;
+  TASSIGN(v591, v592);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v591, v551, v589);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v593 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v594 = (uint64_t) v266;
+  TASSIGN(v593, v594);
+  TADD(v593, v561, v571);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v595 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v596 = (uint64_t) v265;
+  TASSIGN(v595, v596);
+  pipe_barrier(PIPE_V);
+  TADD(v595, v581, v591);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v597 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v598 = (uint64_t) v266;
+  TASSIGN(v597, v598);
+  pipe_barrier(PIPE_V);
+  TADD(v597, v593, v595);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v599 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v600 = (uint64_t) v266;
+  TASSIGN(v599, v600);
+  pipe_barrier(PIPE_V);
+  TADDS(v599, v597, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v601 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v602 = (uint64_t) v270;
+  TASSIGN(v601, v602);
+  pipe_barrier(PIPE_V);
+  TDIV(v601, v561, v599);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v603 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v604 = (uint64_t) v269;
+  TASSIGN(v603, v604);
+  TDIV(v603, v571, v599);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v605 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v606 = (uint64_t) v268;
+  TASSIGN(v605, v606);
+  TDIV(v605, v581, v599);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v607 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v608 = (uint64_t) v267;
+  TASSIGN(v607, v608);
+  TDIV(v607, v591, v599);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v609 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v610 = (uint64_t) v272;
+  TASSIGN(v609, v610);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v609, v601, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v611 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v612 = (uint64_t) v272;
+  TASSIGN(v611, v612);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v613 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v614 = (uint64_t) v264;
+  TASSIGN(v613, v614);
+  pipe_barrier(PIPE_V);
+  TADDS(v613, v611, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v615 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v616 = (uint64_t) v264;
+  TASSIGN(v615, v616);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v617 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v618 = (uint64_t) v270;
+  TASSIGN(v617, v618);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v617, v601, v615);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v619 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v620 = (uint64_t) v272;
+  TASSIGN(v619, v620);
+  TROWSUM(v619, v603, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v621 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v622 = (uint64_t) v272;
+  TASSIGN(v621, v622);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v623 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v624 = (uint64_t) v264;
+  TASSIGN(v623, v624);
+  pipe_barrier(PIPE_V);
+  TADDS(v623, v621, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v625 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v626 = (uint64_t) v264;
+  TASSIGN(v625, v626);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v627 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v628 = (uint64_t) v269;
+  TASSIGN(v627, v628);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v627, v603, v625);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v629 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v630 = (uint64_t) v272;
+  TASSIGN(v629, v630);
+  TROWSUM(v629, v605, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v631 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v632 = (uint64_t) v272;
+  TASSIGN(v631, v632);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v633 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v634 = (uint64_t) v264;
+  TASSIGN(v633, v634);
+  pipe_barrier(PIPE_V);
+  TADDS(v633, v631, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v635 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v636 = (uint64_t) v264;
+  TASSIGN(v635, v636);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v637 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v638 = (uint64_t) v268;
+  TASSIGN(v637, v638);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v637, v605, v635);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v639 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v640 = (uint64_t) v272;
+  TASSIGN(v639, v640);
+  TROWSUM(v639, v607, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v641 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v642 = (uint64_t) v272;
+  TASSIGN(v641, v642);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v643 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v644 = (uint64_t) v264;
+  TASSIGN(v643, v644);
+  pipe_barrier(PIPE_V);
+  TADDS(v643, v641, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v645 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v646 = (uint64_t) v264;
+  TASSIGN(v645, v646);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v647 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v648 = (uint64_t) v267;
+  TASSIGN(v647, v648);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v647, v607, v645);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v649 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v650 = (uint64_t) v266;
+  TASSIGN(v649, v650);
+  TADD(v649, v617, v627);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v651 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v652 = (uint64_t) v265;
+  TASSIGN(v651, v652);
+  pipe_barrier(PIPE_V);
+  TADD(v651, v637, v647);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v653 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v654 = (uint64_t) v266;
+  TASSIGN(v653, v654);
+  pipe_barrier(PIPE_V);
+  TADD(v653, v649, v651);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v655 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v656 = (uint64_t) v266;
+  TASSIGN(v655, v656);
+  pipe_barrier(PIPE_V);
+  TADDS(v655, v653, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v657 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v658 = (uint64_t) v270;
+  TASSIGN(v657, v658);
+  pipe_barrier(PIPE_V);
+  TDIV(v657, v617, v655);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v659 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v660 = (uint64_t) v269;
+  TASSIGN(v659, v660);
+  TDIV(v659, v627, v655);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v661 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v662 = (uint64_t) v268;
+  TASSIGN(v661, v662);
+  TDIV(v661, v637, v655);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v663 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v664 = (uint64_t) v267;
+  TASSIGN(v663, v664);
+  TDIV(v663, v647, v655);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v665 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v666 = (uint64_t) v272;
+  TASSIGN(v665, v666);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v665, v657, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v667 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v668 = (uint64_t) v272;
+  TASSIGN(v667, v668);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v669 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v670 = (uint64_t) v264;
+  TASSIGN(v669, v670);
+  pipe_barrier(PIPE_V);
+  TADDS(v669, v667, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v671 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v672 = (uint64_t) v264;
+  TASSIGN(v671, v672);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v673 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v674 = (uint64_t) v270;
+  TASSIGN(v673, v674);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v673, v657, v671);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v675 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v676 = (uint64_t) v272;
+  TASSIGN(v675, v676);
+  TROWSUM(v675, v659, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v677 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v678 = (uint64_t) v272;
+  TASSIGN(v677, v678);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v679 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v680 = (uint64_t) v264;
+  TASSIGN(v679, v680);
+  pipe_barrier(PIPE_V);
+  TADDS(v679, v677, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v681 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v682 = (uint64_t) v264;
+  TASSIGN(v681, v682);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v683 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v684 = (uint64_t) v269;
+  TASSIGN(v683, v684);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v683, v659, v681);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v685 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v686 = (uint64_t) v272;
+  TASSIGN(v685, v686);
+  TROWSUM(v685, v661, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v687 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v688 = (uint64_t) v272;
+  TASSIGN(v687, v688);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v689 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v690 = (uint64_t) v264;
+  TASSIGN(v689, v690);
+  pipe_barrier(PIPE_V);
+  TADDS(v689, v687, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v691 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v692 = (uint64_t) v264;
+  TASSIGN(v691, v692);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v693 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v694 = (uint64_t) v268;
+  TASSIGN(v693, v694);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v693, v661, v691);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v695 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v696 = (uint64_t) v272;
+  TASSIGN(v695, v696);
+  TROWSUM(v695, v663, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v697 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v698 = (uint64_t) v272;
+  TASSIGN(v697, v698);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v699 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v700 = (uint64_t) v264;
+  TASSIGN(v699, v700);
+  pipe_barrier(PIPE_V);
+  TADDS(v699, v697, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v701 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v702 = (uint64_t) v264;
+  TASSIGN(v701, v702);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v703 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v704 = (uint64_t) v267;
+  TASSIGN(v703, v704);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v703, v663, v701);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v705 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v706 = (uint64_t) v266;
+  TASSIGN(v705, v706);
+  TADD(v705, v673, v683);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v707 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v708 = (uint64_t) v265;
+  TASSIGN(v707, v708);
+  pipe_barrier(PIPE_V);
+  TADD(v707, v693, v703);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v709 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v710 = (uint64_t) v266;
+  TASSIGN(v709, v710);
+  pipe_barrier(PIPE_V);
+  TADD(v709, v705, v707);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v711 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v712 = (uint64_t) v266;
+  TASSIGN(v711, v712);
+  pipe_barrier(PIPE_V);
+  TADDS(v711, v709, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v713 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v714 = (uint64_t) v270;
+  TASSIGN(v713, v714);
+  pipe_barrier(PIPE_V);
+  TDIV(v713, v673, v711);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v715 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v716 = (uint64_t) v269;
+  TASSIGN(v715, v716);
+  TDIV(v715, v683, v711);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v717 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v718 = (uint64_t) v268;
+  TASSIGN(v717, v718);
+  TDIV(v717, v693, v711);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v719 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v720 = (uint64_t) v267;
+  TASSIGN(v719, v720);
+  TDIV(v719, v703, v711);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v721 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v722 = (uint64_t) v272;
+  TASSIGN(v721, v722);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v721, v713, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v723 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v724 = (uint64_t) v272;
+  TASSIGN(v723, v724);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v725 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v726 = (uint64_t) v264;
+  TASSIGN(v725, v726);
+  pipe_barrier(PIPE_V);
+  TADDS(v725, v723, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v727 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v728 = (uint64_t) v264;
+  TASSIGN(v727, v728);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v729 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v730 = (uint64_t) v270;
+  TASSIGN(v729, v730);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v729, v713, v727);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v731 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v732 = (uint64_t) v272;
+  TASSIGN(v731, v732);
+  TROWSUM(v731, v715, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v733 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v734 = (uint64_t) v272;
+  TASSIGN(v733, v734);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v735 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v736 = (uint64_t) v264;
+  TASSIGN(v735, v736);
+  pipe_barrier(PIPE_V);
+  TADDS(v735, v733, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v737 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v738 = (uint64_t) v264;
+  TASSIGN(v737, v738);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v739 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v740 = (uint64_t) v269;
+  TASSIGN(v739, v740);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v739, v715, v737);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v741 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v742 = (uint64_t) v272;
+  TASSIGN(v741, v742);
+  TROWSUM(v741, v717, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v743 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v744 = (uint64_t) v272;
+  TASSIGN(v743, v744);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v745 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v746 = (uint64_t) v264;
+  TASSIGN(v745, v746);
+  pipe_barrier(PIPE_V);
+  TADDS(v745, v743, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v747 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v748 = (uint64_t) v264;
+  TASSIGN(v747, v748);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v749 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v750 = (uint64_t) v268;
+  TASSIGN(v749, v750);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v749, v717, v747);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v751 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v752 = (uint64_t) v272;
+  TASSIGN(v751, v752);
+  TROWSUM(v751, v719, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v753 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v754 = (uint64_t) v272;
+  TASSIGN(v753, v754);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v755 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v756 = (uint64_t) v264;
+  TASSIGN(v755, v756);
+  pipe_barrier(PIPE_V);
+  TADDS(v755, v753, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v757 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v758 = (uint64_t) v264;
+  TASSIGN(v757, v758);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v759 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v760 = (uint64_t) v267;
+  TASSIGN(v759, v760);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v759, v719, v757);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v761 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v762 = (uint64_t) v266;
+  TASSIGN(v761, v762);
+  TADD(v761, v729, v739);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v763 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v764 = (uint64_t) v265;
+  TASSIGN(v763, v764);
+  pipe_barrier(PIPE_V);
+  TADD(v763, v749, v759);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v765 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v766 = (uint64_t) v266;
+  TASSIGN(v765, v766);
+  pipe_barrier(PIPE_V);
+  TADD(v765, v761, v763);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v767 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v768 = (uint64_t) v266;
+  TASSIGN(v767, v768);
+  pipe_barrier(PIPE_V);
+  TADDS(v767, v765, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v769 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v770 = (uint64_t) v270;
+  TASSIGN(v769, v770);
+  pipe_barrier(PIPE_V);
+  TDIV(v769, v729, v767);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v771 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v772 = (uint64_t) v269;
+  TASSIGN(v771, v772);
+  TDIV(v771, v739, v767);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v773 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v774 = (uint64_t) v268;
+  TASSIGN(v773, v774);
+  TDIV(v773, v749, v767);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v775 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v776 = (uint64_t) v267;
+  TASSIGN(v775, v776);
+  TDIV(v775, v759, v767);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v777 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v778 = (uint64_t) v272;
+  TASSIGN(v777, v778);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v777, v769, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v779 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v780 = (uint64_t) v272;
+  TASSIGN(v779, v780);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v781 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v782 = (uint64_t) v264;
+  TASSIGN(v781, v782);
+  pipe_barrier(PIPE_V);
+  TADDS(v781, v779, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v783 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v784 = (uint64_t) v264;
+  TASSIGN(v783, v784);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v785 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v786 = (uint64_t) v270;
+  TASSIGN(v785, v786);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v785, v769, v783);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v787 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v788 = (uint64_t) v272;
+  TASSIGN(v787, v788);
+  TROWSUM(v787, v771, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v789 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v790 = (uint64_t) v272;
+  TASSIGN(v789, v790);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v791 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v792 = (uint64_t) v264;
+  TASSIGN(v791, v792);
+  pipe_barrier(PIPE_V);
+  TADDS(v791, v789, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v793 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v794 = (uint64_t) v264;
+  TASSIGN(v793, v794);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v795 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v796 = (uint64_t) v269;
+  TASSIGN(v795, v796);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v795, v771, v793);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v797 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v798 = (uint64_t) v272;
+  TASSIGN(v797, v798);
+  TROWSUM(v797, v773, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v799 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v800 = (uint64_t) v272;
+  TASSIGN(v799, v800);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v801 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v802 = (uint64_t) v264;
+  TASSIGN(v801, v802);
+  pipe_barrier(PIPE_V);
+  TADDS(v801, v799, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v803 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v804 = (uint64_t) v264;
+  TASSIGN(v803, v804);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v805 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v806 = (uint64_t) v268;
+  TASSIGN(v805, v806);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v805, v773, v803);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v807 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v808 = (uint64_t) v272;
+  TASSIGN(v807, v808);
+  TROWSUM(v807, v775, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v809 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v810 = (uint64_t) v272;
+  TASSIGN(v809, v810);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v811 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v812 = (uint64_t) v264;
+  TASSIGN(v811, v812);
+  pipe_barrier(PIPE_V);
+  TADDS(v811, v809, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v813 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v814 = (uint64_t) v264;
+  TASSIGN(v813, v814);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v815 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v816 = (uint64_t) v267;
+  TASSIGN(v815, v816);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v815, v775, v813);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v817 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v818 = (uint64_t) v266;
+  TASSIGN(v817, v818);
+  TADD(v817, v785, v795);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v819 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v820 = (uint64_t) v265;
+  TASSIGN(v819, v820);
+  pipe_barrier(PIPE_V);
+  TADD(v819, v805, v815);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v821 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v822 = (uint64_t) v266;
+  TASSIGN(v821, v822);
+  pipe_barrier(PIPE_V);
+  TADD(v821, v817, v819);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v823 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v824 = (uint64_t) v266;
+  TASSIGN(v823, v824);
+  pipe_barrier(PIPE_V);
+  TADDS(v823, v821, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v825 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v826 = (uint64_t) v270;
+  TASSIGN(v825, v826);
+  pipe_barrier(PIPE_V);
+  TDIV(v825, v785, v823);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v827 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v828 = (uint64_t) v269;
+  TASSIGN(v827, v828);
+  TDIV(v827, v795, v823);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v829 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v830 = (uint64_t) v268;
+  TASSIGN(v829, v830);
+  TDIV(v829, v805, v823);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v831 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v832 = (uint64_t) v267;
+  TASSIGN(v831, v832);
+  TDIV(v831, v815, v823);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v833 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v834 = (uint64_t) v272;
+  TASSIGN(v833, v834);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v833, v825, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v835 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v836 = (uint64_t) v272;
+  TASSIGN(v835, v836);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v837 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v838 = (uint64_t) v264;
+  TASSIGN(v837, v838);
+  pipe_barrier(PIPE_V);
+  TADDS(v837, v835, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v839 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v840 = (uint64_t) v264;
+  TASSIGN(v839, v840);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v841 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v842 = (uint64_t) v270;
+  TASSIGN(v841, v842);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v841, v825, v839);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v843 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v844 = (uint64_t) v272;
+  TASSIGN(v843, v844);
+  TROWSUM(v843, v827, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v845 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v846 = (uint64_t) v272;
+  TASSIGN(v845, v846);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v847 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v848 = (uint64_t) v264;
+  TASSIGN(v847, v848);
+  pipe_barrier(PIPE_V);
+  TADDS(v847, v845, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v849 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v850 = (uint64_t) v264;
+  TASSIGN(v849, v850);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v851 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v852 = (uint64_t) v269;
+  TASSIGN(v851, v852);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v851, v827, v849);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v853 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v854 = (uint64_t) v272;
+  TASSIGN(v853, v854);
+  TROWSUM(v853, v829, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v855 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v856 = (uint64_t) v272;
+  TASSIGN(v855, v856);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v857 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v858 = (uint64_t) v264;
+  TASSIGN(v857, v858);
+  pipe_barrier(PIPE_V);
+  TADDS(v857, v855, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v859 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v860 = (uint64_t) v264;
+  TASSIGN(v859, v860);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v861 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v862 = (uint64_t) v268;
+  TASSIGN(v861, v862);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v861, v829, v859);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v863 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v864 = (uint64_t) v272;
+  TASSIGN(v863, v864);
+  TROWSUM(v863, v831, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v865 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v866 = (uint64_t) v272;
+  TASSIGN(v865, v866);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v867 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v868 = (uint64_t) v264;
+  TASSIGN(v867, v868);
+  pipe_barrier(PIPE_V);
+  TADDS(v867, v865, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v869 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v870 = (uint64_t) v264;
+  TASSIGN(v869, v870);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v871 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v872 = (uint64_t) v267;
+  TASSIGN(v871, v872);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v871, v831, v869);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v873 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v874 = (uint64_t) v266;
+  TASSIGN(v873, v874);
+  TADD(v873, v841, v851);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v875 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v876 = (uint64_t) v265;
+  TASSIGN(v875, v876);
+  pipe_barrier(PIPE_V);
+  TADD(v875, v861, v871);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v877 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v878 = (uint64_t) v266;
+  TASSIGN(v877, v878);
+  pipe_barrier(PIPE_V);
+  TADD(v877, v873, v875);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v879 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v880 = (uint64_t) v266;
+  TASSIGN(v879, v880);
+  pipe_barrier(PIPE_V);
+  TADDS(v879, v877, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v881 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v882 = (uint64_t) v270;
+  TASSIGN(v881, v882);
+  pipe_barrier(PIPE_V);
+  TDIV(v881, v841, v879);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v883 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v884 = (uint64_t) v269;
+  TASSIGN(v883, v884);
+  TDIV(v883, v851, v879);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v885 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v886 = (uint64_t) v268;
+  TASSIGN(v885, v886);
+  TDIV(v885, v861, v879);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v887 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v888 = (uint64_t) v267;
+  TASSIGN(v887, v888);
+  TDIV(v887, v871, v879);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v889 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v890 = (uint64_t) v272;
+  TASSIGN(v889, v890);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v889, v881, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v891 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v892 = (uint64_t) v272;
+  TASSIGN(v891, v892);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v893 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v894 = (uint64_t) v264;
+  TASSIGN(v893, v894);
+  pipe_barrier(PIPE_V);
+  TADDS(v893, v891, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v895 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v896 = (uint64_t) v264;
+  TASSIGN(v895, v896);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v897 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v898 = (uint64_t) v270;
+  TASSIGN(v897, v898);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v897, v881, v895);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v899 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v900 = (uint64_t) v272;
+  TASSIGN(v899, v900);
+  TROWSUM(v899, v883, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v901 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v902 = (uint64_t) v272;
+  TASSIGN(v901, v902);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v903 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v904 = (uint64_t) v264;
+  TASSIGN(v903, v904);
+  pipe_barrier(PIPE_V);
+  TADDS(v903, v901, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v905 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v906 = (uint64_t) v264;
+  TASSIGN(v905, v906);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v907 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v908 = (uint64_t) v269;
+  TASSIGN(v907, v908);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v907, v883, v905);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v909 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v910 = (uint64_t) v272;
+  TASSIGN(v909, v910);
+  TROWSUM(v909, v885, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v911 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v912 = (uint64_t) v272;
+  TASSIGN(v911, v912);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v913 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v914 = (uint64_t) v264;
+  TASSIGN(v913, v914);
+  pipe_barrier(PIPE_V);
+  TADDS(v913, v911, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v915 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v916 = (uint64_t) v264;
+  TASSIGN(v915, v916);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v917 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v918 = (uint64_t) v268;
+  TASSIGN(v917, v918);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v917, v885, v915);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v919 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v920 = (uint64_t) v272;
+  TASSIGN(v919, v920);
+  TROWSUM(v919, v887, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v921 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v922 = (uint64_t) v272;
+  TASSIGN(v921, v922);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v923 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v924 = (uint64_t) v264;
+  TASSIGN(v923, v924);
+  pipe_barrier(PIPE_V);
+  TADDS(v923, v921, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v925 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v926 = (uint64_t) v264;
+  TASSIGN(v925, v926);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v927 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v928 = (uint64_t) v267;
+  TASSIGN(v927, v928);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v927, v887, v925);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v929 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v930 = (uint64_t) v266;
+  TASSIGN(v929, v930);
+  TADD(v929, v897, v907);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v931 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v932 = (uint64_t) v265;
+  TASSIGN(v931, v932);
+  pipe_barrier(PIPE_V);
+  TADD(v931, v917, v927);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v933 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v934 = (uint64_t) v266;
+  TASSIGN(v933, v934);
+  pipe_barrier(PIPE_V);
+  TADD(v933, v929, v931);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v935 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v936 = (uint64_t) v266;
+  TASSIGN(v935, v936);
+  pipe_barrier(PIPE_V);
+  TADDS(v935, v933, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v937 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v938 = (uint64_t) v270;
+  TASSIGN(v937, v938);
+  pipe_barrier(PIPE_V);
+  TDIV(v937, v897, v935);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v939 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v940 = (uint64_t) v269;
+  TASSIGN(v939, v940);
+  TDIV(v939, v907, v935);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v941 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v942 = (uint64_t) v268;
+  TASSIGN(v941, v942);
+  TDIV(v941, v917, v935);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v943 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v944 = (uint64_t) v267;
+  TASSIGN(v943, v944);
+  TDIV(v943, v927, v935);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v945 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v946 = (uint64_t) v272;
+  TASSIGN(v945, v946);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v945, v937, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v947 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v948 = (uint64_t) v272;
+  TASSIGN(v947, v948);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v949 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v950 = (uint64_t) v264;
+  TASSIGN(v949, v950);
+  pipe_barrier(PIPE_V);
+  TADDS(v949, v947, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v951 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v952 = (uint64_t) v264;
+  TASSIGN(v951, v952);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v953 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v954 = (uint64_t) v270;
+  TASSIGN(v953, v954);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v953, v937, v951);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v955 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v956 = (uint64_t) v272;
+  TASSIGN(v955, v956);
+  TROWSUM(v955, v939, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v957 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v958 = (uint64_t) v272;
+  TASSIGN(v957, v958);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v959 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v960 = (uint64_t) v264;
+  TASSIGN(v959, v960);
+  pipe_barrier(PIPE_V);
+  TADDS(v959, v957, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v961 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v962 = (uint64_t) v264;
+  TASSIGN(v961, v962);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v963 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v964 = (uint64_t) v269;
+  TASSIGN(v963, v964);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v963, v939, v961);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v965 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v966 = (uint64_t) v272;
+  TASSIGN(v965, v966);
+  TROWSUM(v965, v941, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v967 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v968 = (uint64_t) v272;
+  TASSIGN(v967, v968);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v969 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v970 = (uint64_t) v264;
+  TASSIGN(v969, v970);
+  pipe_barrier(PIPE_V);
+  TADDS(v969, v967, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v971 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v972 = (uint64_t) v264;
+  TASSIGN(v971, v972);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v973 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v974 = (uint64_t) v268;
+  TASSIGN(v973, v974);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v973, v941, v971);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v975 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v976 = (uint64_t) v272;
+  TASSIGN(v975, v976);
+  TROWSUM(v975, v943, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v977 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v978 = (uint64_t) v272;
+  TASSIGN(v977, v978);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v979 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v980 = (uint64_t) v264;
+  TASSIGN(v979, v980);
+  pipe_barrier(PIPE_V);
+  TADDS(v979, v977, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v981 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v982 = (uint64_t) v264;
+  TASSIGN(v981, v982);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v983 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v984 = (uint64_t) v267;
+  TASSIGN(v983, v984);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v983, v943, v981);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v985 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v986 = (uint64_t) v266;
+  TASSIGN(v985, v986);
+  TADD(v985, v953, v963);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v987 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v988 = (uint64_t) v265;
+  TASSIGN(v987, v988);
+  pipe_barrier(PIPE_V);
+  TADD(v987, v973, v983);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v989 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v990 = (uint64_t) v266;
+  TASSIGN(v989, v990);
+  pipe_barrier(PIPE_V);
+  TADD(v989, v985, v987);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v991 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v992 = (uint64_t) v266;
+  TASSIGN(v991, v992);
+  pipe_barrier(PIPE_V);
+  TADDS(v991, v989, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v993 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v994 = (uint64_t) v270;
+  TASSIGN(v993, v994);
+  pipe_barrier(PIPE_V);
+  TDIV(v993, v953, v991);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v995 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v996 = (uint64_t) v269;
+  TASSIGN(v995, v996);
+  TDIV(v995, v963, v991);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v997 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v998 = (uint64_t) v268;
+  TASSIGN(v997, v998);
+  TDIV(v997, v973, v991);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v999 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1000 = (uint64_t) v267;
+  TASSIGN(v999, v1000);
+  TDIV(v999, v983, v991);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1001 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1002 = (uint64_t) v272;
+  TASSIGN(v1001, v1002);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1001, v993, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1003 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1004 = (uint64_t) v272;
+  TASSIGN(v1003, v1004);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1005 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1006 = (uint64_t) v264;
+  TASSIGN(v1005, v1006);
+  pipe_barrier(PIPE_V);
+  TADDS(v1005, v1003, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1007 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1008 = (uint64_t) v264;
+  TASSIGN(v1007, v1008);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1009 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1010 = (uint64_t) v270;
+  TASSIGN(v1009, v1010);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1009, v993, v1007);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1011 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1012 = (uint64_t) v272;
+  TASSIGN(v1011, v1012);
+  TROWSUM(v1011, v995, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1013 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1014 = (uint64_t) v272;
+  TASSIGN(v1013, v1014);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1015 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1016 = (uint64_t) v264;
+  TASSIGN(v1015, v1016);
+  pipe_barrier(PIPE_V);
+  TADDS(v1015, v1013, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1017 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1018 = (uint64_t) v264;
+  TASSIGN(v1017, v1018);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1019 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1020 = (uint64_t) v269;
+  TASSIGN(v1019, v1020);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1019, v995, v1017);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1021 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1022 = (uint64_t) v272;
+  TASSIGN(v1021, v1022);
+  TROWSUM(v1021, v997, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1023 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1024 = (uint64_t) v272;
+  TASSIGN(v1023, v1024);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1025 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1026 = (uint64_t) v264;
+  TASSIGN(v1025, v1026);
+  pipe_barrier(PIPE_V);
+  TADDS(v1025, v1023, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1027 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1028 = (uint64_t) v264;
+  TASSIGN(v1027, v1028);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1029 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1030 = (uint64_t) v268;
+  TASSIGN(v1029, v1030);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1029, v997, v1027);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1031 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1032 = (uint64_t) v272;
+  TASSIGN(v1031, v1032);
+  TROWSUM(v1031, v999, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1033 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1034 = (uint64_t) v272;
+  TASSIGN(v1033, v1034);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1035 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1036 = (uint64_t) v264;
+  TASSIGN(v1035, v1036);
+  pipe_barrier(PIPE_V);
+  TADDS(v1035, v1033, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1037 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1038 = (uint64_t) v264;
+  TASSIGN(v1037, v1038);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1039 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1040 = (uint64_t) v267;
+  TASSIGN(v1039, v1040);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1039, v999, v1037);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1041 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1042 = (uint64_t) v266;
+  TASSIGN(v1041, v1042);
+  TADD(v1041, v1009, v1019);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1043 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1044 = (uint64_t) v265;
+  TASSIGN(v1043, v1044);
+  pipe_barrier(PIPE_V);
+  TADD(v1043, v1029, v1039);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1045 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1046 = (uint64_t) v266;
+  TASSIGN(v1045, v1046);
+  pipe_barrier(PIPE_V);
+  TADD(v1045, v1041, v1043);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1047 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1048 = (uint64_t) v266;
+  TASSIGN(v1047, v1048);
+  pipe_barrier(PIPE_V);
+  TADDS(v1047, v1045, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1049 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1050 = (uint64_t) v270;
+  TASSIGN(v1049, v1050);
+  pipe_barrier(PIPE_V);
+  TDIV(v1049, v1009, v1047);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1051 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1052 = (uint64_t) v269;
+  TASSIGN(v1051, v1052);
+  TDIV(v1051, v1019, v1047);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1053 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1054 = (uint64_t) v268;
+  TASSIGN(v1053, v1054);
+  TDIV(v1053, v1029, v1047);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1055 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1056 = (uint64_t) v267;
+  TASSIGN(v1055, v1056);
+  TDIV(v1055, v1039, v1047);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1057 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1058 = (uint64_t) v272;
+  TASSIGN(v1057, v1058);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1057, v1049, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1059 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1060 = (uint64_t) v272;
+  TASSIGN(v1059, v1060);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1061 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1062 = (uint64_t) v264;
+  TASSIGN(v1061, v1062);
+  pipe_barrier(PIPE_V);
+  TADDS(v1061, v1059, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1063 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1064 = (uint64_t) v264;
+  TASSIGN(v1063, v1064);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1065 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1066 = (uint64_t) v270;
+  TASSIGN(v1065, v1066);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1065, v1049, v1063);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1067 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1068 = (uint64_t) v272;
+  TASSIGN(v1067, v1068);
+  TROWSUM(v1067, v1051, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1069 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1070 = (uint64_t) v272;
+  TASSIGN(v1069, v1070);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1071 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1072 = (uint64_t) v264;
+  TASSIGN(v1071, v1072);
+  pipe_barrier(PIPE_V);
+  TADDS(v1071, v1069, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1073 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1074 = (uint64_t) v264;
+  TASSIGN(v1073, v1074);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1075 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1076 = (uint64_t) v269;
+  TASSIGN(v1075, v1076);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1075, v1051, v1073);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1077 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1078 = (uint64_t) v272;
+  TASSIGN(v1077, v1078);
+  TROWSUM(v1077, v1053, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1079 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1080 = (uint64_t) v272;
+  TASSIGN(v1079, v1080);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1081 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1082 = (uint64_t) v264;
+  TASSIGN(v1081, v1082);
+  pipe_barrier(PIPE_V);
+  TADDS(v1081, v1079, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1083 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1084 = (uint64_t) v264;
+  TASSIGN(v1083, v1084);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1085 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1086 = (uint64_t) v268;
+  TASSIGN(v1085, v1086);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1085, v1053, v1083);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1087 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1088 = (uint64_t) v272;
+  TASSIGN(v1087, v1088);
+  TROWSUM(v1087, v1055, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1089 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1090 = (uint64_t) v272;
+  TASSIGN(v1089, v1090);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1091 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1092 = (uint64_t) v264;
+  TASSIGN(v1091, v1092);
+  pipe_barrier(PIPE_V);
+  TADDS(v1091, v1089, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1093 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1094 = (uint64_t) v264;
+  TASSIGN(v1093, v1094);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1095 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1096 = (uint64_t) v267;
+  TASSIGN(v1095, v1096);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1095, v1055, v1093);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1097 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1098 = (uint64_t) v266;
+  TASSIGN(v1097, v1098);
+  TADD(v1097, v1065, v1075);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1099 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1100 = (uint64_t) v265;
+  TASSIGN(v1099, v1100);
+  pipe_barrier(PIPE_V);
+  TADD(v1099, v1085, v1095);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1101 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1102 = (uint64_t) v266;
+  TASSIGN(v1101, v1102);
+  pipe_barrier(PIPE_V);
+  TADD(v1101, v1097, v1099);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1103 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1104 = (uint64_t) v266;
+  TASSIGN(v1103, v1104);
+  pipe_barrier(PIPE_V);
+  TADDS(v1103, v1101, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1105 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1106 = (uint64_t) v270;
+  TASSIGN(v1105, v1106);
+  pipe_barrier(PIPE_V);
+  TDIV(v1105, v1065, v1103);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1107 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1108 = (uint64_t) v269;
+  TASSIGN(v1107, v1108);
+  TDIV(v1107, v1075, v1103);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1109 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1110 = (uint64_t) v268;
+  TASSIGN(v1109, v1110);
+  TDIV(v1109, v1085, v1103);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1111 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1112 = (uint64_t) v267;
+  TASSIGN(v1111, v1112);
+  TDIV(v1111, v1095, v1103);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1113 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1114 = (uint64_t) v272;
+  TASSIGN(v1113, v1114);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1113, v1105, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1115 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1116 = (uint64_t) v272;
+  TASSIGN(v1115, v1116);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1117 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1118 = (uint64_t) v264;
+  TASSIGN(v1117, v1118);
+  pipe_barrier(PIPE_V);
+  TADDS(v1117, v1115, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1119 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1120 = (uint64_t) v264;
+  TASSIGN(v1119, v1120);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1121 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1122 = (uint64_t) v270;
+  TASSIGN(v1121, v1122);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1121, v1105, v1119);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1123 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1124 = (uint64_t) v272;
+  TASSIGN(v1123, v1124);
+  TROWSUM(v1123, v1107, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1125 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1126 = (uint64_t) v272;
+  TASSIGN(v1125, v1126);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1127 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1128 = (uint64_t) v264;
+  TASSIGN(v1127, v1128);
+  pipe_barrier(PIPE_V);
+  TADDS(v1127, v1125, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1129 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1130 = (uint64_t) v264;
+  TASSIGN(v1129, v1130);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1131 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1132 = (uint64_t) v269;
+  TASSIGN(v1131, v1132);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1131, v1107, v1129);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1133 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1134 = (uint64_t) v272;
+  TASSIGN(v1133, v1134);
+  TROWSUM(v1133, v1109, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1135 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1136 = (uint64_t) v272;
+  TASSIGN(v1135, v1136);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1137 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1138 = (uint64_t) v264;
+  TASSIGN(v1137, v1138);
+  pipe_barrier(PIPE_V);
+  TADDS(v1137, v1135, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1139 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1140 = (uint64_t) v264;
+  TASSIGN(v1139, v1140);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1141 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1142 = (uint64_t) v268;
+  TASSIGN(v1141, v1142);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1141, v1109, v1139);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1143 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1144 = (uint64_t) v272;
+  TASSIGN(v1143, v1144);
+  TROWSUM(v1143, v1111, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1145 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1146 = (uint64_t) v272;
+  TASSIGN(v1145, v1146);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1147 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1148 = (uint64_t) v264;
+  TASSIGN(v1147, v1148);
+  pipe_barrier(PIPE_V);
+  TADDS(v1147, v1145, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1149 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1150 = (uint64_t) v264;
+  TASSIGN(v1149, v1150);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1151 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1152 = (uint64_t) v267;
+  TASSIGN(v1151, v1152);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1151, v1111, v1149);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1153 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1154 = (uint64_t) v266;
+  TASSIGN(v1153, v1154);
+  TADD(v1153, v1121, v1131);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1155 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1156 = (uint64_t) v265;
+  TASSIGN(v1155, v1156);
+  pipe_barrier(PIPE_V);
+  TADD(v1155, v1141, v1151);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1157 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1158 = (uint64_t) v266;
+  TASSIGN(v1157, v1158);
+  pipe_barrier(PIPE_V);
+  TADD(v1157, v1153, v1155);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1159 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1160 = (uint64_t) v266;
+  TASSIGN(v1159, v1160);
+  pipe_barrier(PIPE_V);
+  TADDS(v1159, v1157, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1161 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1162 = (uint64_t) v270;
+  TASSIGN(v1161, v1162);
+  pipe_barrier(PIPE_V);
+  TDIV(v1161, v1121, v1159);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1163 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1164 = (uint64_t) v269;
+  TASSIGN(v1163, v1164);
+  TDIV(v1163, v1131, v1159);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1165 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1166 = (uint64_t) v268;
+  TASSIGN(v1165, v1166);
+  TDIV(v1165, v1141, v1159);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1167 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1168 = (uint64_t) v267;
+  TASSIGN(v1167, v1168);
+  TDIV(v1167, v1151, v1159);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1169 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1170 = (uint64_t) v272;
+  TASSIGN(v1169, v1170);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1169, v1161, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1171 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1172 = (uint64_t) v272;
+  TASSIGN(v1171, v1172);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1173 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1174 = (uint64_t) v264;
+  TASSIGN(v1173, v1174);
+  pipe_barrier(PIPE_V);
+  TADDS(v1173, v1171, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1175 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1176 = (uint64_t) v264;
+  TASSIGN(v1175, v1176);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1177 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1178 = (uint64_t) v270;
+  TASSIGN(v1177, v1178);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1177, v1161, v1175);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1179 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1180 = (uint64_t) v272;
+  TASSIGN(v1179, v1180);
+  TROWSUM(v1179, v1163, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1181 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1182 = (uint64_t) v272;
+  TASSIGN(v1181, v1182);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1183 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1184 = (uint64_t) v264;
+  TASSIGN(v1183, v1184);
+  pipe_barrier(PIPE_V);
+  TADDS(v1183, v1181, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1185 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1186 = (uint64_t) v264;
+  TASSIGN(v1185, v1186);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1187 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1188 = (uint64_t) v269;
+  TASSIGN(v1187, v1188);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1187, v1163, v1185);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1189 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1190 = (uint64_t) v272;
+  TASSIGN(v1189, v1190);
+  TROWSUM(v1189, v1165, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1191 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1192 = (uint64_t) v272;
+  TASSIGN(v1191, v1192);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1193 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1194 = (uint64_t) v264;
+  TASSIGN(v1193, v1194);
+  pipe_barrier(PIPE_V);
+  TADDS(v1193, v1191, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1195 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1196 = (uint64_t) v264;
+  TASSIGN(v1195, v1196);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1197 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1198 = (uint64_t) v268;
+  TASSIGN(v1197, v1198);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1197, v1165, v1195);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1199 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1200 = (uint64_t) v272;
+  TASSIGN(v1199, v1200);
+  TROWSUM(v1199, v1167, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1201 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1202 = (uint64_t) v272;
+  TASSIGN(v1201, v1202);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1203 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1204 = (uint64_t) v264;
+  TASSIGN(v1203, v1204);
+  pipe_barrier(PIPE_V);
+  TADDS(v1203, v1201, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1205 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1206 = (uint64_t) v264;
+  TASSIGN(v1205, v1206);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1207 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1208 = (uint64_t) v267;
+  TASSIGN(v1207, v1208);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1207, v1167, v1205);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1209 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1210 = (uint64_t) v266;
+  TASSIGN(v1209, v1210);
+  TADD(v1209, v1177, v1187);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1211 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1212 = (uint64_t) v265;
+  TASSIGN(v1211, v1212);
+  pipe_barrier(PIPE_V);
+  TADD(v1211, v1197, v1207);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1213 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1214 = (uint64_t) v266;
+  TASSIGN(v1213, v1214);
+  pipe_barrier(PIPE_V);
+  TADD(v1213, v1209, v1211);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1215 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1216 = (uint64_t) v266;
+  TASSIGN(v1215, v1216);
+  pipe_barrier(PIPE_V);
+  TADDS(v1215, v1213, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1217 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1218 = (uint64_t) v270;
+  TASSIGN(v1217, v1218);
+  pipe_barrier(PIPE_V);
+  TDIV(v1217, v1177, v1215);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1219 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1220 = (uint64_t) v269;
+  TASSIGN(v1219, v1220);
+  TDIV(v1219, v1187, v1215);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1221 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1222 = (uint64_t) v268;
+  TASSIGN(v1221, v1222);
+  TDIV(v1221, v1197, v1215);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1223 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1224 = (uint64_t) v267;
+  TASSIGN(v1223, v1224);
+  TDIV(v1223, v1207, v1215);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1225 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1226 = (uint64_t) v272;
+  TASSIGN(v1225, v1226);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1225, v1217, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1227 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1228 = (uint64_t) v272;
+  TASSIGN(v1227, v1228);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1229 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1230 = (uint64_t) v264;
+  TASSIGN(v1229, v1230);
+  pipe_barrier(PIPE_V);
+  TADDS(v1229, v1227, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1231 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1232 = (uint64_t) v264;
+  TASSIGN(v1231, v1232);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1233 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1234 = (uint64_t) v270;
+  TASSIGN(v1233, v1234);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1233, v1217, v1231);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1235 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1236 = (uint64_t) v272;
+  TASSIGN(v1235, v1236);
+  TROWSUM(v1235, v1219, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1237 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1238 = (uint64_t) v272;
+  TASSIGN(v1237, v1238);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1239 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1240 = (uint64_t) v264;
+  TASSIGN(v1239, v1240);
+  pipe_barrier(PIPE_V);
+  TADDS(v1239, v1237, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1241 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1242 = (uint64_t) v264;
+  TASSIGN(v1241, v1242);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1243 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1244 = (uint64_t) v269;
+  TASSIGN(v1243, v1244);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1243, v1219, v1241);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1245 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1246 = (uint64_t) v272;
+  TASSIGN(v1245, v1246);
+  TROWSUM(v1245, v1221, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1247 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1248 = (uint64_t) v272;
+  TASSIGN(v1247, v1248);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1249 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1250 = (uint64_t) v264;
+  TASSIGN(v1249, v1250);
+  pipe_barrier(PIPE_V);
+  TADDS(v1249, v1247, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1251 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1252 = (uint64_t) v264;
+  TASSIGN(v1251, v1252);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1253 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1254 = (uint64_t) v268;
+  TASSIGN(v1253, v1254);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1253, v1221, v1251);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1255 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1256 = (uint64_t) v272;
+  TASSIGN(v1255, v1256);
+  TROWSUM(v1255, v1223, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1257 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1258 = (uint64_t) v272;
+  TASSIGN(v1257, v1258);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1259 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1260 = (uint64_t) v264;
+  TASSIGN(v1259, v1260);
+  pipe_barrier(PIPE_V);
+  TADDS(v1259, v1257, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1261 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1262 = (uint64_t) v264;
+  TASSIGN(v1261, v1262);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1263 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1264 = (uint64_t) v267;
+  TASSIGN(v1263, v1264);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1263, v1223, v1261);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1265 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1266 = (uint64_t) v266;
+  TASSIGN(v1265, v1266);
+  TADD(v1265, v1233, v1243);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1267 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1268 = (uint64_t) v265;
+  TASSIGN(v1267, v1268);
+  pipe_barrier(PIPE_V);
+  TADD(v1267, v1253, v1263);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1269 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1270 = (uint64_t) v266;
+  TASSIGN(v1269, v1270);
+  pipe_barrier(PIPE_V);
+  TADD(v1269, v1265, v1267);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1271 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1272 = (uint64_t) v266;
+  TASSIGN(v1271, v1272);
+  pipe_barrier(PIPE_V);
+  TADDS(v1271, v1269, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1273 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1274 = (uint64_t) v270;
+  TASSIGN(v1273, v1274);
+  pipe_barrier(PIPE_V);
+  TDIV(v1273, v1233, v1271);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1275 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1276 = (uint64_t) v269;
+  TASSIGN(v1275, v1276);
+  TDIV(v1275, v1243, v1271);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1277 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1278 = (uint64_t) v268;
+  TASSIGN(v1277, v1278);
+  TDIV(v1277, v1253, v1271);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1279 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1280 = (uint64_t) v267;
+  TASSIGN(v1279, v1280);
+  TDIV(v1279, v1263, v1271);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1281 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1282 = (uint64_t) v272;
+  TASSIGN(v1281, v1282);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1281, v1273, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1283 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1284 = (uint64_t) v272;
+  TASSIGN(v1283, v1284);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1285 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1286 = (uint64_t) v264;
+  TASSIGN(v1285, v1286);
+  pipe_barrier(PIPE_V);
+  TADDS(v1285, v1283, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1287 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1288 = (uint64_t) v264;
+  TASSIGN(v1287, v1288);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1289 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1290 = (uint64_t) v270;
+  TASSIGN(v1289, v1290);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1289, v1273, v1287);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1291 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1292 = (uint64_t) v272;
+  TASSIGN(v1291, v1292);
+  TROWSUM(v1291, v1275, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1293 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1294 = (uint64_t) v272;
+  TASSIGN(v1293, v1294);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1295 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1296 = (uint64_t) v264;
+  TASSIGN(v1295, v1296);
+  pipe_barrier(PIPE_V);
+  TADDS(v1295, v1293, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1297 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1298 = (uint64_t) v264;
+  TASSIGN(v1297, v1298);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1299 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1300 = (uint64_t) v269;
+  TASSIGN(v1299, v1300);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1299, v1275, v1297);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1301 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1302 = (uint64_t) v272;
+  TASSIGN(v1301, v1302);
+  TROWSUM(v1301, v1277, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1303 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1304 = (uint64_t) v272;
+  TASSIGN(v1303, v1304);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1305 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1306 = (uint64_t) v264;
+  TASSIGN(v1305, v1306);
+  pipe_barrier(PIPE_V);
+  TADDS(v1305, v1303, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1307 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1308 = (uint64_t) v264;
+  TASSIGN(v1307, v1308);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1309 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1310 = (uint64_t) v268;
+  TASSIGN(v1309, v1310);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1309, v1277, v1307);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1311 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1312 = (uint64_t) v272;
+  TASSIGN(v1311, v1312);
+  TROWSUM(v1311, v1279, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1313 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1314 = (uint64_t) v272;
+  TASSIGN(v1313, v1314);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1315 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1316 = (uint64_t) v264;
+  TASSIGN(v1315, v1316);
+  pipe_barrier(PIPE_V);
+  TADDS(v1315, v1313, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1317 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1318 = (uint64_t) v264;
+  TASSIGN(v1317, v1318);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1319 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1320 = (uint64_t) v267;
+  TASSIGN(v1319, v1320);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1319, v1279, v1317);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1321 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1322 = (uint64_t) v266;
+  TASSIGN(v1321, v1322);
+  TADD(v1321, v1289, v1299);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1323 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1324 = (uint64_t) v265;
+  TASSIGN(v1323, v1324);
+  pipe_barrier(PIPE_V);
+  TADD(v1323, v1309, v1319);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1325 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1326 = (uint64_t) v266;
+  TASSIGN(v1325, v1326);
+  pipe_barrier(PIPE_V);
+  TADD(v1325, v1321, v1323);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1327 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1328 = (uint64_t) v266;
+  TASSIGN(v1327, v1328);
+  pipe_barrier(PIPE_V);
+  TADDS(v1327, v1325, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1329 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1330 = (uint64_t) v270;
+  TASSIGN(v1329, v1330);
+  pipe_barrier(PIPE_V);
+  TDIV(v1329, v1289, v1327);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1331 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1332 = (uint64_t) v269;
+  TASSIGN(v1331, v1332);
+  TDIV(v1331, v1299, v1327);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1333 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1334 = (uint64_t) v268;
+  TASSIGN(v1333, v1334);
+  TDIV(v1333, v1309, v1327);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1335 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1336 = (uint64_t) v267;
+  TASSIGN(v1335, v1336);
+  TDIV(v1335, v1319, v1327);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1337 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1338 = (uint64_t) v272;
+  TASSIGN(v1337, v1338);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1337, v1329, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1339 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1340 = (uint64_t) v272;
+  TASSIGN(v1339, v1340);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1341 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1342 = (uint64_t) v264;
+  TASSIGN(v1341, v1342);
+  pipe_barrier(PIPE_V);
+  TADDS(v1341, v1339, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1343 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1344 = (uint64_t) v264;
+  TASSIGN(v1343, v1344);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1345 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1346 = (uint64_t) v270;
+  TASSIGN(v1345, v1346);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1345, v1329, v1343);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1347 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1348 = (uint64_t) v272;
+  TASSIGN(v1347, v1348);
+  TROWSUM(v1347, v1331, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1349 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1350 = (uint64_t) v272;
+  TASSIGN(v1349, v1350);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1351 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1352 = (uint64_t) v264;
+  TASSIGN(v1351, v1352);
+  pipe_barrier(PIPE_V);
+  TADDS(v1351, v1349, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1353 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1354 = (uint64_t) v264;
+  TASSIGN(v1353, v1354);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1355 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1356 = (uint64_t) v269;
+  TASSIGN(v1355, v1356);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1355, v1331, v1353);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1357 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1358 = (uint64_t) v272;
+  TASSIGN(v1357, v1358);
+  TROWSUM(v1357, v1333, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1359 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1360 = (uint64_t) v272;
+  TASSIGN(v1359, v1360);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1361 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1362 = (uint64_t) v264;
+  TASSIGN(v1361, v1362);
+  pipe_barrier(PIPE_V);
+  TADDS(v1361, v1359, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1363 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1364 = (uint64_t) v264;
+  TASSIGN(v1363, v1364);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1365 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1366 = (uint64_t) v268;
+  TASSIGN(v1365, v1366);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1365, v1333, v1363);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1367 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1368 = (uint64_t) v272;
+  TASSIGN(v1367, v1368);
+  TROWSUM(v1367, v1335, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1369 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1370 = (uint64_t) v272;
+  TASSIGN(v1369, v1370);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1371 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1372 = (uint64_t) v264;
+  TASSIGN(v1371, v1372);
+  pipe_barrier(PIPE_V);
+  TADDS(v1371, v1369, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1373 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1374 = (uint64_t) v264;
+  TASSIGN(v1373, v1374);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1375 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1376 = (uint64_t) v267;
+  TASSIGN(v1375, v1376);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1375, v1335, v1373);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1377 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1378 = (uint64_t) v266;
+  TASSIGN(v1377, v1378);
+  TADD(v1377, v1345, v1355);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1379 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1380 = (uint64_t) v265;
+  TASSIGN(v1379, v1380);
+  pipe_barrier(PIPE_V);
+  TADD(v1379, v1365, v1375);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1381 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1382 = (uint64_t) v266;
+  TASSIGN(v1381, v1382);
+  pipe_barrier(PIPE_V);
+  TADD(v1381, v1377, v1379);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1383 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1384 = (uint64_t) v266;
+  TASSIGN(v1383, v1384);
+  pipe_barrier(PIPE_V);
+  TADDS(v1383, v1381, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1385 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1386 = (uint64_t) v270;
+  TASSIGN(v1385, v1386);
+  pipe_barrier(PIPE_V);
+  TDIV(v1385, v1345, v1383);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1387 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1388 = (uint64_t) v269;
+  TASSIGN(v1387, v1388);
+  TDIV(v1387, v1355, v1383);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1389 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1390 = (uint64_t) v268;
+  TASSIGN(v1389, v1390);
+  TDIV(v1389, v1365, v1383);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1391 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1392 = (uint64_t) v267;
+  TASSIGN(v1391, v1392);
+  TDIV(v1391, v1375, v1383);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1393 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1394 = (uint64_t) v272;
+  TASSIGN(v1393, v1394);
+  pipe_barrier(PIPE_V);
+  TROWSUM(v1393, v1385, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1395 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1396 = (uint64_t) v272;
+  TASSIGN(v1395, v1396);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1397 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1398 = (uint64_t) v264;
+  TASSIGN(v1397, v1398);
+  pipe_barrier(PIPE_V);
+  TADDS(v1397, v1395, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1399 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1400 = (uint64_t) v264;
+  TASSIGN(v1399, v1400);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1401 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1402 = (uint64_t) v270;
+  TASSIGN(v1401, v1402);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1401, v1385, v1399);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1403 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1404 = (uint64_t) v272;
+  TASSIGN(v1403, v1404);
+  TROWSUM(v1403, v1387, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1405 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1406 = (uint64_t) v272;
+  TASSIGN(v1405, v1406);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1407 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1408 = (uint64_t) v264;
+  TASSIGN(v1407, v1408);
+  pipe_barrier(PIPE_V);
+  TADDS(v1407, v1405, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1409 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1410 = (uint64_t) v264;
+  TASSIGN(v1409, v1410);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1411 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1412 = (uint64_t) v269;
+  TASSIGN(v1411, v1412);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1411, v1387, v1409);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1413 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1414 = (uint64_t) v272;
+  TASSIGN(v1413, v1414);
+  TROWSUM(v1413, v1389, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1415 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1416 = (uint64_t) v272;
+  TASSIGN(v1415, v1416);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1417 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1418 = (uint64_t) v264;
+  TASSIGN(v1417, v1418);
+  pipe_barrier(PIPE_V);
+  TADDS(v1417, v1415, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1419 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1420 = (uint64_t) v264;
+  TASSIGN(v1419, v1420);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1421 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1422 = (uint64_t) v268;
+  TASSIGN(v1421, v1422);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1421, v1389, v1419);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1423 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1424 = (uint64_t) v273;
+  TASSIGN(v1423, v1424);
+  TROWSUM(v1423, v1391, v367);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1425 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1426 = (uint64_t) v273;
+  TASSIGN(v1425, v1426);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1427 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v262, v263);
+  uint64_t v1428 = (uint64_t) v264;
+  TASSIGN(v1427, v1428);
+  pipe_barrier(PIPE_V);
+  TADDS(v1427, v1425, v257);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v1429 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v263, v262);
+  uint64_t v1430 = (uint64_t) v264;
+  TASSIGN(v1429, v1430);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1431 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1432 = (uint64_t) v267;
+  TASSIGN(v1431, v1432);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v1431, v1391, v1429);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1433 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1434 = (uint64_t) v266;
+  TASSIGN(v1433, v1434);
+  TADD(v1433, v1401, v1411);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1435 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1436 = (uint64_t) v265;
+  TASSIGN(v1435, v1436);
+  pipe_barrier(PIPE_V);
+  TADD(v1435, v1421, v1431);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1437 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1438 = (uint64_t) v266;
+  TASSIGN(v1437, v1438);
+  pipe_barrier(PIPE_V);
+  TADD(v1437, v1433, v1435);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1439 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1440 = (uint64_t) v266;
+  TASSIGN(v1439, v1440);
+  pipe_barrier(PIPE_V);
+  TADDS(v1439, v1437, v257);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1441 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1442 = (uint64_t) v270;
+  TASSIGN(v1441, v1442);
+  pipe_barrier(PIPE_V);
+  TDIV(v1441, v1401, v1439);
+  set_flag(PIPE_V, PIPE_S, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1443 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1444 = (uint64_t) v269;
+  TASSIGN(v1443, v1444);
+  TDIV(v1443, v1411, v1439);
+  set_flag(PIPE_V, PIPE_S, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1445 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1446 = (uint64_t) v268;
+  TASSIGN(v1445, v1446);
+  TDIV(v1445, v1421, v1439);
+  set_flag(PIPE_V, PIPE_S, EVENT_ID2);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v1447 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v263, v259);
+  uint64_t v1448 = (uint64_t) v267;
+  TASSIGN(v1447, v1448);
+  TDIV(v1447, v1431, v1439);
+  set_flag(PIPE_V, PIPE_S, EVENT_ID3);
+  wait_flag(PIPE_V, PIPE_S, EVENT_ID0);
+  float v1449 = v1441.GetValue(v260);
+  v2[v260] = v1449;
+  wait_flag(PIPE_V, PIPE_S, EVENT_ID1);
+  float v1450 = v1443.GetValue(v260);
+  v2[v261] = v1450;
+  wait_flag(PIPE_V, PIPE_S, EVENT_ID2);
+  float v1451 = v1445.GetValue(v260);
+  v2[v259] = v1451;
+  wait_flag(PIPE_V, PIPE_S, EVENT_ID3);
+  float v1452 = v1447.GetValue(v260);
+  v2[v258] = v1452;
+  float v1453 = v1441.GetValue(v262);
+  v2[v262] = v1453;
+  float v1454 = v1443.GetValue(v262);
+  v2[v256] = v1454;
+  float v1455 = v1445.GetValue(v262);
+  v2[v255] = v1455;
+  float v1456 = v1447.GetValue(v262);
+  v2[v254] = v1456;
+  float v1457 = v1441.GetValue(v253);
+  v2[v253] = v1457;
+  float v1458 = v1443.GetValue(v253);
+  v2[v252] = v1458;
+  float v1459 = v1445.GetValue(v253);
+  v2[v251] = v1459;
+  float v1460 = v1447.GetValue(v253);
+  v2[v250] = v1460;
+  float v1461 = v1441.GetValue(v249);
+  v2[v249] = v1461;
+  float v1462 = v1443.GetValue(v249);
+  v2[v248] = v1462;
+  float v1463 = v1445.GetValue(v249);
+  v2[v247] = v1463;
+  float v1464 = v1447.GetValue(v249);
+  v2[v246] = v1464;
+  float v1465 = v1441.GetValue(v259);
+  v2[v263] = v1465;
+  float v1466 = v1443.GetValue(v259);
+  v2[v245] = v1466;
+  float v1467 = v1445.GetValue(v259);
+  v2[v244] = v1467;
+  float v1468 = v1447.GetValue(v259);
+  v2[v243] = v1468;
+  float v1469 = v1441.GetValue(v255);
+  v2[v242] = v1469;
+  float v1470 = v1443.GetValue(v255);
+  v2[v241] = v1470;
+  float v1471 = v1445.GetValue(v255);
+  v2[v240] = v1471;
+  float v1472 = v1447.GetValue(v255);
+  v2[v239] = v1472;
+  float v1473 = v1441.GetValue(v251);
+  v2[v238] = v1473;
+  float v1474 = v1443.GetValue(v251);
+  v2[v237] = v1474;
+  float v1475 = v1445.GetValue(v251);
+  v2[v236] = v1475;
+  float v1476 = v1447.GetValue(v251);
+  v2[v235] = v1476;
+  float v1477 = v1441.GetValue(v247);
+  v2[v234] = v1477;
+  float v1478 = v1443.GetValue(v247);
+  v2[v233] = v1478;
+  float v1479 = v1445.GetValue(v247);
+  v2[v232] = v1479;
+  float v1480 = v1447.GetValue(v247);
+  v2[v231] = v1480;
+  float v1481 = v1441.GetValue(v263);
+  v2[v230] = v1481;
+  float v1482 = v1443.GetValue(v263);
+  v2[v229] = v1482;
+  float v1483 = v1445.GetValue(v263);
+  v2[v228] = v1483;
+  float v1484 = v1447.GetValue(v263);
+  v2[v227] = v1484;
+  float v1485 = v1441.GetValue(v242);
+  v2[v226] = v1485;
+  float v1486 = v1443.GetValue(v242);
+  v2[v225] = v1486;
+  float v1487 = v1445.GetValue(v242);
+  v2[v224] = v1487;
+  float v1488 = v1447.GetValue(v242);
+  v2[v223] = v1488;
+  float v1489 = v1441.GetValue(v238);
+  v2[v222] = v1489;
+  float v1490 = v1443.GetValue(v238);
+  v2[v221] = v1490;
+  float v1491 = v1445.GetValue(v238);
+  v2[v220] = v1491;
+  float v1492 = v1447.GetValue(v238);
+  v2[v219] = v1492;
+  float v1493 = v1441.GetValue(v234);
+  v2[v218] = v1493;
+  float v1494 = v1443.GetValue(v234);
+  v2[v217] = v1494;
+  float v1495 = v1445.GetValue(v234);
+  v2[v216] = v1495;
+  float v1496 = v1447.GetValue(v234);
+  v2[v215] = v1496;
+  float v1497 = v1441.GetValue(v244);
+  v2[v214] = v1497;
+  float v1498 = v1443.GetValue(v244);
+  v2[v213] = v1498;
+  float v1499 = v1445.GetValue(v244);
+  v2[v212] = v1499;
+  float v1500 = v1447.GetValue(v244);
+  v2[v211] = v1500;
+  float v1501 = v1441.GetValue(v240);
+  v2[v210] = v1501;
+  float v1502 = v1443.GetValue(v240);
+  v2[v209] = v1502;
+  float v1503 = v1445.GetValue(v240);
+  v2[v208] = v1503;
+  float v1504 = v1447.GetValue(v240);
+  v2[v207] = v1504;
+  float v1505 = v1441.GetValue(v236);
+  v2[v206] = v1505;
+  float v1506 = v1443.GetValue(v236);
+  v2[v205] = v1506;
+  float v1507 = v1445.GetValue(v236);
+  v2[v204] = v1507;
+  float v1508 = v1447.GetValue(v236);
+  v2[v203] = v1508;
+  float v1509 = v1441.GetValue(v232);
+  v2[v202] = v1509;
+  float v1510 = v1443.GetValue(v232);
+  v2[v201] = v1510;
+  float v1511 = v1445.GetValue(v232);
+  v2[v200] = v1511;
+  float v1512 = v1447.GetValue(v232);
+  v2[v199] = v1512;
+  float v1513 = v1441.GetValue(v230);
+  v2[v198] = v1513;
+  float v1514 = v1443.GetValue(v230);
+  v2[v197] = v1514;
+  float v1515 = v1445.GetValue(v230);
+  v2[v196] = v1515;
+  float v1516 = v1447.GetValue(v230);
+  v2[v195] = v1516;
+  float v1517 = v1441.GetValue(v226);
+  v2[v194] = v1517;
+  float v1518 = v1443.GetValue(v226);
+  v2[v193] = v1518;
+  float v1519 = v1445.GetValue(v226);
+  v2[v192] = v1519;
+  float v1520 = v1447.GetValue(v226);
+  v2[v191] = v1520;
+  float v1521 = v1441.GetValue(v222);
+  v2[v190] = v1521;
+  float v1522 = v1443.GetValue(v222);
+  v2[v189] = v1522;
+  float v1523 = v1445.GetValue(v222);
+  v2[v188] = v1523;
+  float v1524 = v1447.GetValue(v222);
+  v2[v187] = v1524;
+  float v1525 = v1441.GetValue(v218);
+  v2[v186] = v1525;
+  float v1526 = v1443.GetValue(v218);
+  v2[v185] = v1526;
+  float v1527 = v1445.GetValue(v218);
+  v2[v184] = v1527;
+  float v1528 = v1447.GetValue(v218);
+  v2[v183] = v1528;
+  float v1529 = v1441.GetValue(v228);
+  v2[v182] = v1529;
+  float v1530 = v1443.GetValue(v228);
+  v2[v181] = v1530;
+  float v1531 = v1445.GetValue(v228);
+  v2[v180] = v1531;
+  float v1532 = v1447.GetValue(v228);
+  v2[v179] = v1532;
+  float v1533 = v1441.GetValue(v224);
+  v2[v178] = v1533;
+  float v1534 = v1443.GetValue(v224);
+  v2[v177] = v1534;
+  float v1535 = v1445.GetValue(v224);
+  v2[v176] = v1535;
+  float v1536 = v1447.GetValue(v224);
+  v2[v175] = v1536;
+  float v1537 = v1441.GetValue(v220);
+  v2[v174] = v1537;
+  float v1538 = v1443.GetValue(v220);
+  v2[v173] = v1538;
+  float v1539 = v1445.GetValue(v220);
+  v2[v172] = v1539;
+  float v1540 = v1447.GetValue(v220);
+  v2[v171] = v1540;
+  float v1541 = v1441.GetValue(v216);
+  v2[v170] = v1541;
+  float v1542 = v1443.GetValue(v216);
+  v2[v169] = v1542;
+  float v1543 = v1445.GetValue(v216);
+  v2[v168] = v1543;
+  float v1544 = v1447.GetValue(v216);
+  v2[v167] = v1544;
+  float v1545 = v1441.GetValue(v214);
+  v2[v166] = v1545;
+  float v1546 = v1443.GetValue(v214);
+  v2[v165] = v1546;
+  float v1547 = v1445.GetValue(v214);
+  v2[v164] = v1547;
+  float v1548 = v1447.GetValue(v214);
+  v2[v163] = v1548;
+  float v1549 = v1441.GetValue(v210);
+  v2[v162] = v1549;
+  float v1550 = v1443.GetValue(v210);
+  v2[v161] = v1550;
+  float v1551 = v1445.GetValue(v210);
+  v2[v160] = v1551;
+  float v1552 = v1447.GetValue(v210);
+  v2[v159] = v1552;
+  float v1553 = v1441.GetValue(v206);
+  v2[v158] = v1553;
+  float v1554 = v1443.GetValue(v206);
+  v2[v157] = v1554;
+  float v1555 = v1445.GetValue(v206);
+  v2[v156] = v1555;
+  float v1556 = v1447.GetValue(v206);
+  v2[v155] = v1556;
+  float v1557 = v1441.GetValue(v202);
+  v2[v154] = v1557;
+  float v1558 = v1443.GetValue(v202);
+  v2[v153] = v1558;
+  float v1559 = v1445.GetValue(v202);
+  v2[v152] = v1559;
+  float v1560 = v1447.GetValue(v202);
+  v2[v151] = v1560;
+  float v1561 = v1441.GetValue(v212);
+  v2[v150] = v1561;
+  float v1562 = v1443.GetValue(v212);
+  v2[v149] = v1562;
+  float v1563 = v1445.GetValue(v212);
+  v2[v148] = v1563;
+  float v1564 = v1447.GetValue(v212);
+  v2[v147] = v1564;
+  float v1565 = v1441.GetValue(v208);
+  v2[v146] = v1565;
+  float v1566 = v1443.GetValue(v208);
+  v2[v145] = v1566;
+  float v1567 = v1445.GetValue(v208);
+  v2[v144] = v1567;
+  float v1568 = v1447.GetValue(v208);
+  v2[v143] = v1568;
+  float v1569 = v1441.GetValue(v204);
+  v2[v142] = v1569;
+  float v1570 = v1443.GetValue(v204);
+  v2[v141] = v1570;
+  float v1571 = v1445.GetValue(v204);
+  v2[v140] = v1571;
+  float v1572 = v1447.GetValue(v204);
+  v2[v139] = v1572;
+  float v1573 = v1441.GetValue(v200);
+  v2[v138] = v1573;
+  float v1574 = v1443.GetValue(v200);
+  v2[v137] = v1574;
+  float v1575 = v1445.GetValue(v200);
+  v2[v136] = v1575;
+  float v1576 = v1447.GetValue(v200);
+  v2[v135] = v1576;
+  float v1577 = v1441.GetValue(v198);
+  v2[v134] = v1577;
+  float v1578 = v1443.GetValue(v198);
+  v2[v133] = v1578;
+  float v1579 = v1445.GetValue(v198);
+  v2[v132] = v1579;
+  float v1580 = v1447.GetValue(v198);
+  v2[v131] = v1580;
+  float v1581 = v1441.GetValue(v194);
+  v2[v130] = v1581;
+  float v1582 = v1443.GetValue(v194);
+  v2[v129] = v1582;
+  float v1583 = v1445.GetValue(v194);
+  v2[v128] = v1583;
+  float v1584 = v1447.GetValue(v194);
+  v2[v127] = v1584;
+  float v1585 = v1441.GetValue(v190);
+  v2[v126] = v1585;
+  float v1586 = v1443.GetValue(v190);
+  v2[v125] = v1586;
+  float v1587 = v1445.GetValue(v190);
+  v2[v124] = v1587;
+  float v1588 = v1447.GetValue(v190);
+  v2[v123] = v1588;
+  float v1589 = v1441.GetValue(v186);
+  v2[v122] = v1589;
+  float v1590 = v1443.GetValue(v186);
+  v2[v121] = v1590;
+  float v1591 = v1445.GetValue(v186);
+  v2[v120] = v1591;
+  float v1592 = v1447.GetValue(v186);
+  v2[v119] = v1592;
+  float v1593 = v1441.GetValue(v196);
+  v2[v118] = v1593;
+  float v1594 = v1443.GetValue(v196);
+  v2[v117] = v1594;
+  float v1595 = v1445.GetValue(v196);
+  v2[v116] = v1595;
+  float v1596 = v1447.GetValue(v196);
+  v2[v115] = v1596;
+  float v1597 = v1441.GetValue(v192);
+  v2[v114] = v1597;
+  float v1598 = v1443.GetValue(v192);
+  v2[v113] = v1598;
+  float v1599 = v1445.GetValue(v192);
+  v2[v112] = v1599;
+  float v1600 = v1447.GetValue(v192);
+  v2[v111] = v1600;
+  float v1601 = v1441.GetValue(v188);
+  v2[v110] = v1601;
+  float v1602 = v1443.GetValue(v188);
+  v2[v109] = v1602;
+  float v1603 = v1445.GetValue(v188);
+  v2[v108] = v1603;
+  float v1604 = v1447.GetValue(v188);
+  v2[v107] = v1604;
+  float v1605 = v1441.GetValue(v184);
+  v2[v106] = v1605;
+  float v1606 = v1443.GetValue(v184);
+  v2[v105] = v1606;
+  float v1607 = v1445.GetValue(v184);
+  v2[v104] = v1607;
+  float v1608 = v1447.GetValue(v184);
+  v2[v103] = v1608;
+  float v1609 = v1441.GetValue(v182);
+  v2[v102] = v1609;
+  float v1610 = v1443.GetValue(v182);
+  v2[v101] = v1610;
+  float v1611 = v1445.GetValue(v182);
+  v2[v100] = v1611;
+  float v1612 = v1447.GetValue(v182);
+  v2[v99] = v1612;
+  float v1613 = v1441.GetValue(v178);
+  v2[v98] = v1613;
+  float v1614 = v1443.GetValue(v178);
+  v2[v97] = v1614;
+  float v1615 = v1445.GetValue(v178);
+  v2[v96] = v1615;
+  float v1616 = v1447.GetValue(v178);
+  v2[v95] = v1616;
+  float v1617 = v1441.GetValue(v174);
+  v2[v94] = v1617;
+  float v1618 = v1443.GetValue(v174);
+  v2[v93] = v1618;
+  float v1619 = v1445.GetValue(v174);
+  v2[v92] = v1619;
+  float v1620 = v1447.GetValue(v174);
+  v2[v91] = v1620;
+  float v1621 = v1441.GetValue(v170);
+  v2[v90] = v1621;
+  float v1622 = v1443.GetValue(v170);
+  v2[v89] = v1622;
+  float v1623 = v1445.GetValue(v170);
+  v2[v88] = v1623;
+  float v1624 = v1447.GetValue(v170);
+  v2[v87] = v1624;
+  float v1625 = v1441.GetValue(v180);
+  v2[v86] = v1625;
+  float v1626 = v1443.GetValue(v180);
+  v2[v85] = v1626;
+  float v1627 = v1445.GetValue(v180);
+  v2[v84] = v1627;
+  float v1628 = v1447.GetValue(v180);
+  v2[v83] = v1628;
+  float v1629 = v1441.GetValue(v176);
+  v2[v82] = v1629;
+  float v1630 = v1443.GetValue(v176);
+  v2[v81] = v1630;
+  float v1631 = v1445.GetValue(v176);
+  v2[v80] = v1631;
+  float v1632 = v1447.GetValue(v176);
+  v2[v79] = v1632;
+  float v1633 = v1441.GetValue(v172);
+  v2[v78] = v1633;
+  float v1634 = v1443.GetValue(v172);
+  v2[v77] = v1634;
+  float v1635 = v1445.GetValue(v172);
+  v2[v76] = v1635;
+  float v1636 = v1447.GetValue(v172);
+  v2[v75] = v1636;
+  float v1637 = v1441.GetValue(v168);
+  v2[v74] = v1637;
+  float v1638 = v1443.GetValue(v168);
+  v2[v73] = v1638;
+  float v1639 = v1445.GetValue(v168);
+  v2[v72] = v1639;
+  float v1640 = v1447.GetValue(v168);
+  v2[v71] = v1640;
+  float v1641 = v1441.GetValue(v166);
+  v2[v70] = v1641;
+  float v1642 = v1443.GetValue(v166);
+  v2[v69] = v1642;
+  float v1643 = v1445.GetValue(v166);
+  v2[v68] = v1643;
+  float v1644 = v1447.GetValue(v166);
+  v2[v67] = v1644;
+  float v1645 = v1441.GetValue(v162);
+  v2[v66] = v1645;
+  float v1646 = v1443.GetValue(v162);
+  v2[v65] = v1646;
+  float v1647 = v1445.GetValue(v162);
+  v2[v64] = v1647;
+  float v1648 = v1447.GetValue(v162);
+  v2[v63] = v1648;
+  float v1649 = v1441.GetValue(v158);
+  v2[v62] = v1649;
+  float v1650 = v1443.GetValue(v158);
+  v2[v61] = v1650;
+  float v1651 = v1445.GetValue(v158);
+  v2[v60] = v1651;
+  float v1652 = v1447.GetValue(v158);
+  v2[v59] = v1652;
+  float v1653 = v1441.GetValue(v154);
+  v2[v58] = v1653;
+  float v1654 = v1443.GetValue(v154);
+  v2[v57] = v1654;
+  float v1655 = v1445.GetValue(v154);
+  v2[v56] = v1655;
+  float v1656 = v1447.GetValue(v154);
+  v2[v55] = v1656;
+  float v1657 = v1441.GetValue(v164);
+  v2[v54] = v1657;
+  float v1658 = v1443.GetValue(v164);
+  v2[v53] = v1658;
+  float v1659 = v1445.GetValue(v164);
+  v2[v52] = v1659;
+  float v1660 = v1447.GetValue(v164);
+  v2[v51] = v1660;
+  float v1661 = v1441.GetValue(v160);
+  v2[v50] = v1661;
+  float v1662 = v1443.GetValue(v160);
+  v2[v49] = v1662;
+  float v1663 = v1445.GetValue(v160);
+  v2[v48] = v1663;
+  float v1664 = v1447.GetValue(v160);
+  v2[v47] = v1664;
+  float v1665 = v1441.GetValue(v156);
+  v2[v46] = v1665;
+  float v1666 = v1443.GetValue(v156);
+  v2[v45] = v1666;
+  float v1667 = v1445.GetValue(v156);
+  v2[v44] = v1667;
+  float v1668 = v1447.GetValue(v156);
+  v2[v43] = v1668;
+  float v1669 = v1441.GetValue(v152);
+  v2[v42] = v1669;
+  float v1670 = v1443.GetValue(v152);
+  v2[v41] = v1670;
+  float v1671 = v1445.GetValue(v152);
+  v2[v40] = v1671;
+  float v1672 = v1447.GetValue(v152);
+  v2[v39] = v1672;
+  float v1673 = v1441.GetValue(v150);
+  v2[v38] = v1673;
+  float v1674 = v1443.GetValue(v150);
+  v2[v37] = v1674;
+  float v1675 = v1445.GetValue(v150);
+  v2[v36] = v1675;
+  float v1676 = v1447.GetValue(v150);
+  v2[v35] = v1676;
+  float v1677 = v1441.GetValue(v146);
+  v2[v34] = v1677;
+  float v1678 = v1443.GetValue(v146);
+  v2[v33] = v1678;
+  float v1679 = v1445.GetValue(v146);
+  v2[v32] = v1679;
+  float v1680 = v1447.GetValue(v146);
+  v2[v31] = v1680;
+  float v1681 = v1441.GetValue(v142);
+  v2[v30] = v1681;
+  float v1682 = v1443.GetValue(v142);
+  v2[v29] = v1682;
+  float v1683 = v1445.GetValue(v142);
+  v2[v28] = v1683;
+  float v1684 = v1447.GetValue(v142);
+  v2[v27] = v1684;
+  float v1685 = v1441.GetValue(v138);
+  v2[v26] = v1685;
+  float v1686 = v1443.GetValue(v138);
+  v2[v25] = v1686;
+  float v1687 = v1445.GetValue(v138);
+  v2[v24] = v1687;
+  float v1688 = v1447.GetValue(v138);
+  v2[v23] = v1688;
+  float v1689 = v1441.GetValue(v148);
+  v2[v22] = v1689;
+  float v1690 = v1443.GetValue(v148);
+  v2[v21] = v1690;
+  float v1691 = v1445.GetValue(v148);
+  v2[v20] = v1691;
+  float v1692 = v1447.GetValue(v148);
+  v2[v19] = v1692;
+  float v1693 = v1441.GetValue(v144);
+  v2[v18] = v1693;
+  float v1694 = v1443.GetValue(v144);
+  v2[v17] = v1694;
+  float v1695 = v1445.GetValue(v144);
+  v2[v16] = v1695;
+  float v1696 = v1447.GetValue(v144);
+  v2[v15] = v1696;
+  float v1697 = v1441.GetValue(v140);
+  v2[v14] = v1697;
+  float v1698 = v1443.GetValue(v140);
+  v2[v13] = v1698;
+  float v1699 = v1445.GetValue(v140);
+  v2[v12] = v1699;
+  float v1700 = v1447.GetValue(v140);
+  v2[v11] = v1700;
+  float v1701 = v1441.GetValue(v136);
+  v2[v10] = v1701;
+  float v1702 = v1443.GetValue(v136);
+  v2[v9] = v1702;
+  float v1703 = v1445.GetValue(v136);
+  v2[v8] = v1703;
+  float v1704 = v1447.GetValue(v136);
+  v2[v7] = v1704;
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: comb_logits_inline19__ssa_v1
+    __gm__ Tensor* comb_logits_inline19__ssa_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* comb_logits_inline19__ssa_v1 = reinterpret_cast<__gm__ float*>(comb_logits_inline19__ssa_v1_tensor->buffer.addr) + comb_logits_inline19__ssa_v1_tensor->start_offset;
+
+    // Unpack tensor: comb_flat_inline37__ssa_v0
+    __gm__ Tensor* comb_flat_inline37__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* comb_flat_inline37__ssa_v0 = reinterpret_cast<__gm__ float*>(comb_flat_inline37__ssa_v0_tensor->buffer.addr) + comb_flat_inline37__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    comb_sinkhorn(comb_logits_inline19__ssa_v1, comb_flat_inline37__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/combine.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/combine.cpp
index 7333af6e7..543bb4770 100644
--- a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/combine.cpp
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/combine.cpp
@@ -60,10 +60,10 @@ using namespace pto;
 
 // Demo dimensions — must match dispatch.cpp / main.py.
 static constexpr int N = 2;
-static constexpr int T = 8;
+static constexpr int T = 16;
 static constexpr int TOPK = 2;
-static constexpr int D = 64;
-static constexpr int L = 4;
+static constexpr int D = 4096;
+static constexpr int L = 8;
 static constexpr int R = 32;
 static constexpr int W_PAD = 8;
 static constexpr int IDX_PAD = 8;
@@ -218,7 +218,7 @@ extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ in
             TLOAD(add_bf_tile, src_g);
             pipe_barrier(PIPE_ALL);
 
-            TCVT(add_fp_tile, add_bf_tile, RoundMode::CAST_ROUND);
+            TCVT(add_fp_tile, add_bf_tile, RoundMode::CAST_RINT);
             pipe_barrier(PIPE_V);
             TADD(acc_tile, acc_tile, add_fp_tile);
             pipe_barrier(PIPE_V);
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/dispatch.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/dispatch.cpp
index 547a759b6..c4d63eb6d 100644
--- a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/dispatch.cpp
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/dispatch.cpp
@@ -67,14 +67,16 @@
 
 using namespace pto;
 
-// Demo dimensions — must match main.py.
+// Demo dimensions — must match main.py. Sized to mirror the production
+// moe_expert decode config (D = hidden_size = 4096, L = N_LOCAL_EXPERTS = 8,
+// T = decode tokens per rank = 16, R = RECV_MAX = 32).
 static constexpr int N = 2;
-static constexpr int T = 8;
+static constexpr int T = 16;
 static constexpr int TOPK = 2;
-static constexpr int D = 64;
-static constexpr int L = 4;
+static constexpr int D = 4096;
+static constexpr int L = 8;
 static constexpr int R = 32;
-static constexpr int N_ROUTES = T * TOPK;  // 16
+static constexpr int N_ROUTES = T * TOPK;  // 32
 
 // Weight payload tile width. The protocol contract is one FP32 weight per
 // (e, slot) — recv_w[L, R] FP32. AIV vector tiles have a hardware minimum
@@ -90,15 +92,15 @@ static constexpr int IDX_PAD = 8;
 
 // Window region byte sizes — mirror *_BYTES in main.py.
 //
-// Layout:
-//   pub_counts[N][N][L]            INT32   (64 B)
+// Layout (sizes for the D=4096 / L=8 / T=16 / R=32 demo config):
+//   pub_counts[N][N][L]            INT32   (256 B)
 //   count_done_sig[N]              INT32   (padded slot, 64 B)
-//   recv_x[L][R][D]                BF16    (16 KB)
-//   recv_w[L][R][W_PAD]            FP32    (4  KB; weight at slot [0], rest = 0)
-//   recv_idx[L][R][IDX_PAD]        INT32   (4  KB; r=t*TOPK+k at slot [0], rest = 0)
+//   recv_x[L][R][D]                BF16    (2 MiB)
+//   recv_w[L][R][W_PAD]            FP32    (8  KB; weight at slot [0], rest = 0)
+//   recv_idx[L][R][IDX_PAD]        INT32   (8  KB; r=t*TOPK+k at slot [0], rest = 0)
 //   data_done_sig[N]               INT32   (padded slot, 64 B)
 // ---- Cross-rank visible regions consumed by combine.cpp ----
-//   routed_y_buf[T][TOPK][D]       BF16    (2  KB demo; combine push destination,
+//   routed_y_buf[T][TOPK][D]       BF16    (256 KB; combine push destination,
 //                                            addressed directly by (t, k))
 //   combine_done_sig[N]            INT32   (padded slot, 64 B)
 //
@@ -106,10 +108,10 @@ static constexpr int IDX_PAD = 8;
 // as a host-backed device tensor via the orch.
 static constexpr int kPubCountsBytes = N * N * L * 4;  // N*N*L INT32
 static constexpr int kSignalBytes = 64;
-static constexpr int kRecvXBytes = L * R * D * 2;  // BF16
+static constexpr int kRecvXBytes = L * R * D * 2;  // BF16, 2 MiB
 static constexpr int kRecvWBytes = L * R * W_PAD * 4;
 static constexpr int kRecvIdxBytes = L * R * IDX_PAD * 4;
-static constexpr int kRoutedYBufBytes = T * TOPK * D * 2;  // BF16
+static constexpr int kRoutedYBufBytes = T * TOPK * D * 2;  // BF16, 256 KiB
 
 static constexpr int kOffPubCounts = 0;
 static constexpr int kOffCountDone = kOffPubCounts + kPubCountsBytes;
@@ -179,7 +181,10 @@ extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ in
     // ------------------------------------------------------------------
     // histogram: scalar histogram + (dst, loc_e)-sorted route table.
     //
-    // Scalar GM reads of indices[t * TOPK + k] are fine on AIV.
+    // Scalar GM reads of indices[t * TOPK + k] are fine on AIV. The upstream
+    // router task (write_route_outputs) flushes its scalar stores to L2 via
+    // dcci(..., CACHELINE_OUT) + dsb at its own tail, so dispatch picks up
+    // fresh data here without needing a reader-side cache invalidate.
     // Bucket each route by (dst, loc_e) and stable-sort so the payload_push
     // cursor matches each peer's src-major slot_offset rule.
     // ------------------------------------------------------------------
@@ -194,10 +199,17 @@ extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ in
     int route_loc_e[N_ROUTES];
     int route_r[N_ROUTES];
 
+    // EP routing policy: route slot k of each token goes to peer
+    // (my_rank + k) % N. The router was JIT'd with EP_WORLD_SIZE=1 and writes
+    // local expert IDs in [0, L); we layer the rank component on top here so
+    // the demo spreads tokens across both ranks instead of pinning everything
+    // to rank 0. A production EP system would read a true global ID from the
+    // router output instead.
     for (int r = 0; r < N_ROUTES; ++r) {
-        int eid = indices[r];
-        int dst = eid / L;
-        int loc_e = eid - dst * L;
+        int t = r / TOPK;
+        int k = r - t * TOPK;
+        int dst = (my_rank + k) % N;
+        int loc_e = indices[r];  // local expert id in [0, L)
         send_counts[dst][loc_e] += 1;
         route_dst[r] = dst;
         route_loc_e[r] = loc_e;
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_gate_up_dequant.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_gate_up_dequant.cpp
new file mode 100644
index 000000000..ce3c89b7b
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_gate_up_dequant.cpp
@@ -0,0 +1,204 @@
+// Kernel Function: exp_gate_up_dequant
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_gate_up_dequant(__gm__ int32_t* v1, __gm__ int32_t* v2, __gm__ float* v3, __gm__ float* v4, __gm__ float* v5, __gm__ float* v6, __gm__ float* v7, int32_t v8, int32_t v9) {
+  RoundMode v10 = RoundMode::CAST_NONE;
+  unsigned v11 = 0;
+  const int32_t v12 = 4096;
+  const int32_t v13 = 256;
+  const int32_t v14 = 16;
+  const int32_t v15 = 1;
+  const int64_t v16 = 18432;
+  const int64_t v17 = 2048;
+  const int64_t v18 = 1024;
+  const int64_t v19 = 0;
+  const int64_t v20 = 67584;
+  const int64_t v21 = 51200;
+  const int64_t v22 = 34816;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v23 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v24 = (uint64_t) v22;
+  TASSIGN(v23, v24);
+  pto::Shape<1, 1, 1, 16, 256> v25 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v26 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v27 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v1 + ((v11 + v11 * (unsigned) v12) + v11 * (unsigned) v13 + v11 * (unsigned) v15), v25, v26);
+  TLOAD(v23, v27);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v29 = (uint64_t) v21;
+  TASSIGN(v28, v29);
+  pto::Shape<1, 1, 1, 16, 256> v30 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v31 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v32 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v2 + ((v11 + v11 * (unsigned) v12) + v11 * (unsigned) v13 + v11 * (unsigned) v15), v30, v31);
+  TLOAD(v28, v32);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v15);
+  uint64_t v34 = (uint64_t) v20;
+  TASSIGN(v33, v34);
+  pto::Shape<1, 1, 1, 16, 1> v35 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v36 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v37 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v5 + (v11 + v11 * (unsigned) v15 + v11 * (unsigned) v14), v35, v36);
+  TLOAD(v33, v37);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v39 = (uint64_t) v22;
+  TASSIGN(v38, v39);
+  Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v40 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v41 = (uint64_t) v21;
+  TASSIGN(v40, v41);
+  Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v42 = Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v13);
+  uint64_t v43 = (uint64_t) v19;
+  TASSIGN(v42, v43);
+  pto::Shape<1, 1, 1, 1, 256> v44 = pto::Shape<1, 1, 1, 1, 256>();
+  pto::Stride<4096, 4096, 4096, 4096, 1> v45 = pto::Stride<4096, 4096, 4096, 4096, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND> v46 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND>(v3 + (v11 + (unsigned) v8 * (unsigned) v12 + (unsigned) v9 * (unsigned) v15), v44, v45);
+  TLOAD(v42, v46);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v13);
+  uint64_t v48 = (uint64_t) v18;
+  TASSIGN(v47, v48);
+  pto::Shape<1, 1, 1, 1, 256> v49 = pto::Shape<1, 1, 1, 1, 256>();
+  pto::Stride<4096, 4096, 4096, 4096, 1> v50 = pto::Stride<4096, 4096, 4096, 4096, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND> v51 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND>(v4 + (v11 + (unsigned) v8 * (unsigned) v12 + (unsigned) v9 * (unsigned) v15), v49, v50);
+  TLOAD(v47, v51);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v53 = (uint64_t) v17;
+  TASSIGN(v52, v53);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v52, v38, v10);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v55 = (uint64_t) v16;
+  TASSIGN(v54, v55);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TCVT(v54, v40, v10);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v57 = (uint64_t) v17;
+  TASSIGN(v56, v57);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  TROWEXPANDMUL(v56, v52, v33);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v59 = (uint64_t) v17;
+  TASSIGN(v58, v59);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  TCOLEXPANDMUL(v58, v56, v42);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v60 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v61 = (uint64_t) v16;
+  TASSIGN(v60, v61);
+  TROWEXPANDMUL(v60, v54, v33);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v13);
+  uint64_t v63 = (uint64_t) v16;
+  TASSIGN(v62, v63);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  TCOLEXPANDMUL(v62, v60, v47);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  pto::Shape<1, 1, 1, 16, 256> v64 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v65 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v66 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v6 + (v11 + v11 * (unsigned) v13 + v11 * (unsigned) v15), v64, v65);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v66, v58);
+  pto::Shape<1, 1, 1, 16, 256> v67 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v68 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v69 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v7 + (v11 + v11 * (unsigned) v13 + v11 * (unsigned) v15), v67, v68);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v69, v62);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: gate_acc_inline69__rv_v2
+    __gm__ Tensor* gate_acc_inline69__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int32_t* gate_acc_inline69__rv_v2 = reinterpret_cast<__gm__ int32_t*>(gate_acc_inline69__rv_v2_tensor->buffer.addr) + gate_acc_inline69__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: up_acc_inline46__rv_v2
+    __gm__ Tensor* up_acc_inline46__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int32_t* up_acc_inline46__rv_v2 = reinterpret_cast<__gm__ int32_t*>(up_acc_inline46__rv_v2_tensor->buffer.addr) + up_acc_inline46__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: expert_w1_scale__ssa_v0
+    __gm__ Tensor* expert_w1_scale__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* expert_w1_scale__ssa_v0 = reinterpret_cast<__gm__ float*>(expert_w1_scale__ssa_v0_tensor->buffer.addr) + expert_w1_scale__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: expert_w3_scale__ssa_v0
+    __gm__ Tensor* expert_w3_scale__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* expert_w3_scale__ssa_v0 = reinterpret_cast<__gm__ float*>(expert_w3_scale__ssa_v0_tensor->buffer.addr) + expert_w3_scale__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: recv_x_scale_dq_inline29__ssa_v0
+    __gm__ Tensor* recv_x_scale_dq_inline29__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ float* recv_x_scale_dq_inline29__ssa_v0 = reinterpret_cast<__gm__ float*>(recv_x_scale_dq_inline29__ssa_v0_tensor->buffer.addr) + recv_x_scale_dq_inline29__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[5]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack tensor: ret1__out
+    __gm__ Tensor* ret1__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[6]);
+    __gm__ float* ret1__out = reinterpret_cast<__gm__ float*>(ret1__out_tensor->buffer.addr) + ret1__out_tensor->start_offset;
+
+    // Unpack scalar: local_i_inline67__idx_v0
+    union { uint64_t u64; int64_t val; } local_i_inline67__idx_v0_conv;
+    local_i_inline67__idx_v0_conv.u64 = args[7];
+    int64_t local_i_inline67__idx_v0 = local_i_inline67__idx_v0_conv.val;
+
+    // Unpack scalar: n0_inline72__idx_v0
+    union { uint64_t u64; int64_t val; } n0_inline72__idx_v0_conv;
+    n0_inline72__idx_v0_conv.u64 = args[8];
+    int64_t n0_inline72__idx_v0 = n0_inline72__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_gate_up_dequant(gate_acc_inline69__rv_v2, up_acc_inline46__rv_v2, expert_w1_scale__ssa_v0, expert_w3_scale__ssa_v0, recv_x_scale_dq_inline29__ssa_v0, ret0__out, ret1__out, local_i_inline67__idx_v0, n0_inline72__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_h_q.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_h_q.cpp
new file mode 100644
index 000000000..65aabca5b
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_h_q.cpp
@@ -0,0 +1,213 @@
+// Kernel Function: exp_h_q
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_h_q(__gm__ float* v1, __gm__ int8_t* v2, __gm__ float* v3) {
+  RoundMode v4 = RoundMode::CAST_TRUNC;
+  RoundMode v5 = RoundMode::CAST_ROUND;
+  unsigned v6 = 0;
+  const float v7 = 127.0f;
+  const int32_t v8 = 256;
+  const int32_t v9 = 0;
+  const float v10 = 9.99999974E-5f;
+  const int32_t v11 = 1;
+  const int32_t v12 = 4096;
+  const int32_t v13 = 16;
+  const int64_t v14 = 24640;
+  const int64_t v15 = 16448;
+  const int64_t v16 = 64;
+  const int64_t v17 = 0;
+  const int64_t v18 = 61568;
+  const int64_t v19 = 45184;
+  const int64_t v20 = 28800;
+  const int64_t v21 = 28736;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v22 = (size_t) v12;
+  size_t v23 = (size_t) v9;
+  size_t v24 = (size_t) v8;
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v26 = (uint64_t) v21;
+  TASSIGN(v25, v26);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TEXPANDS(v25, v10);
+  for (size_t v27 = v23; v27 < v22; v27 += v24) {
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v29 = (uint64_t) v20;
+    TASSIGN(v28, v29);
+    pto::Shape<1, 1, 1, 16, 256> v30 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v31 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v32 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v12 + (unsigned) ((int32_t) v27) * (unsigned) v11), v30, v31);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v28, v32);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v34 = (uint64_t) v19;
+    TASSIGN(v33, v34);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    TNEG(v33, v28);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v36 = (uint64_t) v20;
+    TASSIGN(v35, v36);
+    pipe_barrier(PIPE_V);
+    TMAX(v35, v28, v33);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v38 = (uint64_t) v19;
+    TASSIGN(v37, v38);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+    uint64_t v40 = (uint64_t) v18;
+    TASSIGN(v39, v40);
+    pipe_barrier(PIPE_V);
+    TROWMAX(v39, v35, v37);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v42 = (uint64_t) v18;
+    TASSIGN(v41, v42);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v44 = (uint64_t) v21;
+    TASSIGN(v43, v44);
+    pipe_barrier(PIPE_V);
+    TMAX(v43, v25, v41);
+  }
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v46 = (uint64_t) v17;
+  TASSIGN(v45, v46);
+  TEXPANDS(v45, v7);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v48 = (uint64_t) v21;
+  TASSIGN(v47, v48);
+  pipe_barrier(PIPE_V);
+  TDIV(v47, v45, v25);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v49 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v50 = (uint64_t) v17;
+  TASSIGN(v49, v50);
+  pipe_barrier(PIPE_V);
+  TRECIP(v49, v47);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v51 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v52 = (uint64_t) v17;
+  TASSIGN(v51, v52);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v53 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v54 = (uint64_t) v21;
+  TASSIGN(v53, v54);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  for (size_t v55 = v23; v55 < v22; v55 += v24) {
+    int32_t v56 = (int32_t) v55;
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v57 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v58 = (uint64_t) v20;
+    TASSIGN(v57, v58);
+    pto::Shape<1, 1, 1, 16, 256> v59 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v60 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v61 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v12 + (unsigned) v56 * (unsigned) v11), v59, v60);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    TLOAD(v57, v61);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v63 = (uint64_t) v20;
+    TASSIGN(v62, v63);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    TROWEXPANDMUL(v62, v57, v53);
+    Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v65 = (uint64_t) v16;
+    TASSIGN(v64, v65);
+    pipe_barrier(PIPE_V);
+    TCVT(v64, v62, v5);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v67 = (uint64_t) v15;
+    TASSIGN(v66, v67);
+    pipe_barrier(PIPE_V);
+    TCVT(v66, v64, v5);
+    Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v68 = Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v69 = (uint64_t) v14;
+    TASSIGN(v68, v69);
+    pipe_barrier(PIPE_V);
+    wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+    TCVT(v68, v66, v4);
+    set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    pto::Shape<1, 1, 1, 16, 256> v70 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v71 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v72 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v6 + v6 * (unsigned) v12 + (unsigned) v56 * (unsigned) v11), v70, v71);
+    wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    TSTORE(v72, v68);
+    set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  }
+  pto::Shape<1, 1, 1, 16, 1> v73 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v74 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v75 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v3 + (v6 + v6 * (unsigned) v11 + v6 * (unsigned) v13), v73, v74);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v75, v51);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: h_tile_fp32_inline18__rv_v2
+    __gm__ Tensor* h_tile_fp32_inline18__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* h_tile_fp32_inline18__rv_v2 = reinterpret_cast<__gm__ float*>(h_tile_fp32_inline18__rv_v2_tensor->buffer.addr) + h_tile_fp32_inline18__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: h_tile_i8_inline92__ssa_v0
+    __gm__ Tensor* h_tile_i8_inline92__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* h_tile_i8_inline92__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(h_tile_i8_inline92__ssa_v0_tensor->buffer.addr) + h_tile_i8_inline92__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    exp_h_q(h_tile_fp32_inline18__rv_v2, h_tile_i8_inline92__ssa_v0, ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_recv_y_write.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_recv_y_write.cpp
new file mode 100644
index 000000000..0135b048e
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_recv_y_write.cpp
@@ -0,0 +1,108 @@
+// Kernel Function: exp_recv_y_write
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_recv_y_write(__gm__ float* v1, __gm__ bfloat16_t* v2, int32_t v3, int32_t v4) {
+  RoundMode v5 = RoundMode::CAST_RINT;
+  unsigned v6 = 0;
+  const int32_t v7 = 4096;
+  const int32_t v8 = 1;
+  const int32_t v9 = 512;
+  const int32_t v10 = 16;
+  const int64_t v11 = 32768;
+  const int64_t v12 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v13 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v14 = (uint64_t) v12;
+  TASSIGN(v13, v14);
+  pto::Shape<1, 1, 1, 16, 512> v15 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v16 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v17 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v8), v15, v16);
+  TLOAD(v13, v17);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v18 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v19 = (uint64_t) v11;
+  TASSIGN(v18, v19);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v18, v13, v5);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v20 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v21 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v22 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v6 + (unsigned) v3 * (unsigned) v7 + (unsigned) v4 * (unsigned) v8), v20, v21);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v22, v18);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: y_2d_v1_inline17__ssa_v0
+    __gm__ Tensor* y_2d_v1_inline17__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* y_2d_v1_inline17__ssa_v0 = reinterpret_cast<__gm__ float*>(y_2d_v1_inline17__ssa_v0_tensor->buffer.addr) + y_2d_v1_inline17__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: recv_y_flat_inline53__iter_v5
+    __gm__ Tensor* recv_y_flat_inline53__iter_v5_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ bfloat16_t* recv_y_flat_inline53__iter_v5 = reinterpret_cast<__gm__ bfloat16_t*>(recv_y_flat_inline53__iter_v5_tensor->buffer.addr) + recv_y_flat_inline53__iter_v5_tensor->start_offset;
+
+    // Unpack scalar: flat_t0_inline40__ssa_v0
+    union { uint64_t u64; int64_t val; } flat_t0_inline40__ssa_v0_conv;
+    flat_t0_inline40__ssa_v0_conv.u64 = args[2];
+    int64_t flat_t0_inline40__ssa_v0 = flat_t0_inline40__ssa_v0_conv.val;
+
+    // Unpack scalar: d0_inline49__idx_v0
+    union { uint64_t u64; int64_t val; } d0_inline49__idx_v0_conv;
+    d0_inline49__idx_v0_conv.u64 = args[3];
+    int64_t d0_inline49__idx_v0 = d0_inline49__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_recv_y_write(y_2d_v1_inline17__ssa_v0, recv_y_flat_inline53__iter_v5, flat_t0_inline40__ssa_v0, d0_inline49__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_swiglu.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_swiglu.cpp
new file mode 100644
index 000000000..fd893f9ed
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_swiglu.cpp
@@ -0,0 +1,161 @@
+// Kernel Function: exp_swiglu
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_swiglu(__gm__ float* v1, __gm__ float* v2, __gm__ float* v3, __gm__ float* v4, int32_t v5) {
+  unsigned v6 = 0;
+  const float v7 = 1.0f;
+  const int32_t v8 = 1;
+  const int32_t v9 = 256;
+  const int32_t v10 = 16;
+  const int64_t v11 = 0;
+  const int64_t v12 = 49216;
+  const int64_t v13 = 32832;
+  const int64_t v14 = 16448;
+  const int64_t v15 = 64;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v16 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v17 = (uint64_t) v15;
+  TASSIGN(v16, v17);
+  pto::Shape<1, 1, 1, 16, 256> v18 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v19 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v20 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v8), v18, v19);
+  TLOAD(v16, v20);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v21 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v22 = (uint64_t) v14;
+  TASSIGN(v21, v22);
+  pto::Shape<1, 1, 1, 16, 256> v23 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v24 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v25 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v2 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v8), v23, v24);
+  TLOAD(v21, v25);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v27 = (uint64_t) v13;
+  TASSIGN(v26, v27);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TNEG(v26, v16);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v29 = (uint64_t) v13;
+  TASSIGN(v28, v29);
+  pipe_barrier(PIPE_V);
+  TEXP(v28, v26);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v31 = (uint64_t) v13;
+  TASSIGN(v30, v31);
+  pipe_barrier(PIPE_V);
+  TADDS(v30, v28, v7);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v32 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v33 = (uint64_t) v12;
+  TASSIGN(v32, v33);
+  pipe_barrier(PIPE_V);
+  TRECIP(v32, v30);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v35 = (uint64_t) v15;
+  TASSIGN(v34, v35);
+  pipe_barrier(PIPE_V);
+  TMUL(v34, v16, v32);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v37 = (uint64_t) v15;
+  TASSIGN(v36, v37);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TMUL(v36, v34, v21);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v8);
+  uint64_t v39 = (uint64_t) v11;
+  TASSIGN(v38, v39);
+  pto::Shape<1, 1, 1, 16, 1> v40 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 256> v41 = pto::Stride<16, 16, 16, 1, 256>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 256>, pto::Layout::DN> v42 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 256>, pto::Layout::DN>(v3 + (v6 + (unsigned) v5 * (unsigned) v8 + v6 * (unsigned) v9), v40, v41);
+  TLOAD(v38, v42);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v44 = (uint64_t) v15;
+  TASSIGN(v43, v44);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDMUL(v43, v36, v38);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 256> v45 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v46 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v47 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v4 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v8), v45, v46);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v47, v43);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: gate_2d_v1_inline7__ssa_v0
+    __gm__ Tensor* gate_2d_v1_inline7__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* gate_2d_v1_inline7__ssa_v0 = reinterpret_cast<__gm__ float*>(gate_2d_v1_inline7__ssa_v0_tensor->buffer.addr) + gate_2d_v1_inline7__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: up_2d_v1_inline78__ssa_v0
+    __gm__ Tensor* up_2d_v1_inline78__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* up_2d_v1_inline78__ssa_v0 = reinterpret_cast<__gm__ float*>(up_2d_v1_inline78__ssa_v0_tensor->buffer.addr) + up_2d_v1_inline78__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: recv_weights_flat_inline62__ssa_v0
+    __gm__ Tensor* recv_weights_flat_inline62__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* recv_weights_flat_inline62__ssa_v0 = reinterpret_cast<__gm__ float*>(recv_weights_flat_inline62__ssa_v0_tensor->buffer.addr) + recv_weights_flat_inline62__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: flat_t0_inline40__ssa_v0
+    union { uint64_t u64; int64_t val; } flat_t0_inline40__ssa_v0_conv;
+    flat_t0_inline40__ssa_v0_conv.u64 = args[4];
+    int64_t flat_t0_inline40__ssa_v0 = flat_t0_inline40__ssa_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_swiglu(gate_2d_v1_inline7__ssa_v0, up_2d_v1_inline78__ssa_v0, recv_weights_flat_inline62__ssa_v0, ret0__out, flat_t0_inline40__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_swiglu_mask.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_swiglu_mask.cpp
new file mode 100644
index 000000000..69645e11b
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_swiglu_mask.cpp
@@ -0,0 +1,108 @@
+// Kernel Function: exp_swiglu_mask
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_swiglu_mask(__gm__ float* v1, __gm__ float* v2, int32_t v3, int32_t v4) {
+  unsigned v5 = 0;
+  const int32_t v6 = 4096;
+  const int32_t v7 = 1;
+  const int32_t v8 = 256;
+  const int32_t v9 = 16;
+  const int64_t v10 = 16384;
+  const int64_t v11 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v12 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v13 = (uint64_t) v11;
+  TASSIGN(v12, v13);
+  pto::Shape<1, 1, 1, 16, 256> v14 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v15 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v16 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v8 + v5 * (unsigned) v7), v14, v15);
+  TLOAD(v12, v16);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  v12.SetValidShape(v3, v8);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v17 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v9, v8);
+  uint64_t v18 = (uint64_t) v10;
+  TASSIGN(v17, v18);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TFILLPAD(v17, v12);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 256> v19 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v20 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v21 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v5 + v5 * (unsigned) v6 + (unsigned) v4 * (unsigned) v7), v19, v20);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v21, v17);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: h_chunk_inline89__ssa_v0
+    __gm__ Tensor* h_chunk_inline89__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* h_chunk_inline89__ssa_v0 = reinterpret_cast<__gm__ float*>(h_chunk_inline89__ssa_v0_tensor->buffer.addr) + h_chunk_inline89__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: h_tile_fp32_inline18__iter_v1
+    __gm__ Tensor* h_tile_fp32_inline18__iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* h_tile_fp32_inline18__iter_v1 = reinterpret_cast<__gm__ float*>(h_tile_fp32_inline18__iter_v1_tensor->buffer.addr) + h_tile_fp32_inline18__iter_v1_tensor->start_offset;
+
+    // Unpack scalar: valid_rows_inline73__ssa_v0
+    union { uint64_t u64; int64_t val; } valid_rows_inline73__ssa_v0_conv;
+    valid_rows_inline73__ssa_v0_conv.u64 = args[2];
+    int64_t valid_rows_inline73__ssa_v0 = valid_rows_inline73__ssa_v0_conv.val;
+
+    // Unpack scalar: n0_inline72__idx_v0
+    union { uint64_t u64; int64_t val; } n0_inline72__idx_v0_conv;
+    n0_inline72__idx_v0_conv.u64 = args[3];
+    int64_t n0_inline72__idx_v0 = n0_inline72__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_swiglu_mask(h_chunk_inline89__ssa_v0, h_tile_fp32_inline18__iter_v1, valid_rows_inline73__ssa_v0, n0_inline72__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_w2_dequant.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_w2_dequant.cpp
new file mode 100644
index 000000000..0f31e0015
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/exp_w2_dequant.cpp
@@ -0,0 +1,150 @@
+// Kernel Function: exp_w2_dequant
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void exp_w2_dequant(__gm__ int32_t* v1, __gm__ float* v2, __gm__ float* v3, __gm__ float* v4, int32_t v5, int32_t v6) {
+  RoundMode v7 = RoundMode::CAST_NONE;
+  unsigned v8 = 0;
+  const int32_t v9 = 8192;
+  const int32_t v10 = 4096;
+  const int32_t v11 = 512;
+  const int32_t v12 = 16;
+  const int32_t v13 = 1;
+  const int64_t v14 = 34880;
+  const int64_t v15 = 32832;
+  const int64_t v16 = 32768;
+  const int64_t v17 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, int32_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v18 = Tile<TileType::Vec, int32_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v19 = (uint64_t) v17;
+  TASSIGN(v18, v19);
+  pto::Shape<1, 1, 1, 16, 512> v20 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v21 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v22 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v1 + ((v8 + v8 * (unsigned) v9) + v8 * (unsigned) v11 + v8 * (unsigned) v13), v20, v21);
+  TLOAD(v18, v22);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v23 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v13);
+  uint64_t v24 = (uint64_t) v16;
+  TASSIGN(v23, v24);
+  pto::Shape<1, 1, 1, 16, 1> v25 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v26 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v27 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v3 + (v8 + v8 * (unsigned) v13 + v8 * (unsigned) v12), v25, v26);
+  TLOAD(v23, v27);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, int32_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, int32_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v29 = (uint64_t) v17;
+  TASSIGN(v28, v29);
+  Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v31 = (uint64_t) v15;
+  TASSIGN(v30, v31);
+  pto::Shape<1, 1, 1, 1, 512> v32 = pto::Shape<1, 1, 1, 1, 512>();
+  pto::Stride<4096, 4096, 4096, 4096, 1> v33 = pto::Stride<4096, 4096, 4096, 4096, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND> v34 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND>(v2 + (v8 + (unsigned) v5 * (unsigned) v10 + (unsigned) v6 * (unsigned) v13), v32, v33);
+  TLOAD(v30, v34);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v36 = (uint64_t) v14;
+  TASSIGN(v35, v36);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v35, v28, v7);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v38 = (uint64_t) v14;
+  TASSIGN(v37, v38);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TROWEXPANDMUL(v37, v35, v23);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v40 = (uint64_t) v14;
+  TASSIGN(v39, v40);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  TCOLEXPANDMUL(v39, v37, v30);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v41 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v42 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v43 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v4 + (v8 + v8 * (unsigned) v11 + v8 * (unsigned) v13), v41, v42);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v43, v39);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: y_acc_inline109__rv_v2
+    __gm__ Tensor* y_acc_inline109__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int32_t* y_acc_inline109__rv_v2 = reinterpret_cast<__gm__ int32_t*>(y_acc_inline109__rv_v2_tensor->buffer.addr) + y_acc_inline109__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: expert_w2_scale__ssa_v0
+    __gm__ Tensor* expert_w2_scale__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* expert_w2_scale__ssa_v0 = reinterpret_cast<__gm__ float*>(expert_w2_scale__ssa_v0_tensor->buffer.addr) + expert_w2_scale__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: h_tile_scale_dq_inline63__ssa_v0
+    __gm__ Tensor* h_tile_scale_dq_inline63__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* h_tile_scale_dq_inline63__ssa_v0 = reinterpret_cast<__gm__ float*>(h_tile_scale_dq_inline63__ssa_v0_tensor->buffer.addr) + h_tile_scale_dq_inline63__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: local_i_inline67__idx_v0
+    union { uint64_t u64; int64_t val; } local_i_inline67__idx_v0_conv;
+    local_i_inline67__idx_v0_conv.u64 = args[4];
+    int64_t local_i_inline67__idx_v0 = local_i_inline67__idx_v0_conv.val;
+
+    // Unpack scalar: d0_inline49__idx_v0
+    union { uint64_t u64; int64_t val; } d0_inline49__idx_v0_conv;
+    d0_inline49__idx_v0_conv.u64 = args[5];
+    int64_t d0_inline49__idx_v0 = d0_inline49__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    exp_w2_dequant(y_acc_inline109__rv_v2, expert_w2_scale__ssa_v0, h_tile_scale_dq_inline63__ssa_v0, ret0__out, local_i_inline67__idx_v0, d0_inline49__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_add.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_add.cpp
new file mode 100644
index 000000000..e51770d93
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_add.cpp
@@ -0,0 +1,89 @@
+/*
+ * Copyright (c) PyPTO Contributors.
+ * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+ * CANN Open Software License Agreement Version 2.0 (the "License").
+ * Please refer to the License for details. You may not use this file except in compliance with the License.
+ * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+ * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+ * See LICENSE in the root of the software repository for the full text of the License.
+ * -----------------------------------------------------------------------------------------------------------
+ */
+/**
+ * FFN residual add: ``ffn_out = routed_y + sh``.
+ *
+ *   routed_y  : FP32  [T, D]   (combine output)
+ *   sh        : BF16  [T, D]   (moe_expert shared-expert output)
+ *   ffn_out   : BF16  [T, D]
+ */
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#define __aicore__ [aicore]
+#endif
+
+#include <cstdint>
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+using namespace pto;
+
+static constexpr int T = 16;
+static constexpr int D = 4096;
+
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t *args) {
+    __gm__ Tensor *routed_y_tensor = reinterpret_cast<__gm__ Tensor *>(args[0]);
+    __gm__ Tensor *sh_tensor = reinterpret_cast<__gm__ Tensor *>(args[1]);
+    __gm__ Tensor *ffn_out_tensor = reinterpret_cast<__gm__ Tensor *>(args[2]);
+
+    __gm__ float *routed_y =
+        reinterpret_cast<__gm__ float *>(routed_y_tensor->buffer.addr) + routed_y_tensor->start_offset;
+    __gm__ bfloat16_t *sh =
+        reinterpret_cast<__gm__ bfloat16_t *>(sh_tensor->buffer.addr) + sh_tensor->start_offset;
+    __gm__ bfloat16_t *ffn_out =
+        reinterpret_cast<__gm__ bfloat16_t *>(ffn_out_tensor->buffer.addr) + ffn_out_tensor->start_offset;
+
+    using RowFpG = GlobalTensor<float, Shape<1, 1, 1, 1, D>, Stride<D, D, D, D, 1>>;
+    using RowBfG = GlobalTensor<bfloat16_t, Shape<1, 1, 1, 1, D>, Stride<D, D, D, D, 1>>;
+    using RowFpTile = Tile<TileType::Vec, float, 1, D, BLayout::RowMajor, 1, D>;
+    using RowBfTile = Tile<TileType::Vec, bfloat16_t, 1, D, BLayout::RowMajor, 1, D>;
+
+    // 4 tiles laid out within the 192 KiB AIV UB (0x30000 is OOB so the 4th
+    // slot uses a tighter 0x28000 offset instead of the usual 64 KB pitch).
+    RowFpTile routed_fp;
+    RowFpTile sh_fp;
+    RowBfTile sh_bf;
+    RowBfTile out_bf;
+    TASSIGN(routed_fp, 0x0);
+    TASSIGN(sh_fp, 0x10000);
+    TASSIGN(sh_bf, 0x20000);
+    TASSIGN(out_bf, 0x28000);
+
+    for (int t = 0; t < T; ++t) {
+        RowFpG r_g(routed_y + t * D);
+        RowBfG s_g(sh + t * D);
+        RowBfG o_g(ffn_out + t * D);
+
+        TLOAD(routed_fp, r_g);
+        TLOAD(sh_bf, s_g);
+        set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+        wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+
+        TCVT(sh_fp, sh_bf, RoundMode::CAST_RINT);
+        pipe_barrier(PIPE_V);
+        TADD(routed_fp, routed_fp, sh_fp);
+        pipe_barrier(PIPE_V);
+        TCVT(out_bf, routed_fp, RoundMode::CAST_RINT);
+        pipe_barrier(PIPE_V);
+
+        set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+        wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+        TSTORE(o_g, out_bf);
+        set_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+        wait_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+    }
+    pipe_barrier(PIPE_ALL);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_norm_apply.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_norm_apply.cpp
new file mode 100644
index 000000000..885251cb8
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_norm_apply.cpp
@@ -0,0 +1,160 @@
+// Kernel Function: ffn_norm_apply
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void ffn_norm_apply(__gm__ float* v1, __gm__ bfloat16_t* v2, __gm__ float* v3, __gm__ bfloat16_t* v4, __gm__ bfloat16_t* v5, int32_t v6) {
+  RoundMode v7 = RoundMode::CAST_RINT;
+  RoundMode v8 = RoundMode::CAST_ROUND;
+  unsigned v9 = 0;
+  const int32_t v10 = 512;
+  const int32_t v11 = 4096;
+  const int32_t v12 = 16;
+  const int32_t v13 = 1;
+  const int64_t v14 = 49216;
+  const int64_t v15 = 16448;
+  const int64_t v16 = 64;
+  const int64_t v17 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  int32_t v18 = (int32_t) ((uint32_t) v6 * (uint32_t) v10);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v19 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v20 = (uint64_t) v17;
+  TASSIGN(v19, v20);
+  pto::Shape<1, 1, 1, 1, 16> v21 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<16, 16, 16, 16, 1> v22 = pto::Stride<16, 16, 16, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND> v23 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND>(v1 + (v9 + v9 * (unsigned) v12 + v9 * (unsigned) v13), v21, v22);
+  TLOAD(v19, v23);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v24 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v13);
+  uint64_t v25 = (uint64_t) v17;
+  TASSIGN(v24, v25);
+  Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v10);
+  uint64_t v27 = (uint64_t) v16;
+  TASSIGN(v26, v27);
+  pto::Shape<1, 1, 1, 16, 512> v28 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v29 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v30 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v9 + v9 * (unsigned) v11 + (unsigned) v18 * (unsigned) v13), v28, v29);
+  TLOAD(v26, v30);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v31 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v10);
+  uint64_t v32 = (uint64_t) v15;
+  TASSIGN(v31, v32);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v31, v26, v8);
+  Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v10);
+  uint64_t v34 = (uint64_t) v14;
+  TASSIGN(v33, v34);
+  pto::Shape<1, 1, 1, 1, 512> v35 = pto::Shape<1, 1, 1, 1, 512>();
+  pto::Stride<512, 512, 512, 512, 1> v36 = pto::Stride<512, 512, 512, 512, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<512, 512, 512, 512, 1>, pto::Layout::ND> v37 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<512, 512, 512, 512, 1>, pto::Layout::ND>(v3 + (v9 + (unsigned) v18 * (unsigned) v13), v35, v36);
+  TLOAD(v33, v37);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v10);
+  uint64_t v39 = (uint64_t) v14;
+  TASSIGN(v38, v39);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v40 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v10);
+  uint64_t v41 = (uint64_t) v15;
+  TASSIGN(v40, v41);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDMUL(v40, v31, v24);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v42 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v10);
+  uint64_t v43 = (uint64_t) v15;
+  TASSIGN(v42, v43);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TCOLEXPANDMUL(v42, v40, v38);
+  Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v44 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v10);
+  uint64_t v45 = (uint64_t) v16;
+  TASSIGN(v44, v45);
+  pipe_barrier(PIPE_V);
+  TCVT(v44, v42, v7);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v46 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v47 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v48 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v4 + (v9 + v9 * (unsigned) v11 + (unsigned) v18 * (unsigned) v13), v46, v47);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v48, v44);
+  pto::Shape<1, 1, 1, 16, 512> v49 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v50 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v51 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v5 + (v9 + v9 * (unsigned) v11 + (unsigned) v18 * (unsigned) v13), v49, v50);
+  TSTORE(v51, v44);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: inv_rms_inline126__ssa_v1
+    __gm__ Tensor* inv_rms_inline126__ssa_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* inv_rms_inline126__ssa_v1 = reinterpret_cast<__gm__ float*>(inv_rms_inline126__ssa_v1_tensor->buffer.addr) + inv_rms_inline126__ssa_v1_tensor->start_offset;
+
+    // Unpack tensor: x_mixed_flat_inline127__ssa_v0
+    __gm__ Tensor* x_mixed_flat_inline127__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ bfloat16_t* x_mixed_flat_inline127__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(x_mixed_flat_inline127__ssa_v0_tensor->buffer.addr) + x_mixed_flat_inline127__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: norm_w__ssa_v0
+    __gm__ Tensor* norm_w__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* norm_w__ssa_v0 = reinterpret_cast<__gm__ float*>(norm_w__ssa_v0_tensor->buffer.addr) + norm_w__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: x_norm_bf16_inline111__iter_v1
+    __gm__ Tensor* x_norm_bf16_inline111__iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ bfloat16_t* x_norm_bf16_inline111__iter_v1 = reinterpret_cast<__gm__ bfloat16_t*>(x_norm_bf16_inline111__iter_v1_tensor->buffer.addr) + x_norm_bf16_inline111__iter_v1_tensor->start_offset;
+
+    // Unpack tensor: x_norm__iter_v1
+    __gm__ Tensor* x_norm__iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ bfloat16_t* x_norm__iter_v1 = reinterpret_cast<__gm__ bfloat16_t*>(x_norm__iter_v1_tensor->buffer.addr) + x_norm__iter_v1_tensor->start_offset;
+
+    // Unpack scalar: db_inline125__idx_v0
+    union { uint64_t u64; int64_t val; } db_inline125__idx_v0_conv;
+    db_inline125__idx_v0_conv.u64 = args[5];
+    int64_t db_inline125__idx_v0 = db_inline125__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    ffn_norm_apply(inv_rms_inline126__ssa_v1, x_mixed_flat_inline127__ssa_v0, norm_w__ssa_v0, x_norm_bf16_inline111__iter_v1, x_norm__iter_v1, db_inline125__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_norm_rms.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_norm_rms.cpp
new file mode 100644
index 000000000..c648ef28c
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/ffn_norm_rms.cpp
@@ -0,0 +1,279 @@
+// Kernel Function: ffn_norm_rms
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void ffn_norm_rms(__gm__ bfloat16_t* v1, __gm__ float* v2) {
+  RoundMode v3 = RoundMode::CAST_ROUND;
+  unsigned v4 = 0;
+  const float v5 = 9.99999997E-7f;
+  const float v6 = 2.44140625E-4f;
+  const int32_t v7 = 1536;
+  const int32_t v8 = 1024;
+  const int32_t v9 = 512;
+  const int32_t v10 = 4;
+  const int32_t v11 = 8;
+  const int32_t v12 = 0;
+  const float v13 = 0.0f;
+  const int32_t v14 = 1;
+  const int32_t v15 = 4096;
+  const int32_t v16 = 16;
+  const int64_t v17 = 64;
+  const int64_t v18 = 0;
+  const int64_t v19 = 98496;
+  const int64_t v20 = 65728;
+  const int64_t v21 = 49344;
+  const int64_t v22 = 32960;
+  const int64_t v23 = 16576;
+  const int64_t v24 = 192;
+  const int64_t v25 = 128;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+  uint64_t v27 = (uint64_t) v25;
+  TASSIGN(v26, v27);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+  TEXPANDS(v26, v13);
+  for (size_t v28 = (size_t) v12; v28 < ((size_t) v11); v28 += (size_t) v10) {
+    int32_t v29 = (int32_t) ((uint32_t) ((int32_t) v28) * (uint32_t) v9);
+    Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v31 = (uint64_t) v24;
+    TASSIGN(v30, v31);
+    pto::Shape<1, 1, 1, 16, 512> v32 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v33 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v34 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v4 + v4 * (unsigned) v15 + (unsigned) v29 * (unsigned) v14), v32, v33);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v30, v34);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v36 = (uint64_t) v23;
+    TASSIGN(v35, v36);
+    pto::Shape<1, 1, 1, 16, 512> v37 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v38 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v39 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v4 + v4 * (unsigned) v15 + (unsigned) ((int32_t) (uint32_t) v29 + (uint32_t) v9) * (unsigned) v14), v37, v38);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+    TLOAD(v35, v39);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v40 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v41 = (uint64_t) v22;
+    TASSIGN(v40, v41);
+    pto::Shape<1, 1, 1, 16, 512> v42 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v43 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v44 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v4 + v4 * (unsigned) v15 + (unsigned) ((int32_t) (uint32_t) v29 + (uint32_t) v8) * (unsigned) v14), v42, v43);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    TLOAD(v40, v44);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+    Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v46 = (uint64_t) v21;
+    TASSIGN(v45, v46);
+    pto::Shape<1, 1, 1, 16, 512> v47 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v48 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v49 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v4 + v4 * (unsigned) v15 + (unsigned) ((int32_t) (uint32_t) v29 + (uint32_t) v7) * (unsigned) v14), v47, v48);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+    TLOAD(v45, v49);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v51 = (uint64_t) v20;
+    TASSIGN(v50, v51);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    TCVT(v50, v30, v3);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v53 = (uint64_t) v20;
+    TASSIGN(v52, v53);
+    pipe_barrier(PIPE_V);
+    TMUL(v52, v50, v50);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v55 = (uint64_t) v19;
+    TASSIGN(v54, v55);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v14);
+    uint64_t v57 = (uint64_t) v18;
+    TASSIGN(v56, v57);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v56, v52, v54);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v59 = (uint64_t) v18;
+    TASSIGN(v58, v59);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v60 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v61 = (uint64_t) v17;
+    TASSIGN(v60, v61);
+    pipe_barrier(PIPE_V);
+    TADD(v60, v26, v58);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v63 = (uint64_t) v20;
+    TASSIGN(v62, v63);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    TCVT(v62, v35, v3);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v65 = (uint64_t) v20;
+    TASSIGN(v64, v65);
+    pipe_barrier(PIPE_V);
+    TMUL(v64, v62, v62);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v67 = (uint64_t) v19;
+    TASSIGN(v66, v67);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v68 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v14);
+    uint64_t v69 = (uint64_t) v18;
+    TASSIGN(v68, v69);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v68, v64, v66);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v70 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v71 = (uint64_t) v18;
+    TASSIGN(v70, v71);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v72 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v73 = (uint64_t) v17;
+    TASSIGN(v72, v73);
+    pipe_barrier(PIPE_V);
+    TADD(v72, v60, v70);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v74 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v75 = (uint64_t) v20;
+    TASSIGN(v74, v75);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+    TCVT(v74, v40, v3);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v76 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v77 = (uint64_t) v20;
+    TASSIGN(v76, v77);
+    pipe_barrier(PIPE_V);
+    TMUL(v76, v74, v74);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v78 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v79 = (uint64_t) v19;
+    TASSIGN(v78, v79);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v80 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v14);
+    uint64_t v81 = (uint64_t) v18;
+    TASSIGN(v80, v81);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v80, v76, v78);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v82 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v83 = (uint64_t) v18;
+    TASSIGN(v82, v83);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v84 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v85 = (uint64_t) v17;
+    TASSIGN(v84, v85);
+    pipe_barrier(PIPE_V);
+    TADD(v84, v72, v82);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v86 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v87 = (uint64_t) v20;
+    TASSIGN(v86, v87);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+    TCVT(v86, v45, v3);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v88 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v89 = (uint64_t) v20;
+    TASSIGN(v88, v89);
+    pipe_barrier(PIPE_V);
+    TMUL(v88, v86, v86);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v90 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v9);
+    uint64_t v91 = (uint64_t) v19;
+    TASSIGN(v90, v91);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v92 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v16, v14);
+    uint64_t v93 = (uint64_t) v18;
+    TASSIGN(v92, v93);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v92, v88, v90);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v94 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v95 = (uint64_t) v18;
+    TASSIGN(v94, v95);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v96 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+    uint64_t v97 = (uint64_t) v25;
+    TASSIGN(v96, v97);
+    pipe_barrier(PIPE_V);
+    TADD(v96, v84, v94);
+  }
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v98 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+  uint64_t v99 = (uint64_t) v25;
+  TASSIGN(v98, v99);
+  pipe_barrier(PIPE_V);
+  TMULS(v98, v26, v6);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v100 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+  uint64_t v101 = (uint64_t) v25;
+  TASSIGN(v100, v101);
+  pipe_barrier(PIPE_V);
+  TADDS(v100, v98, v5);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v102 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+  uint64_t v103 = (uint64_t) v25;
+  TASSIGN(v102, v103);
+  pipe_barrier(PIPE_V);
+  TSQRT(v102, v100);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v104 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v16);
+  uint64_t v105 = (uint64_t) v17;
+  TASSIGN(v104, v105);
+  pipe_barrier(PIPE_V);
+  TRECIP(v104, v102);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v106 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<16, 16, 16, 16, 1> v107 = pto::Stride<16, 16, 16, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND> v108 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND>(v2 + (v4 + v4 * (unsigned) v16 + v4 * (unsigned) v14), v106, v107);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v108, v104);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_mixed_flat_inline127__ssa_v0
+    __gm__ Tensor* x_mixed_flat_inline127__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ bfloat16_t* x_mixed_flat_inline127__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(x_mixed_flat_inline127__ssa_v0_tensor->buffer.addr) + x_mixed_flat_inline127__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    ffn_norm_rms(x_mixed_flat_inline127__ssa_v0, ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/gate_dot.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/gate_dot.cpp
new file mode 100644
index 000000000..6bfda5720
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/gate_dot.cpp
@@ -0,0 +1,83 @@
+// Kernel Function: gate_dot
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void gate_dot(__gm__ float* v1) {
+  unsigned v2 = 0;
+  const float v3 = -1.00000002E+30f;
+  const int32_t v4 = 1;
+  const int32_t v5 = 32;
+  const int32_t v6 = 16;
+  const int64_t v7 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v8 = Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v6, v5);
+  uint64_t v9 = (uint64_t) v7;
+  TASSIGN(v8, v9);
+  TEXPANDS(v8, v3);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 32> v10 = pto::Shape<1, 1, 1, 16, 32>();
+  pto::Stride<512, 512, 512, 32, 1> v11 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v12 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v1 + (v2 + v2 * (unsigned) v5 + v2 * (unsigned) v4), v10, v11);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v12, v8);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    gate_dot(ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/gate_dot_0.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/gate_dot_0.cpp
new file mode 100644
index 000000000..b59045037
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/gate_dot_0.cpp
@@ -0,0 +1,317 @@
+// Kernel Function: gate_dot_0
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void gate_dot_0(__gm__ float* v1, __gm__ bfloat16_t* v2, __gm__ float* v3, __gm__ float* v4, __gm__ float* v5) {
+  RoundMode v6 = RoundMode::CAST_ROUND;
+  unsigned v7 = 0;
+  const int32_t v8 = 480;
+  const int32_t v9 = 15;
+  const int32_t v10 = 448;
+  const int32_t v11 = 14;
+  const int32_t v12 = 416;
+  const int32_t v13 = 13;
+  const int32_t v14 = 384;
+  const int32_t v15 = 12;
+  const int32_t v16 = 352;
+  const int32_t v17 = 11;
+  const int32_t v18 = 320;
+  const int32_t v19 = 10;
+  const int32_t v20 = 288;
+  const int32_t v21 = 9;
+  const int32_t v22 = 256;
+  const int32_t v23 = 224;
+  const int32_t v24 = 7;
+  const int32_t v25 = 192;
+  const int32_t v26 = 6;
+  const int32_t v27 = 160;
+  const int32_t v28 = 5;
+  const int32_t v29 = 128;
+  const int32_t v30 = 4;
+  const int32_t v31 = 96;
+  const int32_t v32 = 3;
+  const int32_t v33 = 64;
+  const int32_t v34 = 2;
+  const int32_t v35 = 32;
+  const float v36 = 1.0f;
+  const float v37 = 0.0f;
+  const int32_t v38 = 0;
+  const int32_t v39 = 512;
+  const int32_t v40 = 8;
+  const int32_t v41 = 4096;
+  const int32_t v42 = 16;
+  const int32_t v43 = 1;
+  const int64_t v44 = 32896;
+  const int64_t v45 = 32832;
+  const int64_t v46 = 32768;
+  const int64_t v47 = 0;
+  const int64_t v48 = 82176;
+  const int64_t v49 = 49408;
+  const int64_t v50 = 33024;
+  const int64_t v51 = 32960;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v52 = (size_t) v43;
+  size_t v53 = (size_t) v40;
+  size_t v54 = (size_t) v38;
+  set_flag(PIPE_S, PIPE_V, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  for (size_t v55 = v54; v55 < v53; v55 += v52) {
+    int32_t v56 = (int32_t) v55;
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v57 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v58 = (uint64_t) v51;
+    TASSIGN(v57, v58);
+    wait_flag(PIPE_S, PIPE_V, EVENT_ID0);
+    TEXPANDS(v57, v37);
+    for (size_t v59 = v54; v59 < v53; v59 += v52) {
+      int32_t v60 = (int32_t) ((uint32_t) ((int32_t) v59) * (uint32_t) v39);
+      Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v61 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v42, v39);
+      uint64_t v62 = (uint64_t) v50;
+      TASSIGN(v61, v62);
+      pto::Shape<1, 1, 1, 16, 512> v63 = pto::Shape<1, 1, 1, 16, 512>();
+      pto::Stride<65536, 65536, 65536, 4096, 1> v64 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+      GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v65 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v7 + v7 * (unsigned) v41 + (unsigned) v60 * (unsigned) v43), v63, v64);
+      wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+      TLOAD(v61, v65);
+      set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+      Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v42, v39);
+      uint64_t v67 = (uint64_t) v49;
+      TASSIGN(v66, v67);
+      wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+      TCVT(v66, v61, v6);
+      set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+      Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v68 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v39);
+      uint64_t v69 = (uint64_t) v48;
+      TASSIGN(v68, v69);
+      pto::Shape<1, 1, 1, 1, 512> v70 = pto::Shape<1, 1, 1, 1, 512>();
+      pto::Stride<4096, 4096, 4096, 4096, 1> v71 = pto::Stride<4096, 4096, 4096, 4096, 1>();
+      GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND> v72 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND>(v3 + (v7 + (unsigned) v56 * (unsigned) v41 + (unsigned) v60 * (unsigned) v43), v70, v71);
+      wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+      TLOAD(v68, v72);
+      set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+      Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v73 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v42, v39);
+      uint64_t v74 = (uint64_t) v49;
+      TASSIGN(v73, v74);
+      wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+      pipe_barrier(PIPE_V);
+      TCOLEXPANDMUL(v73, v66, v68);
+      set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+      Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v75 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v42, v39);
+      uint64_t v76 = (uint64_t) v47;
+      TASSIGN(v75, v76);
+      Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v77 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v42, v43);
+      uint64_t v78 = (uint64_t) v46;
+      TASSIGN(v77, v78);
+      pipe_barrier(PIPE_V);
+      TROWSUM(v77, v73, v75);
+      Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v79 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+      uint64_t v80 = (uint64_t) v46;
+      TASSIGN(v79, v80);
+      Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v81 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+      uint64_t v82 = (uint64_t) v51;
+      TASSIGN(v81, v82);
+      pipe_barrier(PIPE_V);
+      TADD(v81, v57, v79);
+    };
+    set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    pto::Shape<1, 1, 1, 1, 16> v83 = pto::Shape<1, 1, 1, 1, 16>();
+    pto::Stride<16, 16, 16, 16, 1> v84 = pto::Stride<16, 16, 16, 16, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND> v85 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND>(v1 + (v7 + v7 * (unsigned) v42 + v7 * (unsigned) v43), v83, v84);
+    wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    TSTORE(v85, v57);
+    set_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+    float v86 = v4[v55];
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v87 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v88 = (uint64_t) v51;
+    TASSIGN(v87, v88);
+    wait_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v87, v85);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v89 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v90 = (uint64_t) v45;
+    TASSIGN(v89, v90);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+    TMULS(v89, v87, v37);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v91 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v92 = (uint64_t) v45;
+    TASSIGN(v91, v92);
+    pipe_barrier(PIPE_V);
+    TMAX(v91, v87, v89);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v93 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v94 = (uint64_t) v44;
+    TASSIGN(v93, v94);
+    TNEG(v93, v87);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v95 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v96 = (uint64_t) v51;
+    TASSIGN(v95, v96);
+    pipe_barrier(PIPE_V);
+    TMAX(v95, v87, v93);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v97 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v98 = (uint64_t) v51;
+    TASSIGN(v97, v98);
+    pipe_barrier(PIPE_V);
+    TNEG(v97, v95);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v99 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v100 = (uint64_t) v51;
+    TASSIGN(v99, v100);
+    pipe_barrier(PIPE_V);
+    TEXP(v99, v97);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v101 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v102 = (uint64_t) v51;
+    TASSIGN(v101, v102);
+    pipe_barrier(PIPE_V);
+    TADDS(v101, v99, v36);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v103 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v104 = (uint64_t) v51;
+    TASSIGN(v103, v104);
+    pipe_barrier(PIPE_V);
+    TLOG(v103, v101);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v105 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v106 = (uint64_t) v51;
+    TASSIGN(v105, v106);
+    pipe_barrier(PIPE_V);
+    TADD(v105, v91, v103);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v107 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v108 = (uint64_t) v51;
+    TASSIGN(v107, v108);
+    pipe_barrier(PIPE_V);
+    TSQRT(v107, v105);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v109 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v110 = (uint64_t) v51;
+    TASSIGN(v109, v110);
+    pipe_barrier(PIPE_V);
+    TADDS(v109, v107, v86);
+    set_flag(PIPE_V, PIPE_S, EVENT_ID0);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v111 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v43, v42);
+    uint64_t v112 = (uint64_t) v51;
+    TASSIGN(v111, v112);
+    wait_flag(PIPE_V, PIPE_S, EVENT_ID0);
+    float v113 = v111.GetValue(v38);
+    v5[v55] = v113;
+    float v114 = v111.GetValue(v43);
+    int32_t v115 = (int32_t) ((uint32_t) v56 + (uint32_t) v35);
+    v5[v115] = v114;
+    float v116 = v111.GetValue(v34);
+    int32_t v117 = (int32_t) ((uint32_t) v56 + (uint32_t) v33);
+    v5[v117] = v116;
+    float v118 = v111.GetValue(v32);
+    int32_t v119 = (int32_t) ((uint32_t) v56 + (uint32_t) v31);
+    v5[v119] = v118;
+    float v120 = v111.GetValue(v30);
+    int32_t v121 = (int32_t) ((uint32_t) v56 + (uint32_t) v29);
+    v5[v121] = v120;
+    float v122 = v111.GetValue(v28);
+    int32_t v123 = (int32_t) ((uint32_t) v56 + (uint32_t) v27);
+    v5[v123] = v122;
+    float v124 = v111.GetValue(v26);
+    int32_t v125 = (int32_t) ((uint32_t) v56 + (uint32_t) v25);
+    v5[v125] = v124;
+    float v126 = v111.GetValue(v24);
+    int32_t v127 = (int32_t) ((uint32_t) v56 + (uint32_t) v23);
+    v5[v127] = v126;
+    float v128 = v111.GetValue(v40);
+    int32_t v129 = (int32_t) ((uint32_t) v56 + (uint32_t) v22);
+    v5[v129] = v128;
+    float v130 = v111.GetValue(v21);
+    int32_t v131 = (int32_t) ((uint32_t) v56 + (uint32_t) v20);
+    v5[v131] = v130;
+    float v132 = v111.GetValue(v19);
+    int32_t v133 = (int32_t) ((uint32_t) v56 + (uint32_t) v18);
+    v5[v133] = v132;
+    float v134 = v111.GetValue(v17);
+    int32_t v135 = (int32_t) ((uint32_t) v56 + (uint32_t) v16);
+    v5[v135] = v134;
+    float v136 = v111.GetValue(v15);
+    int32_t v137 = (int32_t) ((uint32_t) v56 + (uint32_t) v14);
+    v5[v137] = v136;
+    float v138 = v111.GetValue(v13);
+    int32_t v139 = (int32_t) ((uint32_t) v56 + (uint32_t) v12);
+    v5[v139] = v138;
+    float v140 = v111.GetValue(v11);
+    int32_t v141 = (int32_t) ((uint32_t) v56 + (uint32_t) v10);
+    v5[v141] = v140;
+    float v142 = v111.GetValue(v9);
+    set_flag(PIPE_S, PIPE_V, EVENT_ID0);
+    int32_t v143 = (int32_t) ((uint32_t) v56 + (uint32_t) v8);
+    v5[v143] = v142;
+  }
+  wait_flag(PIPE_S, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: score_acc_buf_inline117__ssa_v0
+    __gm__ Tensor* score_acc_buf_inline117__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* score_acc_buf_inline117__ssa_v0 = reinterpret_cast<__gm__ float*>(score_acc_buf_inline117__ssa_v0_tensor->buffer.addr) + score_acc_buf_inline117__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: x_norm_bf16_inline111__rv_v2
+    __gm__ Tensor* x_norm_bf16_inline111__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ bfloat16_t* x_norm_bf16_inline111__rv_v2 = reinterpret_cast<__gm__ bfloat16_t*>(x_norm_bf16_inline111__rv_v2_tensor->buffer.addr) + x_norm_bf16_inline111__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: gate_w__ssa_v0
+    __gm__ Tensor* gate_w__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* gate_w__ssa_v0 = reinterpret_cast<__gm__ float*>(gate_w__ssa_v0_tensor->buffer.addr) + gate_w__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: gate_bias__ssa_v0
+    __gm__ Tensor* gate_bias__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* gate_bias__ssa_v0 = reinterpret_cast<__gm__ float*>(gate_bias__ssa_v0_tensor->buffer.addr) + gate_bias__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: biased_flat_inline133__ssa_v0
+    __gm__ Tensor* biased_flat_inline133__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ float* biased_flat_inline133__ssa_v0 = reinterpret_cast<__gm__ float*>(biased_flat_inline133__ssa_v0_tensor->buffer.addr) + biased_flat_inline133__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    gate_dot_0(score_acc_buf_inline117__ssa_v0, x_norm_bf16_inline111__rv_v2, gate_w__ssa_v0, gate_bias__ssa_v0, biased_flat_inline133__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/hc_post.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/hc_post.cpp
new file mode 100644
index 000000000..b4900a193
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/hc_post.cpp
@@ -0,0 +1,188 @@
+// Kernel Function: hc_post
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void hc_post(__gm__ bfloat16_t* v1, __gm__ float* v2, __gm__ bfloat16_t* v3, __gm__ float* v4, __gm__ bfloat16_t* v5, int32_t v6, int32_t v7) {
+  RoundMode v8 = RoundMode::CAST_ROUND;
+  unsigned v9 = 0;
+  const int32_t v10 = 512;
+  const int32_t v11 = 8;
+  const int32_t v12 = 4;
+  const int32_t v13 = 0;
+  const int32_t v14 = 4096;
+  const int32_t v15 = 1;
+  const int32_t v16 = 16384;
+  const int32_t v17 = 16;
+  const int64_t v18 = 3072;
+  const int64_t v19 = 1024;
+  const int64_t v20 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v21 = (size_t) v15;
+  size_t v22 = (size_t) v13;
+  set_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  for (size_t v23 = v22; v23 < ((size_t) v17); v23 += v21) {
+    int32_t v24 = (int32_t) ((uint32_t) ((int32_t) (uint32_t) v6 * (uint32_t) v17) + (uint32_t) ((int32_t) v23));
+    __gm__ bfloat16_t* v25;
+    if (v24 < v17) {
+      float v26 = v2[(int32_t) ((uint32_t) ((int32_t) (uint32_t) v24 * (uint32_t) v12) + (uint32_t) v7)];
+      for (size_t v27 = v22; v27 < ((size_t) v11); v27 += v21) {
+        int32_t v28 = (int32_t) ((uint32_t) ((int32_t) v27) * (uint32_t) v10);
+        Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v29 = Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+        uint64_t v30 = (uint64_t) v20;
+        TASSIGN(v29, v30);
+        pto::Shape<1, 1, 1, 1, 512> v31 = pto::Shape<1, 1, 1, 1, 512>();
+        pto::Stride<4096, 4096, 4096, 4096, 1> v32 = pto::Stride<4096, 4096, 4096, 4096, 1>();
+        GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND> v33 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND>(v3 + (v9 + (unsigned) v24 * (unsigned) v14 + (unsigned) v28 * (unsigned) v15), v31, v32);
+        wait_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+        TLOAD(v29, v33);
+        set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+        Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+        uint64_t v35 = (uint64_t) v19;
+        TASSIGN(v34, v35);
+        wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+        TCVT(v34, v29, v8);
+        set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+        Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+        uint64_t v37 = (uint64_t) v19;
+        TASSIGN(v36, v37);
+        pipe_barrier(PIPE_V);
+        TMULS(v36, v34, v26);
+        wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+        for (size_t v38 = v22; v38 < ((size_t) v12); v38 += v21) {
+          int32_t v39 = (int32_t) v38;
+          float v40 = v4[(int32_t) ((uint32_t) ((int32_t) (uint32_t) ((int32_t) (uint32_t) v24 * (uint32_t) v17) + (uint32_t) ((int32_t) (uint32_t) v39 * (uint32_t) v12)) + (uint32_t) v7)];
+          Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+          uint64_t v42 = (uint64_t) v20;
+          TASSIGN(v41, v42);
+          pto::Shape<1, 1, 1, 1, 512> v43 = pto::Shape<1, 1, 1, 1, 512>();
+          pto::Stride<16384, 16384, 16384, 16384, 1> v44 = pto::Stride<16384, 16384, 16384, 16384, 1>();
+          GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND> v45 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND>(v5 + (v9 + (unsigned) v24 * (unsigned) v16 + (unsigned) ((int32_t) (uint32_t) ((int32_t) (uint32_t) v39 * (uint32_t) v14) + (uint32_t) v28) * (unsigned) v15), v43, v44);
+          wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+          TLOAD(v41, v45);
+          set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+          Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v46 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+          uint64_t v47 = (uint64_t) v18;
+          TASSIGN(v46, v47);
+          wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+          pipe_barrier(PIPE_V);
+          TCVT(v46, v41, v8);
+          set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+          Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v48 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+          uint64_t v49 = (uint64_t) v18;
+          TASSIGN(v48, v49);
+          pipe_barrier(PIPE_V);
+          TMULS(v48, v46, v40);
+          Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+          uint64_t v51 = (uint64_t) v19;
+          TASSIGN(v50, v51);
+          pipe_barrier(PIPE_V);
+          TADD(v50, v36, v48);
+        };
+        Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v10);
+        uint64_t v53 = (uint64_t) v20;
+        TASSIGN(v52, v53);
+        pipe_barrier(PIPE_V);
+        TCVT(v52, v36, v8);
+        set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+        pto::Shape<1, 1, 1, 1, 512> v54 = pto::Shape<1, 1, 1, 1, 512>();
+        pto::Stride<16384, 16384, 16384, 16384, 1> v55 = pto::Stride<16384, 16384, 16384, 16384, 1>();
+        GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND> v56 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND>(v1 + (v9 + (unsigned) v24 * (unsigned) v16 + (unsigned) ((int32_t) (uint32_t) ((int32_t) (uint32_t) v7 * (uint32_t) v14) + (uint32_t) v28) * (unsigned) v15), v54, v55);
+        wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+        TSTORE(v56, v52);
+        set_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+      };
+      v25 = v1;
+    } else {
+      v25 = v1;
+    };
+  }
+  wait_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: y_flat_inline7__co_l0_iter_v3
+    __gm__ Tensor* y_flat_inline7__co_l0_iter_v3_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ bfloat16_t* y_flat_inline7__co_l0_iter_v3 = reinterpret_cast<__gm__ bfloat16_t*>(y_flat_inline7__co_l0_iter_v3_tensor->buffer.addr) + y_flat_inline7__co_l0_iter_v3_tensor->start_offset;
+
+    // Unpack tensor: post_flat_inline8__ssa_v0
+    __gm__ Tensor* post_flat_inline8__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* post_flat_inline8__ssa_v0 = reinterpret_cast<__gm__ float*>(post_flat_inline8__ssa_v0_tensor->buffer.addr) + post_flat_inline8__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: x_flat_inline2__ssa_v0
+    __gm__ Tensor* x_flat_inline2__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ bfloat16_t* x_flat_inline2__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(x_flat_inline2__ssa_v0_tensor->buffer.addr) + x_flat_inline2__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: comb_flat_inline9__ssa_v0
+    __gm__ Tensor* comb_flat_inline9__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* comb_flat_inline9__ssa_v0 = reinterpret_cast<__gm__ float*>(comb_flat_inline9__ssa_v0_tensor->buffer.addr) + comb_flat_inline9__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: residual_flat_inline6__ssa_v0
+    __gm__ Tensor* residual_flat_inline6__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ bfloat16_t* residual_flat_inline6__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(residual_flat_inline6__ssa_v0_tensor->buffer.addr) + residual_flat_inline6__ssa_v0_tensor->start_offset;
+
+    // Unpack scalar: t_inline10__co_idx_v0
+    union { uint64_t u64; int64_t val; } t_inline10__co_idx_v0_conv;
+    t_inline10__co_idx_v0_conv.u64 = args[5];
+    int64_t t_inline10__co_idx_v0 = t_inline10__co_idx_v0_conv.val;
+
+    // Unpack scalar: out_h_inline3__idx_v0
+    union { uint64_t u64; int64_t val; } out_h_inline3__idx_v0_conv;
+    out_h_inline3__idx_v0_conv.u64 = args[6];
+    int64_t out_h_inline3__idx_v0 = out_h_inline3__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    hc_post(y_flat_inline7__co_l0_iter_v3, post_flat_inline8__ssa_v0, x_flat_inline2__ssa_v0, comb_flat_inline9__ssa_v0, residual_flat_inline6__ssa_v0, t_inline10__co_idx_v0, out_h_inline3__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/linear.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/linear.cpp
new file mode 100644
index 000000000..e81e8a388
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/linear.cpp
@@ -0,0 +1,244 @@
+// Kernel Function: linear
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void linear(__gm__ float* v1, __gm__ float* v2, __gm__ float* v3, __gm__ float* v4) {
+  unsigned v5 = 0;
+  const int32_t v6 = 480;
+  const int32_t v7 = 15;
+  const int32_t v8 = 448;
+  const int32_t v9 = 14;
+  const int32_t v10 = 416;
+  const int32_t v11 = 13;
+  const int32_t v12 = 384;
+  const int32_t v13 = 12;
+  const int32_t v14 = 352;
+  const int32_t v15 = 11;
+  const int32_t v16 = 320;
+  const int32_t v17 = 10;
+  const int32_t v18 = 288;
+  const int32_t v19 = 9;
+  const int32_t v20 = 256;
+  const int32_t v21 = 8;
+  const int32_t v22 = 224;
+  const int32_t v23 = 7;
+  const int32_t v24 = 192;
+  const int32_t v25 = 6;
+  const int32_t v26 = 160;
+  const int32_t v27 = 5;
+  const int32_t v28 = 128;
+  const int32_t v29 = 4;
+  const int32_t v30 = 96;
+  const int32_t v31 = 3;
+  const int32_t v32 = 64;
+  const int32_t v33 = 2;
+  const int32_t v34 = 32;
+  const float v35 = 0.0f;
+  const int32_t v36 = 0;
+  const int32_t v37 = 512;
+  const int32_t v38 = 24;
+  const int32_t v39 = 1;
+  const int32_t v40 = 16384;
+  const int32_t v41 = 16;
+  const int64_t v42 = 0;
+  const int64_t v43 = 35008;
+  const int64_t v44 = 32960;
+  const int64_t v45 = 192;
+  const int64_t v46 = 128;
+  const int64_t v47 = 64;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v48 = (size_t) v39;
+  size_t v49 = (size_t) v36;
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v41);
+  uint64_t v51 = (uint64_t) v47;
+  TASSIGN(v50, v51);
+  pto::Shape<1, 1, 1, 1, 16> v52 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<16, 16, 16, 16, 1> v53 = pto::Stride<16, 16, 16, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND> v54 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND>(v3 + (v5 + v5 * (unsigned) v41 + v5 * (unsigned) v39), v52, v53);
+  set_flag(PIPE_S, PIPE_V, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v50, v54);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  for (size_t v55 = v49; v55 < ((size_t) v38); v55 += v48) {
+    int32_t v56 = (int32_t) v55;
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v57 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v41);
+    uint64_t v58 = (uint64_t) v46;
+    TASSIGN(v57, v58);
+    wait_flag(PIPE_S, PIPE_V, EVENT_ID0);
+    TEXPANDS(v57, v35);
+    for (size_t v59 = v49; v59 < ((size_t) v34); v59 += v48) {
+      int32_t v60 = (int32_t) ((uint32_t) ((int32_t) v59) * (uint32_t) v37);
+      Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v61 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v41, v37);
+      uint64_t v62 = (uint64_t) v45;
+      TASSIGN(v61, v62);
+      pto::Shape<1, 1, 1, 16, 512> v63 = pto::Shape<1, 1, 1, 16, 512>();
+      pto::Stride<262144, 262144, 262144, 16384, 1> v64 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+      GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v65 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v40 + (unsigned) v60 * (unsigned) v39), v63, v64);
+      wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+      TLOAD(v61, v65);
+      Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v37);
+      uint64_t v67 = (uint64_t) v44;
+      TASSIGN(v66, v67);
+      pto::Shape<1, 1, 1, 1, 512> v68 = pto::Shape<1, 1, 1, 1, 512>();
+      pto::Stride<16384, 16384, 16384, 16384, 1> v69 = pto::Stride<16384, 16384, 16384, 16384, 1>();
+      GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND> v70 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND>(v2 + (v5 + (unsigned) v56 * (unsigned) v40 + (unsigned) v60 * (unsigned) v39), v68, v69);
+      TLOAD(v66, v70);
+      set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+      Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v71 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v41, v37);
+      uint64_t v72 = (uint64_t) v45;
+      TASSIGN(v71, v72);
+      wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+      TCOLEXPANDMUL(v71, v61, v66);
+      Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v73 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v41, v37);
+      uint64_t v74 = (uint64_t) v43;
+      TASSIGN(v73, v74);
+      Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v75 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v41, v39);
+      uint64_t v76 = (uint64_t) v42;
+      TASSIGN(v75, v76);
+      pipe_barrier(PIPE_V);
+      TROWSUM(v75, v71, v73);
+      set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+      Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v77 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v41);
+      uint64_t v78 = (uint64_t) v42;
+      TASSIGN(v77, v78);
+      Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v79 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v41);
+      uint64_t v80 = (uint64_t) v46;
+      TASSIGN(v79, v80);
+      pipe_barrier(PIPE_V);
+      TADD(v79, v57, v77);
+    };
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v81 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v41);
+    uint64_t v82 = (uint64_t) v46;
+    TASSIGN(v81, v82);
+    pipe_barrier(PIPE_V);
+    TMUL(v81, v57, v50);
+    set_flag(PIPE_V, PIPE_S, EVENT_ID0);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v83 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v39, v41);
+    uint64_t v84 = (uint64_t) v46;
+    TASSIGN(v83, v84);
+    wait_flag(PIPE_V, PIPE_S, EVENT_ID0);
+    float v85 = v83.GetValue(v36);
+    v4[v55] = v85;
+    float v86 = v83.GetValue(v39);
+    int32_t v87 = (int32_t) ((uint32_t) v56 + (uint32_t) v34);
+    v4[v87] = v86;
+    float v88 = v83.GetValue(v33);
+    int32_t v89 = (int32_t) ((uint32_t) v56 + (uint32_t) v32);
+    v4[v89] = v88;
+    float v90 = v83.GetValue(v31);
+    int32_t v91 = (int32_t) ((uint32_t) v56 + (uint32_t) v30);
+    v4[v91] = v90;
+    float v92 = v83.GetValue(v29);
+    int32_t v93 = (int32_t) ((uint32_t) v56 + (uint32_t) v28);
+    v4[v93] = v92;
+    float v94 = v83.GetValue(v27);
+    int32_t v95 = (int32_t) ((uint32_t) v56 + (uint32_t) v26);
+    v4[v95] = v94;
+    float v96 = v83.GetValue(v25);
+    int32_t v97 = (int32_t) ((uint32_t) v56 + (uint32_t) v24);
+    v4[v97] = v96;
+    float v98 = v83.GetValue(v23);
+    int32_t v99 = (int32_t) ((uint32_t) v56 + (uint32_t) v22);
+    v4[v99] = v98;
+    float v100 = v83.GetValue(v21);
+    int32_t v101 = (int32_t) ((uint32_t) v56 + (uint32_t) v20);
+    v4[v101] = v100;
+    float v102 = v83.GetValue(v19);
+    int32_t v103 = (int32_t) ((uint32_t) v56 + (uint32_t) v18);
+    v4[v103] = v102;
+    float v104 = v83.GetValue(v17);
+    int32_t v105 = (int32_t) ((uint32_t) v56 + (uint32_t) v16);
+    v4[v105] = v104;
+    float v106 = v83.GetValue(v15);
+    int32_t v107 = (int32_t) ((uint32_t) v56 + (uint32_t) v14);
+    v4[v107] = v106;
+    float v108 = v83.GetValue(v13);
+    int32_t v109 = (int32_t) ((uint32_t) v56 + (uint32_t) v12);
+    v4[v109] = v108;
+    float v110 = v83.GetValue(v11);
+    int32_t v111 = (int32_t) ((uint32_t) v56 + (uint32_t) v10);
+    v4[v111] = v110;
+    float v112 = v83.GetValue(v9);
+    int32_t v113 = (int32_t) ((uint32_t) v56 + (uint32_t) v8);
+    v4[v113] = v112;
+    float v114 = v83.GetValue(v7);
+    set_flag(PIPE_S, PIPE_V, EVENT_ID0);
+    int32_t v115 = (int32_t) ((uint32_t) v56 + (uint32_t) v6);
+    v4[v115] = v114;
+  }
+  wait_flag(PIPE_S, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_flat_fp32_inline49__rv_v2
+    __gm__ Tensor* x_flat_fp32_inline49__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* x_flat_fp32_inline49__rv_v2 = reinterpret_cast<__gm__ float*>(x_flat_fp32_inline49__rv_v2_tensor->buffer.addr) + x_flat_fp32_inline49__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: hc_ffn_fn__ssa_v0
+    __gm__ Tensor* hc_ffn_fn__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* hc_ffn_fn__ssa_v0 = reinterpret_cast<__gm__ float*>(hc_ffn_fn__ssa_v0_tensor->buffer.addr) + hc_ffn_fn__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: inv_rms_inline75__ssa_v1
+    __gm__ Tensor* inv_rms_inline75__ssa_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* inv_rms_inline75__ssa_v1 = reinterpret_cast<__gm__ float*>(inv_rms_inline75__ssa_v1_tensor->buffer.addr) + inv_rms_inline75__ssa_v1_tensor->start_offset;
+
+    // Unpack tensor: mixes_flat_inline45__ssa_v0
+    __gm__ Tensor* mixes_flat_inline45__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* mixes_flat_inline45__ssa_v0 = reinterpret_cast<__gm__ float*>(mixes_flat_inline45__ssa_v0_tensor->buffer.addr) + mixes_flat_inline45__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    linear(x_flat_fp32_inline49__rv_v2, hc_ffn_fn__ssa_v0, inv_rms_inline75__ssa_v1, mixes_flat_inline45__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/local_expert.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/local_expert.cpp
deleted file mode 100644
index 2d519b9b2..000000000
--- a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/local_expert.cpp
+++ /dev/null
@@ -1,111 +0,0 @@
-/*
- * Copyright (c) PyPTO Contributors.
- * This program is free software, you can redistribute it and/or modify it under the terms and conditions of
- * CANN Open Software License Agreement Version 2.0 (the "License").
- * Please refer to the License for details. You may not use this file except in compliance with the License.
- * THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
- * INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
- * See LICENSE in the root of the software repository for the full text of the License.
- * -----------------------------------------------------------------------------------------------------------
- */
-/**
- * Local-expert kernel — placeholder for the production moe_expert in this
- * 2-rank demo. Replaced by the real moe_expert without changing the
- * surrounding orchestration / combine wiring.
- *
- * Behavior (one element-wise multiply per row):
- *   recv_y[e, s, :] = recv_x[e, s, :] * recv_w[e, s]      for s in [0, recv_count[e])
- *
- *   recv_x          : BF16  [L, R, D]   (dispatch OUTPUT_EXISTING, reused as INPUT)
- *   recv_w          : FP32  [L, R]      (dispatch OUTPUT_EXISTING, reused as INPUT)
- *   recv_count      : INT32 [L, 1]      (dispatch OUTPUT_EXISTING, reused as INPUT)
- *   recv_y          : BF16  [L, R, D]   (this kernel's OUTPUT_EXISTING)
- *
- * Per-expert n_rows = recv_count[e] decides how many rows to process. We
- * skip padding rows entirely — they stay whatever recv_y was previously
- * initialized to, but combine reads them only via pub_counts-driven slabs
- * that never reach into padding, so the value doesn't matter.
- *
- * Pure local — no CommRemotePtr, no signals, no scratch. BF16 round-trip
- * happens once per row at `cast(x*w, bf16)`.
- */
-
-#ifndef __gm__
-#define __gm__
-#endif
-
-#ifndef __aicore__
-#define __aicore__ [aicore]
-#endif
-
-#include <cstdint>
-
-#include <pto/pto-inst.hpp>
-#include "platform_comm/comm_context.h"
-#include "tensor.h"
-
-using namespace pto;
-
-// Demo dimensions — must match dispatch.cpp / main.py.
-static constexpr int L = 4;
-static constexpr int R = 32;
-static constexpr int D = 64;
-
-extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t *args) {
-    __gm__ Tensor *recv_x_tensor = reinterpret_cast<__gm__ Tensor *>(args[0]);
-    __gm__ Tensor *recv_w_tensor = reinterpret_cast<__gm__ Tensor *>(args[1]);
-    __gm__ Tensor *recv_count_tensor = reinterpret_cast<__gm__ Tensor *>(args[2]);
-    __gm__ Tensor *recv_y_tensor = reinterpret_cast<__gm__ Tensor *>(args[3]);
-    __gm__ CommContext *comm_ctx = reinterpret_cast<__gm__ CommContext *>(args[4]);
-    (void)comm_ctx;  // unused; kept for ABI symmetry with dispatch / combine
-
-    __gm__ bfloat16_t *recv_x =
-        reinterpret_cast<__gm__ bfloat16_t *>(recv_x_tensor->buffer.addr) + recv_x_tensor->start_offset;
-    __gm__ float *recv_w = reinterpret_cast<__gm__ float *>(recv_w_tensor->buffer.addr) + recv_w_tensor->start_offset;
-    __gm__ int32_t *recv_count =
-        reinterpret_cast<__gm__ int32_t *>(recv_count_tensor->buffer.addr) + recv_count_tensor->start_offset;
-    __gm__ bfloat16_t *recv_y =
-        reinterpret_cast<__gm__ bfloat16_t *>(recv_y_tensor->buffer.addr) + recv_y_tensor->start_offset;
-
-    using XShape = Shape<1, 1, 1, 1, D>;
-    using XStride = Stride<D, D, D, D, 1>;
-    using XGlobalBF = GlobalTensor<bfloat16_t, XShape, XStride>;
-    using XTileBF = Tile<TileType::Vec, bfloat16_t, 1, D, BLayout::RowMajor>;
-    using XTileF = Tile<TileType::Vec, float, 1, D, BLayout::RowMajor>;
-
-    XTileBF x_bf;
-    XTileF x_f;
-    TASSIGN(x_bf, 0x0);
-    TASSIGN(x_f, 0x10000);
-
-    for (int e = 0; e < L; ++e) {
-        int n_rows = recv_count[e];
-        for (int slot = 0; slot < n_rows; ++slot) {
-            int row = e * R + slot;
-            float w = recv_w[row];
-
-            __gm__ bfloat16_t *x_src = recv_x + row * D;
-            __gm__ bfloat16_t *y_dst = recv_y + row * D;
-            XGlobalBF x_g(x_src);
-            XGlobalBF y_g(y_dst);
-
-            TLOAD(x_bf, x_g);
-            set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
-            wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
-
-            TCVT(x_f, x_bf, RoundMode::CAST_ROUND);
-            pipe_barrier(PIPE_V);
-            TMULS(x_f, x_f, w);
-            pipe_barrier(PIPE_V);
-            TCVT(x_bf, x_f, RoundMode::CAST_ROUND);
-            pipe_barrier(PIPE_V);
-
-            set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
-            wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
-            TSTORE(y_g, x_bf);
-            set_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
-            wait_flag(PIPE_MTE3, PIPE_MTE2, EVENT_ID0);
-        }
-    }
-    pipe_barrier(PIPE_ALL);
-}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/mix_x.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/mix_x.cpp
new file mode 100644
index 000000000..46e1605c4
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/mix_x.cpp
@@ -0,0 +1,154 @@
+// Kernel Function: mix_x
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void mix_x(__gm__ bfloat16_t* v1, __gm__ float* v2, __gm__ float* v3, int32_t v4) {
+  RoundMode v5 = RoundMode::CAST_RINT;
+  unsigned v6 = 0;
+  const int32_t v7 = 4;
+  const float v8 = 0.0f;
+  const int32_t v9 = 512;
+  const int32_t v10 = 8;
+  const int32_t v11 = 0;
+  const int32_t v12 = 16384;
+  const int32_t v13 = 1;
+  const int32_t v14 = 4096;
+  const int32_t v15 = 16;
+  const int64_t v16 = 4096;
+  const int64_t v17 = 2048;
+  const int64_t v18 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v19 = (size_t) v13;
+  size_t v20 = (size_t) v11;
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  for (size_t v21 = v20; v21 < ((size_t) v15); v21 += v19) {
+    int32_t v22 = (int32_t) ((uint32_t) ((int32_t) (uint32_t) v4 * (uint32_t) v15) + (uint32_t) ((int32_t) v21));
+    __gm__ bfloat16_t* v23;
+    if (v22 < v15) {
+      for (size_t v24 = v20; v24 < ((size_t) v10); v24 += v19) {
+        int32_t v25 = (int32_t) ((uint32_t) ((int32_t) v24) * (uint32_t) v9);
+        Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v9);
+        uint64_t v27 = (uint64_t) v18;
+        TASSIGN(v26, v27);
+        pipe_barrier(PIPE_V);
+        TEXPANDS(v26, v8);
+        for (size_t v28 = v20; v28 < ((size_t) v7); v28 += v19) {
+          int32_t v29 = (int32_t) v28;
+          float v30 = v2[(int32_t) ((uint32_t) ((int32_t) (uint32_t) v22 * (uint32_t) v10) + (uint32_t) v29)];
+          Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v31 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v9);
+          uint64_t v32 = (uint64_t) v17;
+          TASSIGN(v31, v32);
+          pto::Shape<1, 1, 1, 1, 512> v33 = pto::Shape<1, 1, 1, 1, 512>();
+          pto::Stride<16384, 16384, 16384, 16384, 1> v34 = pto::Stride<16384, 16384, 16384, 16384, 1>();
+          GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND> v35 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<16384, 16384, 16384, 16384, 1>, pto::Layout::ND>(v3 + (v6 + (unsigned) v22 * (unsigned) v12 + (unsigned) ((int32_t) (uint32_t) ((int32_t) (uint32_t) v29 * (uint32_t) v14) + (uint32_t) v25) * (unsigned) v13), v33, v34);
+          wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+          TLOAD(v31, v35);
+          set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+          Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v9);
+          uint64_t v37 = (uint64_t) v17;
+          TASSIGN(v36, v37);
+          wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+          TMULS(v36, v31, v30);
+          Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v9);
+          uint64_t v39 = (uint64_t) v18;
+          TASSIGN(v38, v39);
+          pipe_barrier(PIPE_V);
+          TADD(v38, v26, v36);
+          set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+        };
+        Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v40 = Tile<TileType::Vec, bfloat16_t, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v9);
+        uint64_t v41 = (uint64_t) v16;
+        TASSIGN(v40, v41);
+        pipe_barrier(PIPE_V);
+        wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+        TCVT(v40, v26, v5);
+        set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+        pto::Shape<1, 1, 1, 1, 512> v42 = pto::Shape<1, 1, 1, 1, 512>();
+        pto::Stride<4096, 4096, 4096, 4096, 1> v43 = pto::Stride<4096, 4096, 4096, 4096, 1>();
+        GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND> v44 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<4096, 4096, 4096, 4096, 1>, pto::Layout::ND>(v1 + (v6 + (unsigned) v22 * (unsigned) v14 + (unsigned) v25 * (unsigned) v13), v42, v43);
+        wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+        TSTORE(v44, v40);
+        set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+      };
+      v23 = v1;
+    } else {
+      v23 = v1;
+    };
+  }
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_mixed_view_inline55__co_l0_iter_v1
+    __gm__ Tensor* x_mixed_view_inline55__co_l0_iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ bfloat16_t* x_mixed_view_inline55__co_l0_iter_v1 = reinterpret_cast<__gm__ bfloat16_t*>(x_mixed_view_inline55__co_l0_iter_v1_tensor->buffer.addr) + x_mixed_view_inline55__co_l0_iter_v1_tensor->start_offset;
+
+    // Unpack tensor: pre_val_flat_inline44__ssa_v0
+    __gm__ Tensor* pre_val_flat_inline44__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* pre_val_flat_inline44__ssa_v0 = reinterpret_cast<__gm__ float*>(pre_val_flat_inline44__ssa_v0_tensor->buffer.addr) + pre_val_flat_inline44__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: x_flat_fp32_inline49__rv_v2
+    __gm__ Tensor* x_flat_fp32_inline49__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* x_flat_fp32_inline49__rv_v2 = reinterpret_cast<__gm__ float*>(x_flat_fp32_inline49__rv_v2_tensor->buffer.addr) + x_flat_fp32_inline49__rv_v2_tensor->start_offset;
+
+    // Unpack scalar: t_inline5__co_idx_v0
+    union { uint64_t u64; int64_t val; } t_inline5__co_idx_v0_conv;
+    t_inline5__co_idx_v0_conv.u64 = args[3];
+    int64_t t_inline5__co_idx_v0 = t_inline5__co_idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    mix_x(x_mixed_view_inline55__co_l0_iter_v1, pre_val_flat_inline44__ssa_v0, x_flat_fp32_inline49__rv_v2, t_inline5__co_idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/recv_x_q.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/recv_x_q.cpp
new file mode 100644
index 000000000..bbe47ab23
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/recv_x_q.cpp
@@ -0,0 +1,241 @@
+// Kernel Function: recv_x_q
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void recv_x_q(__gm__ bfloat16_t* v1, __gm__ int8_t* v2, __gm__ float* v3, int32_t v4, int32_t v5) {
+  RoundMode v6 = RoundMode::CAST_TRUNC;
+  RoundMode v7 = RoundMode::CAST_ROUND;
+  unsigned v8 = 0;
+  const int32_t v9 = 131072;
+  const float v10 = 127.0f;
+  const int32_t v11 = 256;
+  const int32_t v12 = 0;
+  const float v13 = 9.99999974E-5f;
+  const int32_t v14 = 16;
+  const int32_t v15 = 1;
+  const int32_t v16 = 4096;
+  const int64_t v17 = 24640;
+  const int64_t v18 = 16448;
+  const int64_t v19 = 64;
+  const int64_t v20 = 0;
+  const int64_t v21 = 69760;
+  const int64_t v22 = 53376;
+  const int64_t v23 = 36992;
+  const int64_t v24 = 28800;
+  const int64_t v25 = 28736;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v26 = (size_t) v16;
+  size_t v27 = (size_t) v12;
+  size_t v28 = (size_t) v11;
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v29 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v14);
+  uint64_t v30 = (uint64_t) v25;
+  TASSIGN(v29, v30);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TEXPANDS(v29, v13);
+  for (size_t v31 = v27; v31 < v26; v31 += v28) {
+    Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v32 = Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v33 = (uint64_t) v24;
+    TASSIGN(v32, v33);
+    pto::Shape<1, 1, 1, 16, 256> v34 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<131072, 131072, 131072, 4096, 1> v35 = pto::Stride<131072, 131072, 131072, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<131072, 131072, 131072, 4096, 1>, pto::Layout::ND> v36 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<131072, 131072, 131072, 4096, 1>, pto::Layout::ND>(v1 + ((v8 + (unsigned) v4 * (unsigned) v9) + (unsigned) v5 * (unsigned) v16 + (unsigned) ((int32_t) v31) * (unsigned) v15), v34, v35);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v32, v36);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v38 = (uint64_t) v23;
+    TASSIGN(v37, v38);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    TCVT(v37, v32, v7);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v40 = (uint64_t) v23;
+    TASSIGN(v39, v40);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v42 = (uint64_t) v22;
+    TASSIGN(v41, v42);
+    pipe_barrier(PIPE_V);
+    TNEG(v41, v39);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v44 = (uint64_t) v23;
+    TASSIGN(v43, v44);
+    pipe_barrier(PIPE_V);
+    TMAX(v43, v39, v41);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v46 = (uint64_t) v22;
+    TASSIGN(v45, v46);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v15);
+    uint64_t v48 = (uint64_t) v21;
+    TASSIGN(v47, v48);
+    pipe_barrier(PIPE_V);
+    TROWMAX(v47, v43, v45);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v49 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v14);
+    uint64_t v50 = (uint64_t) v21;
+    TASSIGN(v49, v50);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v51 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v14);
+    uint64_t v52 = (uint64_t) v25;
+    TASSIGN(v51, v52);
+    pipe_barrier(PIPE_V);
+    TMAX(v51, v29, v49);
+  }
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v53 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v14);
+  uint64_t v54 = (uint64_t) v20;
+  TASSIGN(v53, v54);
+  TEXPANDS(v53, v10);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v55 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v14);
+  uint64_t v56 = (uint64_t) v25;
+  TASSIGN(v55, v56);
+  pipe_barrier(PIPE_V);
+  TDIV(v55, v53, v29);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v57 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v14);
+  uint64_t v58 = (uint64_t) v20;
+  TASSIGN(v57, v58);
+  pipe_barrier(PIPE_V);
+  TRECIP(v57, v55);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v59 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v15);
+  uint64_t v60 = (uint64_t) v20;
+  TASSIGN(v59, v60);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v61 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v15);
+  uint64_t v62 = (uint64_t) v25;
+  TASSIGN(v61, v62);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  for (size_t v63 = v27; v63 < v26; v63 += v28) {
+    int32_t v64 = (int32_t) v63;
+    Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v65 = Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v66 = (uint64_t) v24;
+    TASSIGN(v65, v66);
+    pto::Shape<1, 1, 1, 16, 256> v67 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<131072, 131072, 131072, 4096, 1> v68 = pto::Stride<131072, 131072, 131072, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<131072, 131072, 131072, 4096, 1>, pto::Layout::ND> v69 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<131072, 131072, 131072, 4096, 1>, pto::Layout::ND>(v1 + ((v8 + (unsigned) v4 * (unsigned) v9) + (unsigned) v5 * (unsigned) v16 + (unsigned) v64 * (unsigned) v15), v67, v68);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    TLOAD(v65, v69);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v70 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v71 = (uint64_t) v23;
+    TASSIGN(v70, v71);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    TCVT(v70, v65, v7);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v72 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v73 = (uint64_t) v23;
+    TASSIGN(v72, v73);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v74 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v75 = (uint64_t) v23;
+    TASSIGN(v74, v75);
+    pipe_barrier(PIPE_V);
+    TROWEXPANDMUL(v74, v72, v61);
+    Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v76 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v77 = (uint64_t) v19;
+    TASSIGN(v76, v77);
+    pipe_barrier(PIPE_V);
+    TCVT(v76, v74, v7);
+    Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v78 = Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v79 = (uint64_t) v18;
+    TASSIGN(v78, v79);
+    pipe_barrier(PIPE_V);
+    TCVT(v78, v76, v7);
+    Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v80 = Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v14, v11);
+    uint64_t v81 = (uint64_t) v17;
+    TASSIGN(v80, v81);
+    pipe_barrier(PIPE_V);
+    wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+    TCVT(v80, v78, v6);
+    set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    pto::Shape<1, 1, 1, 16, 256> v82 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v83 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v84 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v8 + v8 * (unsigned) v16 + (unsigned) v64 * (unsigned) v15), v82, v83);
+    wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    TSTORE(v84, v80);
+    set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  }
+  pto::Shape<1, 1, 1, 16, 1> v85 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v86 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v87 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v3 + (v8 + v8 * (unsigned) v15 + v8 * (unsigned) v14), v85, v86);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v87, v59);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: recv_x__ssa_v0
+    __gm__ Tensor* recv_x__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ bfloat16_t* recv_x__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(recv_x__ssa_v0_tensor->buffer.addr) + recv_x__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: recv_x_tile_i8_inline75__ssa_v0
+    __gm__ Tensor* recv_x_tile_i8_inline75__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* recv_x_tile_i8_inline75__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(recv_x_tile_i8_inline75__ssa_v0_tensor->buffer.addr) + recv_x_tile_i8_inline75__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: local_i_inline67__idx_v0
+    union { uint64_t u64; int64_t val; } local_i_inline67__idx_v0_conv;
+    local_i_inline67__idx_v0_conv.u64 = args[3];
+    int64_t local_i_inline67__idx_v0 = local_i_inline67__idx_v0_conv.val;
+
+    // Unpack scalar: t0_inline47__ssa_v0
+    union { uint64_t u64; int64_t val; } t0_inline47__ssa_v0_conv;
+    t0_inline47__ssa_v0_conv.u64 = args[4];
+    int64_t t0_inline47__ssa_v0 = t0_inline47__ssa_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    recv_x_q(recv_x__ssa_v0, recv_x_tile_i8_inline75__ssa_v0, ret0__out, local_i_inline67__idx_v0, t0_inline47__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/rms.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/rms.cpp
new file mode 100644
index 000000000..3ae1379a6
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/rms.cpp
@@ -0,0 +1,245 @@
+// Kernel Function: rms
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void rms(__gm__ float* v1, __gm__ float* v2) {
+  unsigned v3 = 0;
+  const float v4 = 9.99999997E-7f;
+  const float v5 = 6.10351563E-5f;
+  const int32_t v6 = 1536;
+  const int32_t v7 = 1024;
+  const int32_t v8 = 512;
+  const int32_t v9 = 4;
+  const int32_t v10 = 32;
+  const int32_t v11 = 0;
+  const float v12 = 0.0f;
+  const int32_t v13 = 1;
+  const int32_t v14 = 16384;
+  const int32_t v15 = 16;
+  const int64_t v16 = 0;
+  const int64_t v17 = 163968;
+  const int64_t v18 = 131200;
+  const int64_t v19 = 98432;
+  const int64_t v20 = 65664;
+  const int64_t v21 = 32896;
+  const int64_t v22 = 128;
+  const int64_t v23 = 64;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v24 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+  uint64_t v25 = (uint64_t) v23;
+  TASSIGN(v24, v25);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TEXPANDS(v24, v12);
+  for (size_t v26 = (size_t) v11; v26 < ((size_t) v10); v26 += (size_t) v9) {
+    int32_t v27 = (int32_t) ((uint32_t) ((int32_t) v26) * (uint32_t) v8);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v29 = (uint64_t) v22;
+    TASSIGN(v28, v29);
+    pto::Shape<1, 1, 1, 16, 512> v30 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<262144, 262144, 262144, 16384, 1> v31 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v32 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v1 + (v3 + v3 * (unsigned) v14 + (unsigned) v27 * (unsigned) v13), v30, v31);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v28, v32);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v34 = (uint64_t) v21;
+    TASSIGN(v33, v34);
+    pto::Shape<1, 1, 1, 16, 512> v35 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<262144, 262144, 262144, 16384, 1> v36 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v37 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v1 + (v3 + v3 * (unsigned) v14 + (unsigned) ((int32_t) (uint32_t) v27 + (uint32_t) v8) * (unsigned) v13), v35, v36);
+    TLOAD(v33, v37);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v39 = (uint64_t) v20;
+    TASSIGN(v38, v39);
+    pto::Shape<1, 1, 1, 16, 512> v40 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<262144, 262144, 262144, 16384, 1> v41 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v42 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v1 + (v3 + v3 * (unsigned) v14 + (unsigned) ((int32_t) (uint32_t) v27 + (uint32_t) v7) * (unsigned) v13), v40, v41);
+    TLOAD(v38, v42);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v44 = (uint64_t) v19;
+    TASSIGN(v43, v44);
+    pto::Shape<1, 1, 1, 16, 512> v45 = pto::Shape<1, 1, 1, 16, 512>();
+    pto::Stride<262144, 262144, 262144, 16384, 1> v46 = pto::Stride<262144, 262144, 262144, 16384, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND> v47 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<262144, 262144, 262144, 16384, 1>, pto::Layout::ND>(v1 + (v3 + v3 * (unsigned) v14 + (unsigned) ((int32_t) (uint32_t) v27 + (uint32_t) v6) * (unsigned) v13), v45, v46);
+    TLOAD(v43, v47);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v48 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v49 = (uint64_t) v22;
+    TASSIGN(v48, v49);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    TMUL(v48, v28, v28);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v51 = (uint64_t) v18;
+    TASSIGN(v50, v51);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v13);
+    uint64_t v53 = (uint64_t) v17;
+    TASSIGN(v52, v53);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v52, v48, v50);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v55 = (uint64_t) v17;
+    TASSIGN(v54, v55);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v57 = (uint64_t) v16;
+    TASSIGN(v56, v57);
+    pipe_barrier(PIPE_V);
+    TADD(v56, v24, v54);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v59 = (uint64_t) v22;
+    TASSIGN(v58, v59);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    TMUL(v58, v33, v33);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v60 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v61 = (uint64_t) v21;
+    TASSIGN(v60, v61);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v13);
+    uint64_t v63 = (uint64_t) v17;
+    TASSIGN(v62, v63);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v62, v58, v60);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v65 = (uint64_t) v17;
+    TASSIGN(v64, v65);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v67 = (uint64_t) v16;
+    TASSIGN(v66, v67);
+    pipe_barrier(PIPE_V);
+    TADD(v66, v56, v64);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v68 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v69 = (uint64_t) v22;
+    TASSIGN(v68, v69);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+    TMUL(v68, v38, v38);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v70 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v71 = (uint64_t) v21;
+    TASSIGN(v70, v71);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v72 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v13);
+    uint64_t v73 = (uint64_t) v17;
+    TASSIGN(v72, v73);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v72, v68, v70);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v74 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v75 = (uint64_t) v17;
+    TASSIGN(v74, v75);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v76 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v77 = (uint64_t) v16;
+    TASSIGN(v76, v77);
+    pipe_barrier(PIPE_V);
+    TADD(v76, v66, v74);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v78 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v79 = (uint64_t) v22;
+    TASSIGN(v78, v79);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+    TMUL(v78, v43, v43);
+    Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v80 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v8);
+    uint64_t v81 = (uint64_t) v21;
+    TASSIGN(v80, v81);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v82 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v15, v13);
+    uint64_t v83 = (uint64_t) v17;
+    TASSIGN(v82, v83);
+    pipe_barrier(PIPE_V);
+    TROWSUM(v82, v78, v80);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v84 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v85 = (uint64_t) v17;
+    TASSIGN(v84, v85);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v86 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+    uint64_t v87 = (uint64_t) v23;
+    TASSIGN(v86, v87);
+    pipe_barrier(PIPE_V);
+    TADD(v86, v76, v84);
+  }
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v88 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+  uint64_t v89 = (uint64_t) v23;
+  TASSIGN(v88, v89);
+  pipe_barrier(PIPE_V);
+  TMULS(v88, v24, v5);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v90 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+  uint64_t v91 = (uint64_t) v23;
+  TASSIGN(v90, v91);
+  pipe_barrier(PIPE_V);
+  TADDS(v90, v88, v4);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v92 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+  uint64_t v93 = (uint64_t) v23;
+  TASSIGN(v92, v93);
+  pipe_barrier(PIPE_V);
+  TSQRT(v92, v90);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v94 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v15);
+  uint64_t v95 = (uint64_t) v16;
+  TASSIGN(v94, v95);
+  pipe_barrier(PIPE_V);
+  TRECIP(v94, v92);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v96 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<16, 16, 16, 16, 1> v97 = pto::Stride<16, 16, 16, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND> v98 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND>(v2 + (v3 + v3 * (unsigned) v15 + v3 * (unsigned) v13), v96, v97);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v98, v94);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_flat_fp32_inline49__rv_v2
+    __gm__ Tensor* x_flat_fp32_inline49__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* x_flat_fp32_inline49__rv_v2 = reinterpret_cast<__gm__ float*>(x_flat_fp32_inline49__rv_v2_tensor->buffer.addr) + x_flat_fp32_inline49__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    rms(x_flat_fp32_inline49__rv_v2, ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_extract_top2.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_extract_top2.cpp
new file mode 100644
index 000000000..1a1fe0447
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_extract_top2.cpp
@@ -0,0 +1,796 @@
+// Kernel Function: route_extract_top2
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void route_extract_top2(__gm__ float* v1, __gm__ float* v2, __gm__ int32_t* v3) {
+  unsigned v4 = 15;
+  unsigned v5 = 14;
+  unsigned v6 = 13;
+  unsigned v7 = 12;
+  unsigned v8 = 11;
+  unsigned v9 = 10;
+  unsigned v10 = 9;
+  unsigned v11 = 8;
+  unsigned v12 = 7;
+  unsigned v13 = 6;
+  unsigned v14 = 5;
+  unsigned v15 = 4;
+  unsigned v16 = 3;
+  unsigned v17 = 2;
+  unsigned v18 = 1;
+  unsigned v19 = 0;
+  const int32_t v20 = 2;
+  const float v21 = 0.0f;
+  const int32_t v22 = 64;
+  const int32_t v23 = 1;
+  const int32_t v24 = 32;
+  const int32_t v25 = 16;
+  const int64_t v26 = 2304;
+  const int64_t v27 = 2240;
+  const int64_t v28 = 2176;
+  const int64_t v29 = 2048;
+  const int64_t v30 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v31 = Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v25, v24);
+  uint64_t v32 = (uint64_t) v30;
+  TASSIGN(v31, v32);
+  TEXPANDS(v31, v21);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 32> v33 = pto::Shape<1, 1, 1, 16, 32>();
+  pto::Stride<512, 512, 512, 32, 1> v34 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v35 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v1 + (v19 + v19 * (unsigned) v24 + v19 * (unsigned) v23), v33, v34);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v35, v31);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v37 = (uint64_t) v29;
+  TASSIGN(v36, v37);
+  pto::Shape<1, 1, 1, 1, 32> v38 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v39 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v40 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v19 * (unsigned) v22 + v19 * (unsigned) v23), v38, v39);
+  TLOAD(v36, v40);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v42 = (uint64_t) v28;
+  TASSIGN(v41, v42);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v41, v36);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v44 = (uint64_t) v27;
+  TASSIGN(v43, v44);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v43, v36);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v46 = (uint64_t) v28;
+  TASSIGN(v45, v46);
+  pipe_barrier(PIPE_V);
+  TMOV(v45, v41);
+  v45.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v47 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v48 = (uint64_t) v26;
+  TASSIGN(v47, v48);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v47, v45);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  pto::Shape<1, 1, 1, 1, 16> v49 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v50 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v51 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v19 * (unsigned) v24 + v19 * (unsigned) v23), v49, v50);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  pipe_barrier(PIPE_MTE3);
+  TSTORE(v51, v47);
+  pto::Shape<1, 1, 1, 1, 16> v52 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v53 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v54 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v19 * (unsigned) v24 + v19 * (unsigned) v23), v52, v53);
+  TSTORE(v54, v43);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v55 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v56 = (uint64_t) v29;
+  TASSIGN(v55, v56);
+  pto::Shape<1, 1, 1, 1, 32> v57 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v58 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v59 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v18 * (unsigned) v22 + v19 * (unsigned) v23), v57, v58);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v55, v59);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v60 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v61 = (uint64_t) v28;
+  TASSIGN(v60, v61);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v60, v55);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v63 = (uint64_t) v27;
+  TASSIGN(v62, v63);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v62, v55);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v65 = (uint64_t) v28;
+  TASSIGN(v64, v65);
+  pipe_barrier(PIPE_V);
+  TMOV(v64, v60);
+  v64.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v66 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v67 = (uint64_t) v26;
+  TASSIGN(v66, v67);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v66, v64);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID2);
+  pto::Shape<1, 1, 1, 1, 16> v68 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v69 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v70 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v18 * (unsigned) v24 + v19 * (unsigned) v23), v68, v69);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID2);
+  TSTORE(v70, v66);
+  pto::Shape<1, 1, 1, 1, 16> v71 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v72 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v73 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v18 * (unsigned) v24 + v19 * (unsigned) v23), v71, v72);
+  TSTORE(v73, v62);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v74 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v75 = (uint64_t) v29;
+  TASSIGN(v74, v75);
+  pto::Shape<1, 1, 1, 1, 32> v76 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v77 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v78 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v17 * (unsigned) v22 + v19 * (unsigned) v23), v76, v77);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  TLOAD(v74, v78);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v79 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v80 = (uint64_t) v28;
+  TASSIGN(v79, v80);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v79, v74);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v81 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v82 = (uint64_t) v27;
+  TASSIGN(v81, v82);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID1);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v81, v74);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v83 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v84 = (uint64_t) v28;
+  TASSIGN(v83, v84);
+  pipe_barrier(PIPE_V);
+  TMOV(v83, v79);
+  v83.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v85 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v86 = (uint64_t) v26;
+  TASSIGN(v85, v86);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v85, v83);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID3);
+  pto::Shape<1, 1, 1, 1, 16> v87 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v88 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v89 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v17 * (unsigned) v24 + v19 * (unsigned) v23), v87, v88);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID3);
+  TSTORE(v89, v85);
+  pto::Shape<1, 1, 1, 1, 16> v90 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v91 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v92 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v17 * (unsigned) v24 + v19 * (unsigned) v23), v90, v91);
+  TSTORE(v92, v81);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v93 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v94 = (uint64_t) v29;
+  TASSIGN(v93, v94);
+  pto::Shape<1, 1, 1, 1, 32> v95 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v96 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v97 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v16 * (unsigned) v22 + v19 * (unsigned) v23), v95, v96);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  TLOAD(v93, v97);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v98 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v99 = (uint64_t) v28;
+  TASSIGN(v98, v99);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v98, v93);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v100 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v101 = (uint64_t) v27;
+  TASSIGN(v100, v101);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID2);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v100, v93);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v102 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v103 = (uint64_t) v28;
+  TASSIGN(v102, v103);
+  pipe_barrier(PIPE_V);
+  TMOV(v102, v98);
+  v102.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v104 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v105 = (uint64_t) v26;
+  TASSIGN(v104, v105);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v104, v102);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID4);
+  pto::Shape<1, 1, 1, 1, 16> v106 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v107 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v108 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v16 * (unsigned) v24 + v19 * (unsigned) v23), v106, v107);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID4);
+  TSTORE(v108, v104);
+  pto::Shape<1, 1, 1, 1, 16> v109 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v110 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v111 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v16 * (unsigned) v24 + v19 * (unsigned) v23), v109, v110);
+  TSTORE(v111, v100);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v112 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v113 = (uint64_t) v29;
+  TASSIGN(v112, v113);
+  pto::Shape<1, 1, 1, 1, 32> v114 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v115 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v116 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v15 * (unsigned) v22 + v19 * (unsigned) v23), v114, v115);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+  TLOAD(v112, v116);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v117 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v118 = (uint64_t) v28;
+  TASSIGN(v117, v118);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v117, v112);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v119 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v120 = (uint64_t) v27;
+  TASSIGN(v119, v120);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID3);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v119, v112);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID4);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v121 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v122 = (uint64_t) v28;
+  TASSIGN(v121, v122);
+  pipe_barrier(PIPE_V);
+  TMOV(v121, v117);
+  v121.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v123 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v124 = (uint64_t) v26;
+  TASSIGN(v123, v124);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v123, v121);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID5);
+  pto::Shape<1, 1, 1, 1, 16> v125 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v126 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v127 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v15 * (unsigned) v24 + v19 * (unsigned) v23), v125, v126);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID5);
+  TSTORE(v127, v123);
+  pto::Shape<1, 1, 1, 1, 16> v128 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v129 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v130 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v15 * (unsigned) v24 + v19 * (unsigned) v23), v128, v129);
+  TSTORE(v130, v119);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID4);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v131 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v132 = (uint64_t) v29;
+  TASSIGN(v131, v132);
+  pto::Shape<1, 1, 1, 1, 32> v133 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v134 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v135 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v14 * (unsigned) v22 + v19 * (unsigned) v23), v133, v134);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID4);
+  TLOAD(v131, v135);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID5);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v136 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v137 = (uint64_t) v28;
+  TASSIGN(v136, v137);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID5);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v136, v131);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v138 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v139 = (uint64_t) v27;
+  TASSIGN(v138, v139);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID4);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v138, v131);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID5);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v140 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v141 = (uint64_t) v28;
+  TASSIGN(v140, v141);
+  pipe_barrier(PIPE_V);
+  TMOV(v140, v136);
+  v140.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v142 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v143 = (uint64_t) v26;
+  TASSIGN(v142, v143);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v142, v140);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID6);
+  pto::Shape<1, 1, 1, 1, 16> v144 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v145 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v146 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v14 * (unsigned) v24 + v19 * (unsigned) v23), v144, v145);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID6);
+  TSTORE(v146, v142);
+  pto::Shape<1, 1, 1, 1, 16> v147 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v148 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v149 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v14 * (unsigned) v24 + v19 * (unsigned) v23), v147, v148);
+  TSTORE(v149, v138);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID5);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v150 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v151 = (uint64_t) v29;
+  TASSIGN(v150, v151);
+  pto::Shape<1, 1, 1, 1, 32> v152 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v153 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v154 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v13 * (unsigned) v22 + v19 * (unsigned) v23), v152, v153);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID5);
+  TLOAD(v150, v154);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID6);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v155 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v156 = (uint64_t) v28;
+  TASSIGN(v155, v156);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID6);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v155, v150);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v157 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v158 = (uint64_t) v27;
+  TASSIGN(v157, v158);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID5);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v157, v150);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID6);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v159 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v160 = (uint64_t) v28;
+  TASSIGN(v159, v160);
+  pipe_barrier(PIPE_V);
+  TMOV(v159, v155);
+  v159.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v161 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v162 = (uint64_t) v26;
+  TASSIGN(v161, v162);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v161, v159);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID7);
+  pto::Shape<1, 1, 1, 1, 16> v163 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v164 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v165 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v13 * (unsigned) v24 + v19 * (unsigned) v23), v163, v164);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID7);
+  TSTORE(v165, v161);
+  pto::Shape<1, 1, 1, 1, 16> v166 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v167 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v168 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v13 * (unsigned) v24 + v19 * (unsigned) v23), v166, v167);
+  TSTORE(v168, v157);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID6);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v169 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v170 = (uint64_t) v29;
+  TASSIGN(v169, v170);
+  pto::Shape<1, 1, 1, 1, 32> v171 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v172 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v173 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v12 * (unsigned) v22 + v19 * (unsigned) v23), v171, v172);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID6);
+  TLOAD(v169, v173);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v174 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v175 = (uint64_t) v28;
+  TASSIGN(v174, v175);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v174, v169);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v176 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v177 = (uint64_t) v27;
+  TASSIGN(v176, v177);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID6);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v176, v169);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID7);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v178 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v179 = (uint64_t) v28;
+  TASSIGN(v178, v179);
+  pipe_barrier(PIPE_V);
+  TMOV(v178, v174);
+  v178.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v180 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v181 = (uint64_t) v26;
+  TASSIGN(v180, v181);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v180, v178);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v182 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v183 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v184 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v12 * (unsigned) v24 + v19 * (unsigned) v23), v182, v183);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v184, v180);
+  pto::Shape<1, 1, 1, 1, 16> v185 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v186 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v187 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v12 * (unsigned) v24 + v19 * (unsigned) v23), v185, v186);
+  TSTORE(v187, v176);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID7);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v188 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v189 = (uint64_t) v29;
+  TASSIGN(v188, v189);
+  pto::Shape<1, 1, 1, 1, 32> v190 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v191 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v192 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v11 * (unsigned) v22 + v19 * (unsigned) v23), v190, v191);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID7);
+  TLOAD(v188, v192);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v193 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v194 = (uint64_t) v28;
+  TASSIGN(v193, v194);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v193, v188);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v195 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v196 = (uint64_t) v27;
+  TASSIGN(v195, v196);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID7);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v195, v188);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v197 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v198 = (uint64_t) v28;
+  TASSIGN(v197, v198);
+  pipe_barrier(PIPE_V);
+  TMOV(v197, v193);
+  v197.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v199 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v200 = (uint64_t) v26;
+  TASSIGN(v199, v200);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v199, v197);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v201 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v202 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v203 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v11 * (unsigned) v24 + v19 * (unsigned) v23), v201, v202);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v203, v199);
+  pto::Shape<1, 1, 1, 1, 16> v204 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v205 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v206 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v11 * (unsigned) v24 + v19 * (unsigned) v23), v204, v205);
+  TSTORE(v206, v195);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v207 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v208 = (uint64_t) v29;
+  TASSIGN(v207, v208);
+  pto::Shape<1, 1, 1, 1, 32> v209 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v210 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v211 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v10 * (unsigned) v22 + v19 * (unsigned) v23), v209, v210);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v207, v211);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v212 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v213 = (uint64_t) v28;
+  TASSIGN(v212, v213);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v212, v207);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v214 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v215 = (uint64_t) v27;
+  TASSIGN(v214, v215);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v214, v207);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v216 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v217 = (uint64_t) v28;
+  TASSIGN(v216, v217);
+  pipe_barrier(PIPE_V);
+  TMOV(v216, v212);
+  v216.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v218 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v219 = (uint64_t) v26;
+  TASSIGN(v218, v219);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v218, v216);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v220 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v221 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v222 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v10 * (unsigned) v24 + v19 * (unsigned) v23), v220, v221);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v222, v218);
+  pto::Shape<1, 1, 1, 1, 16> v223 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v224 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v225 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v10 * (unsigned) v24 + v19 * (unsigned) v23), v223, v224);
+  TSTORE(v225, v214);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v226 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v227 = (uint64_t) v29;
+  TASSIGN(v226, v227);
+  pto::Shape<1, 1, 1, 1, 32> v228 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v229 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v230 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v9 * (unsigned) v22 + v19 * (unsigned) v23), v228, v229);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v226, v230);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v231 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v232 = (uint64_t) v28;
+  TASSIGN(v231, v232);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v231, v226);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v233 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v234 = (uint64_t) v27;
+  TASSIGN(v233, v234);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v233, v226);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v235 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v236 = (uint64_t) v28;
+  TASSIGN(v235, v236);
+  pipe_barrier(PIPE_V);
+  TMOV(v235, v231);
+  v235.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v237 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v238 = (uint64_t) v26;
+  TASSIGN(v237, v238);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v237, v235);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v239 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v240 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v241 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v9 * (unsigned) v24 + v19 * (unsigned) v23), v239, v240);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v241, v237);
+  pto::Shape<1, 1, 1, 1, 16> v242 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v243 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v244 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v9 * (unsigned) v24 + v19 * (unsigned) v23), v242, v243);
+  TSTORE(v244, v233);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v245 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v246 = (uint64_t) v29;
+  TASSIGN(v245, v246);
+  pto::Shape<1, 1, 1, 1, 32> v247 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v248 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v249 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v8 * (unsigned) v22 + v19 * (unsigned) v23), v247, v248);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v245, v249);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v250 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v251 = (uint64_t) v28;
+  TASSIGN(v250, v251);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v250, v245);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v252 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v253 = (uint64_t) v27;
+  TASSIGN(v252, v253);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v252, v245);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v254 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v255 = (uint64_t) v28;
+  TASSIGN(v254, v255);
+  pipe_barrier(PIPE_V);
+  TMOV(v254, v250);
+  v254.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v256 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v257 = (uint64_t) v26;
+  TASSIGN(v256, v257);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v256, v254);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v258 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v259 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v260 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v8 * (unsigned) v24 + v19 * (unsigned) v23), v258, v259);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v260, v256);
+  pto::Shape<1, 1, 1, 1, 16> v261 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v262 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v263 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v8 * (unsigned) v24 + v19 * (unsigned) v23), v261, v262);
+  TSTORE(v263, v252);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v264 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v265 = (uint64_t) v29;
+  TASSIGN(v264, v265);
+  pto::Shape<1, 1, 1, 1, 32> v266 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v267 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v268 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v7 * (unsigned) v22 + v19 * (unsigned) v23), v266, v267);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v264, v268);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v269 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v270 = (uint64_t) v28;
+  TASSIGN(v269, v270);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v269, v264);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v271 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v272 = (uint64_t) v27;
+  TASSIGN(v271, v272);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v271, v264);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v273 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v274 = (uint64_t) v28;
+  TASSIGN(v273, v274);
+  pipe_barrier(PIPE_V);
+  TMOV(v273, v269);
+  v273.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v275 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v276 = (uint64_t) v26;
+  TASSIGN(v275, v276);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v275, v273);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v277 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v278 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v279 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v7 * (unsigned) v24 + v19 * (unsigned) v23), v277, v278);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v279, v275);
+  pto::Shape<1, 1, 1, 1, 16> v280 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v281 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v282 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v7 * (unsigned) v24 + v19 * (unsigned) v23), v280, v281);
+  TSTORE(v282, v271);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v283 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v284 = (uint64_t) v29;
+  TASSIGN(v283, v284);
+  pto::Shape<1, 1, 1, 1, 32> v285 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v286 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v287 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v6 * (unsigned) v22 + v19 * (unsigned) v23), v285, v286);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v283, v287);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v288 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v289 = (uint64_t) v28;
+  TASSIGN(v288, v289);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v288, v283);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v290 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v291 = (uint64_t) v27;
+  TASSIGN(v290, v291);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v290, v283);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v292 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v293 = (uint64_t) v28;
+  TASSIGN(v292, v293);
+  pipe_barrier(PIPE_V);
+  TMOV(v292, v288);
+  v292.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v294 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v295 = (uint64_t) v26;
+  TASSIGN(v294, v295);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v294, v292);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v296 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v297 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v298 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v6 * (unsigned) v24 + v19 * (unsigned) v23), v296, v297);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v298, v294);
+  pto::Shape<1, 1, 1, 1, 16> v299 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v300 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v301 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v6 * (unsigned) v24 + v19 * (unsigned) v23), v299, v300);
+  TSTORE(v301, v290);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v302 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v303 = (uint64_t) v29;
+  TASSIGN(v302, v303);
+  pto::Shape<1, 1, 1, 1, 32> v304 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v305 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v306 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v5 * (unsigned) v22 + v19 * (unsigned) v23), v304, v305);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v302, v306);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v307 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v308 = (uint64_t) v28;
+  TASSIGN(v307, v308);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v307, v302);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v309 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v310 = (uint64_t) v27;
+  TASSIGN(v309, v310);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v309, v302);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v311 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v312 = (uint64_t) v28;
+  TASSIGN(v311, v312);
+  pipe_barrier(PIPE_V);
+  TMOV(v311, v307);
+  v311.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v313 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v314 = (uint64_t) v26;
+  TASSIGN(v313, v314);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v313, v311);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v315 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v316 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v317 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v5 * (unsigned) v24 + v19 * (unsigned) v23), v315, v316);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v317, v313);
+  pto::Shape<1, 1, 1, 1, 16> v318 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v319 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v320 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v5 * (unsigned) v24 + v19 * (unsigned) v23), v318, v319);
+  TSTORE(v320, v309);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v321 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v24);
+  uint64_t v322 = (uint64_t) v29;
+  TASSIGN(v321, v322);
+  pto::Shape<1, 1, 1, 1, 32> v323 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<64, 64, 64, 64, 1> v324 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v325 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v19 + v4 * (unsigned) v22 + v19 * (unsigned) v23), v323, v324);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v321, v325);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v326 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v327 = (uint64_t) v28;
+  TASSIGN(v326, v327);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  pipe_barrier(PIPE_V);
+  TGATHER<Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P0101>(v326, v321);
+  Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v328 = Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v25);
+  uint64_t v329 = (uint64_t) v27;
+  TASSIGN(v328, v329);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TGATHER<Tile<TileType::Vec, int32_t, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, MaskPattern::P1010>(v328, v321);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v330 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v23, v20);
+  uint64_t v331 = (uint64_t) v28;
+  TASSIGN(v330, v331);
+  pipe_barrier(PIPE_V);
+  TMOV(v330, v326);
+  v330.SetValidShape(v23, v20);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null> v332 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Zero, CompactMode::Null>(v23, v25);
+  uint64_t v333 = (uint64_t) v26;
+  TASSIGN(v332, v333);
+  pipe_barrier(PIPE_V);
+  TFILLPAD(v332, v330);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 16> v334 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v335 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v336 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v19 + v4 * (unsigned) v24 + v19 * (unsigned) v23), v334, v335);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v336, v332);
+  pto::Shape<1, 1, 1, 1, 16> v337 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<32, 32, 32, 32, 1> v338 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v339 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v3 + (v19 + v4 * (unsigned) v24 + v19 * (unsigned) v23), v337, v338);
+  TSTORE(v339, v328);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: topk_vals_pad_inline110__ssa_v0
+    __gm__ Tensor* topk_vals_pad_inline110__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* topk_vals_pad_inline110__ssa_v0 = reinterpret_cast<__gm__ float*>(topk_vals_pad_inline110__ssa_v0_tensor->buffer.addr) + topk_vals_pad_inline110__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sorted_rows_inline97__ssa_v16
+    __gm__ Tensor* sorted_rows_inline97__ssa_v16_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* sorted_rows_inline97__ssa_v16 = reinterpret_cast<__gm__ float*>(sorted_rows_inline97__ssa_v16_tensor->buffer.addr) + sorted_rows_inline97__ssa_v16_tensor->start_offset;
+
+    // Unpack tensor: topk_idx_pad_inline104__ssa_v0
+    __gm__ Tensor* topk_idx_pad_inline104__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ int32_t* topk_idx_pad_inline104__ssa_v0 = reinterpret_cast<__gm__ int32_t*>(topk_idx_pad_inline104__ssa_v0_tensor->buffer.addr) + topk_idx_pad_inline104__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    route_extract_top2(topk_vals_pad_inline110__ssa_v0, sorted_rows_inline97__ssa_v16, topk_idx_pad_inline104__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_normalize_weights.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_normalize_weights.cpp
new file mode 100644
index 000000000..c925b3e7f
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_normalize_weights.cpp
@@ -0,0 +1,115 @@
+// Kernel Function: route_normalize_weights
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void route_normalize_weights(__gm__ float* v1, __gm__ float* v2) {
+  unsigned v3 = 0;
+  const float v4 = 1.0f;
+  const int32_t v5 = 128;
+  const int32_t v6 = 1;
+  const int32_t v7 = 32;
+  const int32_t v8 = 16;
+  const int64_t v9 = 10240;
+  const int64_t v10 = 2048;
+  const int64_t v11 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v12 = Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v8, v7);
+  uint64_t v13 = (uint64_t) v11;
+  TASSIGN(v12, v13);
+  pto::Shape<1, 1, 1, 16, 32> v14 = pto::Shape<1, 1, 1, 16, 32>();
+  pto::Stride<512, 512, 512, 32, 1> v15 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v16 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v1 + (v3 + v3 * (unsigned) v7 + v3 * (unsigned) v6), v14, v15);
+  TLOAD(v12, v16);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 128, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v17 = Tile<TileType::Vec, float, 16, 128, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v8, v5);
+  uint64_t v18 = (uint64_t) v10;
+  TASSIGN(v17, v18);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v19 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v8, v6);
+  uint64_t v20 = (uint64_t) v9;
+  TASSIGN(v19, v20);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TROWSUM(v19, v12, v17);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v21 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v8, v6);
+  uint64_t v22 = (uint64_t) v9;
+  TASSIGN(v21, v22);
+  Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v23 = Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v8, v7);
+  uint64_t v24 = (uint64_t) v11;
+  TASSIGN(v23, v24);
+  pipe_barrier(PIPE_V);
+  TROWEXPANDDIV(v23, v12, v21);
+  Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Vec, float, 16, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v8, v7);
+  uint64_t v26 = (uint64_t) v11;
+  TASSIGN(v25, v26);
+  pipe_barrier(PIPE_V);
+  TMULS(v25, v23, v4);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 32> v27 = pto::Shape<1, 1, 1, 16, 32>();
+  pto::Stride<512, 512, 512, 32, 1> v28 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v29 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 32>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v2 + (v3 + v3 * (unsigned) v7 + v3 * (unsigned) v6), v27, v28);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v29, v25);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: topk_vals_pad_inline110__ssa_v17
+    __gm__ Tensor* topk_vals_pad_inline110__ssa_v17_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* topk_vals_pad_inline110__ssa_v17 = reinterpret_cast<__gm__ float*>(topk_vals_pad_inline110__ssa_v17_tensor->buffer.addr) + topk_vals_pad_inline110__ssa_v17_tensor->start_offset;
+
+    // Unpack tensor: weight_out_pad_inline108__ssa_v0
+    __gm__ Tensor* weight_out_pad_inline108__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* weight_out_pad_inline108__ssa_v0 = reinterpret_cast<__gm__ float*>(weight_out_pad_inline108__ssa_v0_tensor->buffer.addr) + weight_out_pad_inline108__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    route_normalize_weights(topk_vals_pad_inline110__ssa_v17, weight_out_pad_inline108__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_sort_top2.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_sort_top2.cpp
new file mode 100644
index 000000000..d4ea158f0
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/route_sort_top2.cpp
@@ -0,0 +1,554 @@
+// Kernel Function: route_sort_top2
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void route_sort_top2(__gm__ float* v1, __gm__ float* v2) {
+  unsigned v3 = 15;
+  unsigned v4 = 14;
+  unsigned v5 = 13;
+  unsigned v6 = 12;
+  unsigned v7 = 11;
+  unsigned v8 = 10;
+  unsigned v9 = 9;
+  unsigned v10 = 8;
+  unsigned v11 = 7;
+  unsigned v12 = 6;
+  unsigned v13 = 5;
+  unsigned v14 = 4;
+  unsigned v15 = 3;
+  unsigned v16 = 2;
+  unsigned v17 = 1;
+  unsigned v18 = 0;
+  const int32_t v19 = 0;
+  const int32_t v20 = 64;
+  const int32_t v21 = 1;
+  const int32_t v22 = 32;
+  const int64_t v23 = 256;
+  const int64_t v24 = 128;
+  const int64_t v25 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  uint32_t v26 = (uint32_t) ((int32_t) v19);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v27 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v28 = (uint64_t) v25;
+  TASSIGN(v27, v28);
+  pto::Shape<1, 1, 1, 1, 32> v29 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v30 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v31 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v18 * (unsigned) v22 + v18 * (unsigned) v21), v29, v30);
+  TLOAD(v27, v31);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v32 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v33 = (uint64_t) v24;
+  TASSIGN(v32, v33);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v32, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v35 = (uint64_t) v23;
+  TASSIGN(v34, v35);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TSORT32(v34, v27, v32);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v36 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v37 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v38 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v18 * (unsigned) v20 + v18 * (unsigned) v21), v36, v37);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v38, v34);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v40 = (uint64_t) v25;
+  TASSIGN(v39, v40);
+  pto::Shape<1, 1, 1, 1, 32> v41 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v42 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v43 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v17 * (unsigned) v22 + v18 * (unsigned) v21), v41, v42);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v39, v43);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v44 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v45 = (uint64_t) v24;
+  TASSIGN(v44, v45);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v44, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v46 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v47 = (uint64_t) v23;
+  TASSIGN(v46, v47);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v46, v39, v44);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  pto::Shape<1, 1, 1, 1, 64> v48 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v49 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v50 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v17 * (unsigned) v20 + v18 * (unsigned) v21), v48, v49);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v50, v46);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v51 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v52 = (uint64_t) v25;
+  TASSIGN(v51, v52);
+  pto::Shape<1, 1, 1, 1, 32> v53 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v54 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v55 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v16 * (unsigned) v22 + v18 * (unsigned) v21), v53, v54);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  TLOAD(v51, v55);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v57 = (uint64_t) v24;
+  TASSIGN(v56, v57);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v56, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v59 = (uint64_t) v23;
+  TASSIGN(v58, v59);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID1);
+  TSORT32(v58, v51, v56);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID2);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  pto::Shape<1, 1, 1, 1, 64> v60 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v61 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v62 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v16 * (unsigned) v20 + v18 * (unsigned) v21), v60, v61);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID2);
+  TSTORE(v62, v58);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v63 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v64 = (uint64_t) v25;
+  TASSIGN(v63, v64);
+  pto::Shape<1, 1, 1, 1, 32> v65 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v66 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v67 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v15 * (unsigned) v22 + v18 * (unsigned) v21), v65, v66);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  TLOAD(v63, v67);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v68 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v69 = (uint64_t) v24;
+  TASSIGN(v68, v69);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v68, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v70 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v71 = (uint64_t) v23;
+  TASSIGN(v70, v71);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID2);
+  TSORT32(v70, v63, v68);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID3);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+  pto::Shape<1, 1, 1, 1, 64> v72 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v73 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v74 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v15 * (unsigned) v20 + v18 * (unsigned) v21), v72, v73);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID3);
+  TSTORE(v74, v70);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v75 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v76 = (uint64_t) v25;
+  TASSIGN(v75, v76);
+  pto::Shape<1, 1, 1, 1, 32> v77 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v78 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v79 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v14 * (unsigned) v22 + v18 * (unsigned) v21), v77, v78);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID3);
+  TLOAD(v75, v79);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v80 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v81 = (uint64_t) v24;
+  TASSIGN(v80, v81);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v80, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v82 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v83 = (uint64_t) v23;
+  TASSIGN(v82, v83);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID3);
+  TSORT32(v82, v75, v80);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID4);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID4);
+  pto::Shape<1, 1, 1, 1, 64> v84 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v85 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v86 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v14 * (unsigned) v20 + v18 * (unsigned) v21), v84, v85);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID4);
+  TSTORE(v86, v82);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID4);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v87 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v88 = (uint64_t) v25;
+  TASSIGN(v87, v88);
+  pto::Shape<1, 1, 1, 1, 32> v89 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v90 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v91 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v13 * (unsigned) v22 + v18 * (unsigned) v21), v89, v90);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID4);
+  TLOAD(v87, v91);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID5);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v92 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v93 = (uint64_t) v24;
+  TASSIGN(v92, v93);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v92, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v94 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v95 = (uint64_t) v23;
+  TASSIGN(v94, v95);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID5);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID4);
+  TSORT32(v94, v87, v92);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID5);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID5);
+  pto::Shape<1, 1, 1, 1, 64> v96 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v97 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v98 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v13 * (unsigned) v20 + v18 * (unsigned) v21), v96, v97);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID5);
+  TSTORE(v98, v94);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID5);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v99 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v100 = (uint64_t) v25;
+  TASSIGN(v99, v100);
+  pto::Shape<1, 1, 1, 1, 32> v101 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v102 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v103 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v12 * (unsigned) v22 + v18 * (unsigned) v21), v101, v102);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID5);
+  TLOAD(v99, v103);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID6);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v104 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v105 = (uint64_t) v24;
+  TASSIGN(v104, v105);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v104, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v106 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v107 = (uint64_t) v23;
+  TASSIGN(v106, v107);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID6);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID5);
+  TSORT32(v106, v99, v104);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID6);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID6);
+  pto::Shape<1, 1, 1, 1, 64> v108 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v109 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v110 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v12 * (unsigned) v20 + v18 * (unsigned) v21), v108, v109);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID6);
+  TSTORE(v110, v106);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID6);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v111 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v112 = (uint64_t) v25;
+  TASSIGN(v111, v112);
+  pto::Shape<1, 1, 1, 1, 32> v113 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v114 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v115 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v11 * (unsigned) v22 + v18 * (unsigned) v21), v113, v114);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID6);
+  TLOAD(v111, v115);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v116 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v117 = (uint64_t) v24;
+  TASSIGN(v116, v117);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v116, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v118 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v119 = (uint64_t) v23;
+  TASSIGN(v118, v119);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID6);
+  TSORT32(v118, v111, v116);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID7);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID7);
+  pto::Shape<1, 1, 1, 1, 64> v120 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v121 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v122 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v11 * (unsigned) v20 + v18 * (unsigned) v21), v120, v121);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID7);
+  TSTORE(v122, v118);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID7);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v123 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v124 = (uint64_t) v25;
+  TASSIGN(v123, v124);
+  pto::Shape<1, 1, 1, 1, 32> v125 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v126 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v127 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v10 * (unsigned) v22 + v18 * (unsigned) v21), v125, v126);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID7);
+  TLOAD(v123, v127);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v128 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v129 = (uint64_t) v24;
+  TASSIGN(v128, v129);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v128, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v130 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v131 = (uint64_t) v23;
+  TASSIGN(v130, v131);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID7);
+  TSORT32(v130, v123, v128);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v132 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v133 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v134 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v10 * (unsigned) v20 + v18 * (unsigned) v21), v132, v133);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v134, v130);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v135 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v136 = (uint64_t) v25;
+  TASSIGN(v135, v136);
+  pto::Shape<1, 1, 1, 1, 32> v137 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v138 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v139 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v9 * (unsigned) v22 + v18 * (unsigned) v21), v137, v138);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v135, v139);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v140 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v141 = (uint64_t) v24;
+  TASSIGN(v140, v141);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v140, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v142 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v143 = (uint64_t) v23;
+  TASSIGN(v142, v143);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v142, v135, v140);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v144 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v145 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v146 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v9 * (unsigned) v20 + v18 * (unsigned) v21), v144, v145);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v146, v142);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v147 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v148 = (uint64_t) v25;
+  TASSIGN(v147, v148);
+  pto::Shape<1, 1, 1, 1, 32> v149 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v150 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v151 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v8 * (unsigned) v22 + v18 * (unsigned) v21), v149, v150);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v147, v151);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v152 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v153 = (uint64_t) v24;
+  TASSIGN(v152, v153);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v152, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v154 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v155 = (uint64_t) v23;
+  TASSIGN(v154, v155);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v154, v147, v152);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v156 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v157 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v158 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v8 * (unsigned) v20 + v18 * (unsigned) v21), v156, v157);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v158, v154);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v159 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v160 = (uint64_t) v25;
+  TASSIGN(v159, v160);
+  pto::Shape<1, 1, 1, 1, 32> v161 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v162 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v163 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v7 * (unsigned) v22 + v18 * (unsigned) v21), v161, v162);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v159, v163);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v164 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v165 = (uint64_t) v24;
+  TASSIGN(v164, v165);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v164, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v166 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v167 = (uint64_t) v23;
+  TASSIGN(v166, v167);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v166, v159, v164);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v168 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v169 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v170 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v7 * (unsigned) v20 + v18 * (unsigned) v21), v168, v169);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v170, v166);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v171 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v172 = (uint64_t) v25;
+  TASSIGN(v171, v172);
+  pto::Shape<1, 1, 1, 1, 32> v173 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v174 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v175 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v6 * (unsigned) v22 + v18 * (unsigned) v21), v173, v174);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v171, v175);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v176 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v177 = (uint64_t) v24;
+  TASSIGN(v176, v177);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v176, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v178 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v179 = (uint64_t) v23;
+  TASSIGN(v178, v179);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v178, v171, v176);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v180 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v181 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v182 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v6 * (unsigned) v20 + v18 * (unsigned) v21), v180, v181);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v182, v178);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v183 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v184 = (uint64_t) v25;
+  TASSIGN(v183, v184);
+  pto::Shape<1, 1, 1, 1, 32> v185 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v186 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v187 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v5 * (unsigned) v22 + v18 * (unsigned) v21), v185, v186);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v183, v187);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v188 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v189 = (uint64_t) v24;
+  TASSIGN(v188, v189);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v188, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v190 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v191 = (uint64_t) v23;
+  TASSIGN(v190, v191);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v190, v183, v188);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v192 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v193 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v194 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v5 * (unsigned) v20 + v18 * (unsigned) v21), v192, v193);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v194, v190);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v195 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v196 = (uint64_t) v25;
+  TASSIGN(v195, v196);
+  pto::Shape<1, 1, 1, 1, 32> v197 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v198 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v199 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v4 * (unsigned) v22 + v18 * (unsigned) v21), v197, v198);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v195, v199);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v200 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v201 = (uint64_t) v24;
+  TASSIGN(v200, v201);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v200, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v202 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v203 = (uint64_t) v23;
+  TASSIGN(v202, v203);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v202, v195, v200);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v204 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v205 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v206 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v4 * (unsigned) v20 + v18 * (unsigned) v21), v204, v205);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v206, v202);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v207 = Tile<TileType::Vec, float, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v208 = (uint64_t) v25;
+  TASSIGN(v207, v208);
+  pto::Shape<1, 1, 1, 1, 32> v209 = pto::Shape<1, 1, 1, 1, 32>();
+  pto::Stride<32, 32, 32, 32, 1> v210 = pto::Stride<32, 32, 32, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND> v211 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 32>, pto::Stride<32, 32, 32, 32, 1>, pto::Layout::ND>(v1 + (v18 + v3 * (unsigned) v22 + v18 * (unsigned) v21), v209, v210);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  TLOAD(v207, v211);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v212 = Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v22);
+  uint64_t v213 = (uint64_t) v24;
+  TASSIGN(v212, v213);
+  pipe_barrier(PIPE_V);
+  TCI<Tile<TileType::Vec, uint32_t, 1, 32, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>, uint32_t, 0>(v212, v26);
+  Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v214 = Tile<TileType::Vec, float, 1, 64, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v21, v20);
+  uint64_t v215 = (uint64_t) v23;
+  TASSIGN(v214, v215);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TSORT32(v214, v207, v212);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 1, 64> v216 = pto::Shape<1, 1, 1, 1, 64>();
+  pto::Stride<64, 64, 64, 64, 1> v217 = pto::Stride<64, 64, 64, 64, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND> v218 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 64>, pto::Stride<64, 64, 64, 64, 1>, pto::Layout::ND>(v2 + (v18 + v3 * (unsigned) v20 + v18 * (unsigned) v21), v216, v217);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v218, v214);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: biased_scores_v1_inline136__ssa_v0
+    __gm__ Tensor* biased_scores_v1_inline136__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* biased_scores_v1_inline136__ssa_v0 = reinterpret_cast<__gm__ float*>(biased_scores_v1_inline136__ssa_v0_tensor->buffer.addr) + biased_scores_v1_inline136__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sorted_rows_inline97__ssa_v0
+    __gm__ Tensor* sorted_rows_inline97__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* sorted_rows_inline97__ssa_v0 = reinterpret_cast<__gm__ float*>(sorted_rows_inline97__ssa_v0_tensor->buffer.addr) + sorted_rows_inline97__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    route_sort_top2(biased_scores_v1_inline136__ssa_v0, sorted_rows_inline97__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_gate_up_dequant.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_gate_up_dequant.cpp
new file mode 100644
index 000000000..157809f0b
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_gate_up_dequant.cpp
@@ -0,0 +1,198 @@
+// Kernel Function: sh_gate_up_dequant
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_gate_up_dequant(__gm__ float* v1, __gm__ float* v2, __gm__ int32_t* v3, __gm__ int32_t* v4, __gm__ float* v5, __gm__ float* v6, __gm__ float* v7, int32_t v8) {
+  RoundMode v9 = RoundMode::CAST_NONE;
+  unsigned v10 = 0;
+  const int32_t v11 = 256;
+  const int32_t v12 = 16;
+  const int32_t v13 = 1;
+  const int64_t v14 = 18432;
+  const int64_t v15 = 2048;
+  const int64_t v16 = 1024;
+  const int64_t v17 = 0;
+  const int64_t v18 = 67584;
+  const int64_t v19 = 51200;
+  const int64_t v20 = 34816;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v21 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v22 = (uint64_t) v20;
+  TASSIGN(v21, v22);
+  pto::Shape<1, 1, 1, 16, 256> v23 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v24 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v25 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v3 + (v10 + v10 * (unsigned) v11 + v10 * (unsigned) v13), v23, v24);
+  TLOAD(v21, v25);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v27 = (uint64_t) v19;
+  TASSIGN(v26, v27);
+  pto::Shape<1, 1, 1, 16, 256> v28 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v29 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v30 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v4 + (v10 + v10 * (unsigned) v11 + v10 * (unsigned) v13), v28, v29);
+  TLOAD(v26, v30);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v31 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v13);
+  uint64_t v32 = (uint64_t) v18;
+  TASSIGN(v31, v32);
+  pto::Shape<1, 1, 1, 16, 1> v33 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v34 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v35 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v5 + (v10 + v10 * (unsigned) v13 + v10 * (unsigned) v12), v33, v34);
+  TLOAD(v31, v35);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v37 = (uint64_t) v17;
+  TASSIGN(v36, v37);
+  pto::Shape<1, 1, 1, 1, 256> v38 = pto::Shape<1, 1, 1, 1, 256>();
+  pto::Stride<256, 256, 256, 256, 1> v39 = pto::Stride<256, 256, 256, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<256, 256, 256, 256, 1>, pto::Layout::ND> v40 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<256, 256, 256, 256, 1>, pto::Layout::ND>(v1 + (v10 + (unsigned) v8 * (unsigned) v13), v38, v39);
+  TLOAD(v36, v40);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v42 = (uint64_t) v17;
+  TASSIGN(v41, v42);
+  Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v44 = (uint64_t) v16;
+  TASSIGN(v43, v44);
+  pto::Shape<1, 1, 1, 1, 256> v45 = pto::Shape<1, 1, 1, 1, 256>();
+  pto::Stride<256, 256, 256, 256, 1> v46 = pto::Stride<256, 256, 256, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<256, 256, 256, 256, 1>, pto::Layout::ND> v47 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 256>, pto::Stride<256, 256, 256, 256, 1>, pto::Layout::ND>(v2 + (v10 + (unsigned) v8 * (unsigned) v13), v45, v46);
+  TLOAD(v43, v47);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v48 = Tile<TileType::Vec, float, 1, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v49 = (uint64_t) v16;
+  TASSIGN(v48, v49);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v51 = (uint64_t) v15;
+  TASSIGN(v50, v51);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v50, v21, v9);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v53 = (uint64_t) v14;
+  TASSIGN(v52, v53);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TCVT(v52, v26, v9);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v55 = (uint64_t) v15;
+  TASSIGN(v54, v55);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  TROWEXPANDMUL(v54, v50, v31);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v57 = (uint64_t) v15;
+  TASSIGN(v56, v57);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID3);
+  TCOLEXPANDMUL(v56, v54, v41);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v58 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v59 = (uint64_t) v14;
+  TASSIGN(v58, v59);
+  TROWEXPANDMUL(v58, v52, v31);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v60 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v61 = (uint64_t) v14;
+  TASSIGN(v60, v61);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID4);
+  TCOLEXPANDMUL(v60, v58, v48);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  pto::Shape<1, 1, 1, 16, 256> v62 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v63 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v64 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v6 + (v10 + v10 * (unsigned) v11 + v10 * (unsigned) v13), v62, v63);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v64, v56);
+  pto::Shape<1, 1, 1, 16, 256> v65 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v66 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v67 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v7 + (v10 + v10 * (unsigned) v11 + v10 * (unsigned) v13), v65, v66);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v67, v60);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: shared_w1_scale__ssa_v0
+    __gm__ Tensor* shared_w1_scale__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* shared_w1_scale__ssa_v0 = reinterpret_cast<__gm__ float*>(shared_w1_scale__ssa_v0_tensor->buffer.addr) + shared_w1_scale__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: shared_w3_scale__ssa_v0
+    __gm__ Tensor* shared_w3_scale__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* shared_w3_scale__ssa_v0 = reinterpret_cast<__gm__ float*>(shared_w3_scale__ssa_v0_tensor->buffer.addr) + shared_w3_scale__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sh_gate_acc_inline95__rv_v2
+    __gm__ Tensor* sh_gate_acc_inline95__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ int32_t* sh_gate_acc_inline95__rv_v2 = reinterpret_cast<__gm__ int32_t*>(sh_gate_acc_inline95__rv_v2_tensor->buffer.addr) + sh_gate_acc_inline95__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: sh_up_acc_inline118__rv_v2
+    __gm__ Tensor* sh_up_acc_inline118__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ int32_t* sh_up_acc_inline118__rv_v2 = reinterpret_cast<__gm__ int32_t*>(sh_up_acc_inline118__rv_v2_tensor->buffer.addr) + sh_up_acc_inline118__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: x_local_scale_dq_inline32__ssa_v0
+    __gm__ Tensor* x_local_scale_dq_inline32__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ float* x_local_scale_dq_inline32__ssa_v0 = reinterpret_cast<__gm__ float*>(x_local_scale_dq_inline32__ssa_v0_tensor->buffer.addr) + x_local_scale_dq_inline32__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[5]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack tensor: ret1__out
+    __gm__ Tensor* ret1__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[6]);
+    __gm__ float* ret1__out = reinterpret_cast<__gm__ float*>(ret1__out_tensor->buffer.addr) + ret1__out_tensor->start_offset;
+
+    // Unpack scalar: n0_inline113__idx_v0
+    union { uint64_t u64; int64_t val; } n0_inline113__idx_v0_conv;
+    n0_inline113__idx_v0_conv.u64 = args[7];
+    int64_t n0_inline113__idx_v0 = n0_inline113__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    sh_gate_up_dequant(shared_w1_scale__ssa_v0, shared_w3_scale__ssa_v0, sh_gate_acc_inline95__rv_v2, sh_up_acc_inline118__rv_v2, x_local_scale_dq_inline32__ssa_v0, ret0__out, ret1__out, n0_inline113__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_h_q.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_h_q.cpp
new file mode 100644
index 000000000..358916169
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_h_q.cpp
@@ -0,0 +1,213 @@
+// Kernel Function: sh_h_q
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_h_q(__gm__ float* v1, __gm__ int8_t* v2, __gm__ float* v3) {
+  RoundMode v4 = RoundMode::CAST_TRUNC;
+  RoundMode v5 = RoundMode::CAST_ROUND;
+  unsigned v6 = 0;
+  const float v7 = 127.0f;
+  const int32_t v8 = 256;
+  const int32_t v9 = 0;
+  const float v10 = 9.99999974E-5f;
+  const int32_t v11 = 1;
+  const int32_t v12 = 4096;
+  const int32_t v13 = 16;
+  const int64_t v14 = 24640;
+  const int64_t v15 = 16448;
+  const int64_t v16 = 64;
+  const int64_t v17 = 0;
+  const int64_t v18 = 61568;
+  const int64_t v19 = 45184;
+  const int64_t v20 = 28800;
+  const int64_t v21 = 28736;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v22 = (size_t) v12;
+  size_t v23 = (size_t) v9;
+  size_t v24 = (size_t) v8;
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v26 = (uint64_t) v21;
+  TASSIGN(v25, v26);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TEXPANDS(v25, v10);
+  for (size_t v27 = v23; v27 < v22; v27 += v24) {
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v29 = (uint64_t) v20;
+    TASSIGN(v28, v29);
+    pto::Shape<1, 1, 1, 16, 256> v30 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v31 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v32 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v12 + (unsigned) ((int32_t) v27) * (unsigned) v11), v30, v31);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v28, v32);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v34 = (uint64_t) v19;
+    TASSIGN(v33, v34);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    TNEG(v33, v28);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v36 = (uint64_t) v20;
+    TASSIGN(v35, v36);
+    pipe_barrier(PIPE_V);
+    TMAX(v35, v28, v33);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v38 = (uint64_t) v19;
+    TASSIGN(v37, v38);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+    uint64_t v40 = (uint64_t) v18;
+    TASSIGN(v39, v40);
+    pipe_barrier(PIPE_V);
+    TROWMAX(v39, v35, v37);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v42 = (uint64_t) v18;
+    TASSIGN(v41, v42);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v44 = (uint64_t) v21;
+    TASSIGN(v43, v44);
+    pipe_barrier(PIPE_V);
+    TMAX(v43, v25, v41);
+  }
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v46 = (uint64_t) v17;
+  TASSIGN(v45, v46);
+  TEXPANDS(v45, v7);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v48 = (uint64_t) v21;
+  TASSIGN(v47, v48);
+  pipe_barrier(PIPE_V);
+  TDIV(v47, v45, v25);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v49 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v50 = (uint64_t) v17;
+  TASSIGN(v49, v50);
+  pipe_barrier(PIPE_V);
+  TRECIP(v49, v47);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v51 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v52 = (uint64_t) v17;
+  TASSIGN(v51, v52);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v53 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v54 = (uint64_t) v21;
+  TASSIGN(v53, v54);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  for (size_t v55 = v23; v55 < v22; v55 += v24) {
+    int32_t v56 = (int32_t) v55;
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v57 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v58 = (uint64_t) v20;
+    TASSIGN(v57, v58);
+    pto::Shape<1, 1, 1, 16, 256> v59 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v60 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v61 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v12 + (unsigned) v56 * (unsigned) v11), v59, v60);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    TLOAD(v57, v61);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v62 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v63 = (uint64_t) v20;
+    TASSIGN(v62, v63);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    TROWEXPANDMUL(v62, v57, v53);
+    Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v64 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v65 = (uint64_t) v16;
+    TASSIGN(v64, v65);
+    pipe_barrier(PIPE_V);
+    TCVT(v64, v62, v5);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v66 = Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v67 = (uint64_t) v15;
+    TASSIGN(v66, v67);
+    pipe_barrier(PIPE_V);
+    TCVT(v66, v64, v5);
+    Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v68 = Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v69 = (uint64_t) v14;
+    TASSIGN(v68, v69);
+    pipe_barrier(PIPE_V);
+    wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+    TCVT(v68, v66, v4);
+    set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    pto::Shape<1, 1, 1, 16, 256> v70 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v71 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v72 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v6 + v6 * (unsigned) v12 + (unsigned) v56 * (unsigned) v11), v70, v71);
+    wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    TSTORE(v72, v68);
+    set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  }
+  pto::Shape<1, 1, 1, 16, 1> v73 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v74 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v75 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v3 + (v6 + v6 * (unsigned) v11 + v6 * (unsigned) v13), v73, v74);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v75, v51);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: sh_tile_fp32_inline88__rv_v2
+    __gm__ Tensor* sh_tile_fp32_inline88__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* sh_tile_fp32_inline88__rv_v2 = reinterpret_cast<__gm__ float*>(sh_tile_fp32_inline88__rv_v2_tensor->buffer.addr) + sh_tile_fp32_inline88__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: sh_tile_i8_inline126__ssa_v0
+    __gm__ Tensor* sh_tile_i8_inline126__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* sh_tile_i8_inline126__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(sh_tile_i8_inline126__ssa_v0_tensor->buffer.addr) + sh_tile_i8_inline126__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    sh_h_q(sh_tile_fp32_inline88__rv_v2, sh_tile_i8_inline126__ssa_v0, ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_swiglu.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_swiglu.cpp
new file mode 100644
index 000000000..644df3f38
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_swiglu.cpp
@@ -0,0 +1,143 @@
+// Kernel Function: sh_swiglu
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_swiglu(__gm__ float* v1, __gm__ float* v2, __gm__ float* v3, int32_t v4) {
+  unsigned v5 = 0;
+  const float v6 = 1.0f;
+  const int32_t v7 = 4096;
+  const int32_t v8 = 1;
+  const int32_t v9 = 256;
+  const int32_t v10 = 16;
+  const int64_t v11 = 49152;
+  const int64_t v12 = 32768;
+  const int64_t v13 = 16384;
+  const int64_t v14 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v15 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v16 = (uint64_t) v14;
+  TASSIGN(v15, v16);
+  pto::Shape<1, 1, 1, 16, 256> v17 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v18 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v19 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v9 + v5 * (unsigned) v8), v17, v18);
+  TLOAD(v15, v19);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v20 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v21 = (uint64_t) v13;
+  TASSIGN(v20, v21);
+  pto::Shape<1, 1, 1, 16, 256> v22 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<4096, 4096, 4096, 256, 1> v23 = pto::Stride<4096, 4096, 4096, 256, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND> v24 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<4096, 4096, 4096, 256, 1>, pto::Layout::ND>(v2 + (v5 + v5 * (unsigned) v9 + v5 * (unsigned) v8), v22, v23);
+  TLOAD(v20, v24);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v26 = (uint64_t) v12;
+  TASSIGN(v25, v26);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TNEG(v25, v15);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v27 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v28 = (uint64_t) v12;
+  TASSIGN(v27, v28);
+  pipe_barrier(PIPE_V);
+  TEXP(v27, v25);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v29 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v30 = (uint64_t) v12;
+  TASSIGN(v29, v30);
+  pipe_barrier(PIPE_V);
+  TADDS(v29, v27, v6);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v31 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v32 = (uint64_t) v11;
+  TASSIGN(v31, v32);
+  pipe_barrier(PIPE_V);
+  TRECIP(v31, v29);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v34 = (uint64_t) v14;
+  TASSIGN(v33, v34);
+  pipe_barrier(PIPE_V);
+  TMUL(v33, v15, v31);
+  Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v9);
+  uint64_t v36 = (uint64_t) v14;
+  TASSIGN(v35, v36);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TMUL(v35, v33, v20);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 256> v37 = pto::Shape<1, 1, 1, 16, 256>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v38 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v39 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v3 + (v5 + v5 * (unsigned) v7 + (unsigned) v4 * (unsigned) v8), v37, v38);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v39, v35);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: sh_gate_v1_inline50__ssa_v0
+    __gm__ Tensor* sh_gate_v1_inline50__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* sh_gate_v1_inline50__ssa_v0 = reinterpret_cast<__gm__ float*>(sh_gate_v1_inline50__ssa_v0_tensor->buffer.addr) + sh_gate_v1_inline50__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sh_up_v1_inline9__ssa_v0
+    __gm__ Tensor* sh_up_v1_inline9__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* sh_up_v1_inline9__ssa_v0 = reinterpret_cast<__gm__ float*>(sh_up_v1_inline9__ssa_v0_tensor->buffer.addr) + sh_up_v1_inline9__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sh_tile_fp32_inline88__iter_v1
+    __gm__ Tensor* sh_tile_fp32_inline88__iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* sh_tile_fp32_inline88__iter_v1 = reinterpret_cast<__gm__ float*>(sh_tile_fp32_inline88__iter_v1_tensor->buffer.addr) + sh_tile_fp32_inline88__iter_v1_tensor->start_offset;
+
+    // Unpack scalar: n0_inline113__idx_v0
+    union { uint64_t u64; int64_t val; } n0_inline113__idx_v0_conv;
+    n0_inline113__idx_v0_conv.u64 = args[3];
+    int64_t n0_inline113__idx_v0 = n0_inline113__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    sh_swiglu(sh_gate_v1_inline50__ssa_v0, sh_up_v1_inline9__ssa_v0, sh_tile_fp32_inline88__iter_v1, n0_inline113__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_w2_dequant.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_w2_dequant.cpp
new file mode 100644
index 000000000..2cdb6bce2
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_w2_dequant.cpp
@@ -0,0 +1,143 @@
+// Kernel Function: sh_w2_dequant
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_w2_dequant(__gm__ float* v1, __gm__ int32_t* v2, __gm__ float* v3, __gm__ float* v4, int32_t v5) {
+  RoundMode v6 = RoundMode::CAST_NONE;
+  unsigned v7 = 0;
+  const int32_t v8 = 512;
+  const int32_t v9 = 16;
+  const int32_t v10 = 1;
+  const int64_t v11 = 34880;
+  const int64_t v12 = 32832;
+  const int64_t v13 = 32768;
+  const int64_t v14 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, int32_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v15 = Tile<TileType::Vec, int32_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v16 = (uint64_t) v14;
+  TASSIGN(v15, v16);
+  pto::Shape<1, 1, 1, 16, 512> v17 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v18 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v19 = GlobalTensor<int32_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v2 + (v7 + v7 * (unsigned) v8 + v7 * (unsigned) v10), v17, v18);
+  TLOAD(v15, v19);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v20 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v10);
+  uint64_t v21 = (uint64_t) v13;
+  TASSIGN(v20, v21);
+  pto::Shape<1, 1, 1, 16, 1> v22 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v23 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v24 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v3 + (v7 + v7 * (unsigned) v10 + v7 * (unsigned) v9), v22, v23);
+  TLOAD(v20, v24);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v25 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v8);
+  uint64_t v26 = (uint64_t) v12;
+  TASSIGN(v25, v26);
+  pto::Shape<1, 1, 1, 1, 512> v27 = pto::Shape<1, 1, 1, 1, 512>();
+  pto::Stride<512, 512, 512, 512, 1> v28 = pto::Stride<512, 512, 512, 512, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<512, 512, 512, 512, 1>, pto::Layout::ND> v29 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 512>, pto::Stride<512, 512, 512, 512, 1>, pto::Layout::ND>(v1 + (v7 + (unsigned) v5 * (unsigned) v10), v27, v28);
+  TLOAD(v25, v29);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Vec, float, 1, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v8);
+  uint64_t v31 = (uint64_t) v12;
+  TASSIGN(v30, v31);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v32 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v33 = (uint64_t) v11;
+  TASSIGN(v32, v33);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v32, v15, v6);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v35 = (uint64_t) v11;
+  TASSIGN(v34, v35);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TROWEXPANDMUL(v34, v32, v20);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v37 = (uint64_t) v11;
+  TASSIGN(v36, v37);
+  pipe_barrier(PIPE_V);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID2);
+  TCOLEXPANDMUL(v36, v34, v30);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v38 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v39 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v40 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v4 + (v7 + v7 * (unsigned) v8 + v7 * (unsigned) v10), v38, v39);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v40, v36);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: shared_w2_scale__ssa_v0
+    __gm__ Tensor* shared_w2_scale__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* shared_w2_scale__ssa_v0 = reinterpret_cast<__gm__ float*>(shared_w2_scale__ssa_v0_tensor->buffer.addr) + shared_w2_scale__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sh_y_acc_inline4__rv_v2
+    __gm__ Tensor* sh_y_acc_inline4__rv_v2_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int32_t* sh_y_acc_inline4__rv_v2 = reinterpret_cast<__gm__ int32_t*>(sh_y_acc_inline4__rv_v2_tensor->buffer.addr) + sh_y_acc_inline4__rv_v2_tensor->start_offset;
+
+    // Unpack tensor: sh_tile_scale_dq_inline99__ssa_v0
+    __gm__ Tensor* sh_tile_scale_dq_inline99__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* sh_tile_scale_dq_inline99__ssa_v0 = reinterpret_cast<__gm__ float*>(sh_tile_scale_dq_inline99__ssa_v0_tensor->buffer.addr) + sh_tile_scale_dq_inline99__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: d0_inline64__idx_v0
+    union { uint64_t u64; int64_t val; } d0_inline64__idx_v0_conv;
+    d0_inline64__idx_v0_conv.u64 = args[4];
+    int64_t d0_inline64__idx_v0 = d0_inline64__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    sh_w2_dequant(shared_w2_scale__ssa_v0, sh_y_acc_inline4__rv_v2, sh_tile_scale_dq_inline99__ssa_v0, ret0__out, d0_inline64__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_write.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_write.cpp
new file mode 100644
index 000000000..492672961
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/sh_write.cpp
@@ -0,0 +1,103 @@
+// Kernel Function: sh_write
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void sh_write(__gm__ float* v1, __gm__ bfloat16_t* v2, int32_t v3) {
+  RoundMode v4 = RoundMode::CAST_RINT;
+  unsigned v5 = 0;
+  const int32_t v6 = 4096;
+  const int32_t v7 = 1;
+  const int32_t v8 = 512;
+  const int32_t v9 = 16;
+  const int64_t v10 = 32768;
+  const int64_t v11 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v12 = Tile<TileType::Vec, float, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v13 = (uint64_t) v11;
+  TASSIGN(v12, v13);
+  pto::Shape<1, 1, 1, 16, 512> v14 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<8192, 8192, 8192, 512, 1> v15 = pto::Stride<8192, 8192, 8192, 512, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND> v16 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<8192, 8192, 8192, 512, 1>, pto::Layout::ND>(v1 + (v5 + v5 * (unsigned) v8 + v5 * (unsigned) v7), v14, v15);
+  TLOAD(v12, v16);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v17 = Tile<TileType::Vec, bfloat16_t, 16, 512, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v8);
+  uint64_t v18 = (uint64_t) v10;
+  TASSIGN(v17, v18);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TCVT(v17, v12, v4);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 512> v19 = pto::Shape<1, 1, 1, 16, 512>();
+  pto::Stride<65536, 65536, 65536, 4096, 1> v20 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+  GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v21 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 512>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v5 + v5 * (unsigned) v6 + (unsigned) v3 * (unsigned) v7), v19, v20);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v21, v17);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: sh_y_v1_inline114__ssa_v0
+    __gm__ Tensor* sh_y_v1_inline114__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* sh_y_v1_inline114__ssa_v0 = reinterpret_cast<__gm__ float*>(sh_y_v1_inline114__ssa_v0_tensor->buffer.addr) + sh_y_v1_inline114__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: sh__iter_v1
+    __gm__ Tensor* sh__iter_v1_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ bfloat16_t* sh__iter_v1 = reinterpret_cast<__gm__ bfloat16_t*>(sh__iter_v1_tensor->buffer.addr) + sh__iter_v1_tensor->start_offset;
+
+    // Unpack scalar: d0_inline64__idx_v0
+    union { uint64_t u64; int64_t val; } d0_inline64__idx_v0_conv;
+    d0_inline64__idx_v0_conv.u64 = args[2];
+    int64_t d0_inline64__idx_v0 = d0_inline64__idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    sh_write(sh_y_v1_inline114__ssa_v0, sh__iter_v1, d0_inline64__idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post.cpp
new file mode 100644
index 000000000..453ae14d7
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post.cpp
@@ -0,0 +1,83 @@
+// Kernel Function: split_pre_post
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void split_pre_post(__gm__ float* v1) {
+  unsigned v2 = 0;
+  const float v3 = 1.0f;
+  const int32_t v4 = 1;
+  const int32_t v5 = 8;
+  const int32_t v6 = 16;
+  const int64_t v7 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v8 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v6, v5);
+  uint64_t v9 = (uint64_t) v7;
+  TASSIGN(v8, v9);
+  TEXPANDS(v8, v3);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 8> v10 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<128, 128, 128, 8, 1> v11 = pto::Stride<128, 128, 128, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND> v12 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND>(v1 + (v2 + v2 * (unsigned) v5 + v2 * (unsigned) v4), v10, v11);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v12, v8);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    split_pre_post(ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_0.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_0.cpp
new file mode 100644
index 000000000..bf6a0f230
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_0.cpp
@@ -0,0 +1,163 @@
+// Kernel Function: split_pre_post_0
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void split_pre_post_0(__gm__ float* v1, __gm__ float* v2, __gm__ float* v3, __gm__ float* v4, float v5) {
+  unsigned v6 = 0;
+  const float v7 = 9.99999997E-7f;
+  const float v8 = 1.0f;
+  const int32_t v9 = 1;
+  const int32_t v10 = 32;
+  const int32_t v11 = 8;
+  const int32_t v12 = 16;
+  const int64_t v13 = 1024;
+  const int64_t v14 = 512;
+  const int64_t v15 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v16 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v17 = (uint64_t) v15;
+  TASSIGN(v16, v17);
+  pto::Shape<1, 1, 1, 16, 8> v18 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<512, 512, 512, 32, 1> v19 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v20 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v10 + v6 * (unsigned) v9), v18, v19);
+  TLOAD(v16, v20);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v21 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v22 = (uint64_t) v14;
+  TASSIGN(v21, v22);
+  pto::Shape<1, 1, 1, 16, 8> v23 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<128, 128, 128, 8, 1> v24 = pto::Stride<128, 128, 128, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND> v25 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND>(v2 + (v6 + v6 * (unsigned) v11 + v6 * (unsigned) v9), v23, v24);
+  TLOAD(v21, v25);
+  Tile<TileType::Vec, float, 1, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, float, 1, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v11);
+  uint64_t v27 = (uint64_t) v13;
+  TASSIGN(v26, v27);
+  pto::Shape<1, 1, 1, 1, 8> v28 = pto::Shape<1, 1, 1, 1, 8>();
+  pto::Stride<8, 8, 8, 8, 1> v29 = pto::Stride<8, 8, 8, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 8>, pto::Stride<8, 8, 8, 8, 1>, pto::Layout::ND> v30 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 8>, pto::Stride<8, 8, 8, 8, 1>, pto::Layout::ND>(v3 + (v6 + v6 * (unsigned) v11 + v6 * (unsigned) v9), v28, v29);
+  TLOAD(v26, v30);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v31 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v32 = (uint64_t) v15;
+  TASSIGN(v31, v32);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TMULS(v31, v16, v5);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v34 = (uint64_t) v14;
+  TASSIGN(v33, v34);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TCOLEXPANDMUL(v33, v21, v26);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v36 = (uint64_t) v15;
+  TASSIGN(v35, v36);
+  pipe_barrier(PIPE_V);
+  TADD(v35, v31, v33);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v38 = (uint64_t) v15;
+  TASSIGN(v37, v38);
+  pipe_barrier(PIPE_V);
+  TNEG(v37, v35);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v40 = (uint64_t) v15;
+  TASSIGN(v39, v40);
+  pipe_barrier(PIPE_V);
+  TEXP(v39, v37);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v42 = (uint64_t) v15;
+  TASSIGN(v41, v42);
+  pipe_barrier(PIPE_V);
+  TADDS(v41, v39, v8);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v44 = (uint64_t) v14;
+  TASSIGN(v43, v44);
+  pipe_barrier(PIPE_V);
+  TRECIP(v43, v41);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v12, v11);
+  uint64_t v46 = (uint64_t) v15;
+  TASSIGN(v45, v46);
+  pipe_barrier(PIPE_V);
+  TADDS(v45, v43, v7);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 8> v47 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<128, 128, 128, 8, 1> v48 = pto::Stride<128, 128, 128, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND> v49 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND>(v4 + (v6 + v6 * (unsigned) v11 + v6 * (unsigned) v9), v47, v48);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v49, v45);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: t__tmp_v26
+    __gm__ Tensor* t__tmp_v26_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* t__tmp_v26 = reinterpret_cast<__gm__ float*>(t__tmp_v26_tensor->buffer.addr) + t__tmp_v26_tensor->start_offset;
+
+    // Unpack tensor: ones_hc_inline79__ssa_v0
+    __gm__ Tensor* ones_hc_inline79__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* ones_hc_inline79__ssa_v0 = reinterpret_cast<__gm__ float*>(ones_hc_inline79__ssa_v0_tensor->buffer.addr) + ones_hc_inline79__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: pre_base_inline76__ssa_v0
+    __gm__ Tensor* pre_base_inline76__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* pre_base_inline76__ssa_v0 = reinterpret_cast<__gm__ float*>(pre_base_inline76__ssa_v0_tensor->buffer.addr) + pre_base_inline76__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: scale0_inline21__ssa_v0
+    union { uint64_t u64; float val; } scale0_inline21__ssa_v0_conv;
+    scale0_inline21__ssa_v0_conv.u64 = args[4];
+    float scale0_inline21__ssa_v0 = scale0_inline21__ssa_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    split_pre_post_0(t__tmp_v26, ones_hc_inline79__ssa_v0, pre_base_inline76__ssa_v0, ret0__out, scale0_inline21__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_1.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_1.cpp
new file mode 100644
index 000000000..d60eb2b83
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_1.cpp
@@ -0,0 +1,178 @@
+// Kernel Function: split_pre_post_1
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void split_pre_post_1(__gm__ float* v1, __gm__ float* v2, __gm__ float* v3, __gm__ float* v4, __gm__ float* v5, float v6) {
+  unsigned v7 = 0;
+  const float v8 = 2.0f;
+  const float v9 = 1.0f;
+  const int32_t v10 = 1;
+  const int32_t v11 = 32;
+  const int32_t v12 = 8;
+  const int32_t v13 = 16;
+  const int64_t v14 = 0;
+  const int64_t v15 = 2048;
+  const int64_t v16 = 1536;
+  const int64_t v17 = 1024;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v18 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v19 = (uint64_t) v17;
+  TASSIGN(v18, v19);
+  pto::Shape<1, 1, 1, 16, 8> v20 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<512, 512, 512, 32, 1> v21 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v22 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v1 + (v7 + v7 * (unsigned) v11 + v7 * (unsigned) v10), v20, v21);
+  TLOAD(v18, v22);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v23 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v24 = (uint64_t) v16;
+  TASSIGN(v23, v24);
+  pto::Shape<1, 1, 1, 16, 8> v25 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<128, 128, 128, 8, 1> v26 = pto::Stride<128, 128, 128, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND> v27 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND>(v2 + (v7 + v7 * (unsigned) v12 + v7 * (unsigned) v10), v25, v26);
+  TLOAD(v23, v27);
+  Tile<TileType::Vec, float, 1, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, float, 1, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v10, v12);
+  uint64_t v29 = (uint64_t) v15;
+  TASSIGN(v28, v29);
+  pto::Shape<1, 1, 1, 1, 8> v30 = pto::Shape<1, 1, 1, 1, 8>();
+  pto::Stride<8, 8, 8, 8, 1> v31 = pto::Stride<8, 8, 8, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 8>, pto::Stride<8, 8, 8, 8, 1>, pto::Layout::ND> v32 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 8>, pto::Stride<8, 8, 8, 8, 1>, pto::Layout::ND>(v3 + (v7 + v7 * (unsigned) v12 + v7 * (unsigned) v10), v30, v31);
+  TLOAD(v28, v32);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v33 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v34 = (uint64_t) v17;
+  TASSIGN(v33, v34);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TMULS(v33, v18, v6);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v35 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v36 = (uint64_t) v16;
+  TASSIGN(v35, v36);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TCOLEXPANDMUL(v35, v23, v28);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v37 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v38 = (uint64_t) v17;
+  TASSIGN(v37, v38);
+  pipe_barrier(PIPE_V);
+  TADD(v37, v33, v35);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v39 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v40 = (uint64_t) v17;
+  TASSIGN(v39, v40);
+  pipe_barrier(PIPE_V);
+  TNEG(v39, v37);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v41 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v42 = (uint64_t) v17;
+  TASSIGN(v41, v42);
+  pipe_barrier(PIPE_V);
+  TEXP(v41, v39);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v43 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v44 = (uint64_t) v17;
+  TASSIGN(v43, v44);
+  pipe_barrier(PIPE_V);
+  TADDS(v43, v41, v9);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v45 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v46 = (uint64_t) v16;
+  TASSIGN(v45, v46);
+  pipe_barrier(PIPE_V);
+  TRECIP(v45, v43);
+  Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v47 = Tile<TileType::Vec, float, 16, 8, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v12);
+  uint64_t v48 = (uint64_t) v17;
+  TASSIGN(v47, v48);
+  pipe_barrier(PIPE_V);
+  TMULS(v47, v45, v8);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v49 = Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v13);
+  uint64_t v50 = (uint64_t) v14;
+  TASSIGN(v49, v50);
+  TEXPANDS(v49, v9);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  pto::Shape<1, 1, 1, 16, 8> v51 = pto::Shape<1, 1, 1, 16, 8>();
+  pto::Stride<128, 128, 128, 8, 1> v52 = pto::Stride<128, 128, 128, 8, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND> v53 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 8>, pto::Stride<128, 128, 128, 8, 1>, pto::Layout::ND>(v4 + (v7 + v7 * (unsigned) v12 + v7 * (unsigned) v10), v51, v52);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v53, v47);
+  pto::Shape<1, 1, 1, 16, 16> v54 = pto::Shape<1, 1, 1, 16, 16>();
+  pto::Stride<256, 256, 256, 16, 1> v55 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v56 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v5 + (v7 + v7 * (unsigned) v13 + v7 * (unsigned) v10), v54, v55);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v56, v49);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: t__tmp_v34
+    __gm__ Tensor* t__tmp_v34_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* t__tmp_v34 = reinterpret_cast<__gm__ float*>(t__tmp_v34_tensor->buffer.addr) + t__tmp_v34_tensor->start_offset;
+
+    // Unpack tensor: ones_hc_inline79__ssa_v0
+    __gm__ Tensor* ones_hc_inline79__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* ones_hc_inline79__ssa_v0 = reinterpret_cast<__gm__ float*>(ones_hc_inline79__ssa_v0_tensor->buffer.addr) + ones_hc_inline79__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: post_base_inline50__ssa_v0
+    __gm__ Tensor* post_base_inline50__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* post_base_inline50__ssa_v0 = reinterpret_cast<__gm__ float*>(post_base_inline50__ssa_v0_tensor->buffer.addr) + post_base_inline50__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack tensor: ret1__out
+    __gm__ Tensor* ret1__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[4]);
+    __gm__ float* ret1__out = reinterpret_cast<__gm__ float*>(ret1__out_tensor->buffer.addr) + ret1__out_tensor->start_offset;
+
+    // Unpack scalar: scale1_inline17__ssa_v0
+    union { uint64_t u64; float val; } scale1_inline17__ssa_v0_conv;
+    scale1_inline17__ssa_v0_conv.u64 = args[5];
+    float scale1_inline17__ssa_v0 = scale1_inline17__ssa_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    split_pre_post_1(t__tmp_v34, ones_hc_inline79__ssa_v0, post_base_inline50__ssa_v0, ret0__out, ret1__out, scale1_inline17__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_2.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_2.cpp
new file mode 100644
index 000000000..6f0848188
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/split_pre_post_2.cpp
@@ -0,0 +1,135 @@
+// Kernel Function: split_pre_post_2
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void split_pre_post_2(__gm__ float* v1, __gm__ float* v2, __gm__ float* v3, __gm__ float* v4, float v5) {
+  unsigned v6 = 0;
+  const int32_t v7 = 1;
+  const int32_t v8 = 32;
+  const int32_t v9 = 16;
+  const int64_t v10 = 2048;
+  const int64_t v11 = 1024;
+  const int64_t v12 = 0;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v13 = Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+  uint64_t v14 = (uint64_t) v12;
+  TASSIGN(v13, v14);
+  pto::Shape<1, 1, 1, 16, 16> v15 = pto::Shape<1, 1, 1, 16, 16>();
+  pto::Stride<512, 512, 512, 32, 1> v16 = pto::Stride<512, 512, 512, 32, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND> v17 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<512, 512, 512, 32, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v8 + v6 * (unsigned) v7), v15, v16);
+  TLOAD(v13, v17);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v18 = Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+  uint64_t v19 = (uint64_t) v11;
+  TASSIGN(v18, v19);
+  pto::Shape<1, 1, 1, 16, 16> v20 = pto::Shape<1, 1, 1, 16, 16>();
+  pto::Stride<256, 256, 256, 16, 1> v21 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v22 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v2 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v7), v20, v21);
+  TLOAD(v18, v22);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v23 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v7, v9);
+  uint64_t v24 = (uint64_t) v10;
+  TASSIGN(v23, v24);
+  pto::Shape<1, 1, 1, 1, 16> v25 = pto::Shape<1, 1, 1, 1, 16>();
+  pto::Stride<16, 16, 16, 16, 1> v26 = pto::Stride<16, 16, 16, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND> v27 = GlobalTensor<float, pto::Shape<1, 1, 1, 1, 16>, pto::Stride<16, 16, 16, 16, 1>, pto::Layout::ND>(v3 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v7), v25, v26);
+  TLOAD(v23, v27);
+  set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v28 = Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+  uint64_t v29 = (uint64_t) v12;
+  TASSIGN(v28, v29);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+  TMULS(v28, v13, v5);
+  Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v30 = Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+  uint64_t v31 = (uint64_t) v11;
+  TASSIGN(v30, v31);
+  wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+  TCOLEXPANDMUL(v30, v18, v23);
+  Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v32 = Tile<TileType::Vec, float, 16, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v9, v9);
+  uint64_t v33 = (uint64_t) v12;
+  TASSIGN(v32, v33);
+  pipe_barrier(PIPE_V);
+  TADD(v32, v28, v30);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  pto::Shape<1, 1, 1, 16, 16> v34 = pto::Shape<1, 1, 1, 16, 16>();
+  pto::Stride<256, 256, 256, 16, 1> v35 = pto::Stride<256, 256, 256, 16, 1>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND> v36 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 16>, pto::Stride<256, 256, 256, 16, 1>, pto::Layout::ND>(v4 + (v6 + v6 * (unsigned) v9 + v6 * (unsigned) v7), v34, v35);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+  TSTORE(v36, v32);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: comb_mix_inline41__ssa_v0
+    __gm__ Tensor* comb_mix_inline41__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* comb_mix_inline41__ssa_v0 = reinterpret_cast<__gm__ float*>(comb_mix_inline41__ssa_v0_tensor->buffer.addr) + comb_mix_inline41__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ones_comb_inline67__ssa_v0
+    __gm__ Tensor* ones_comb_inline67__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* ones_comb_inline67__ssa_v0 = reinterpret_cast<__gm__ float*>(ones_comb_inline67__ssa_v0_tensor->buffer.addr) + ones_comb_inline67__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: comb_base_inline70__ssa_v0
+    __gm__ Tensor* comb_base_inline70__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* comb_base_inline70__ssa_v0 = reinterpret_cast<__gm__ float*>(comb_base_inline70__ssa_v0_tensor->buffer.addr) + comb_base_inline70__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Unpack scalar: scale2_inline26__ssa_v0
+    union { uint64_t u64; float val; } scale2_inline26__ssa_v0_conv;
+    scale2_inline26__ssa_v0_conv.u64 = args[4];
+    float scale2_inline26__ssa_v0 = scale2_inline26__ssa_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    split_pre_post_2(comb_mix_inline41__ssa_v0, ones_comb_inline67__ssa_v0, comb_base_inline70__ssa_v0, ret0__out, scale2_inline26__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/write_post.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/write_post.cpp
new file mode 100644
index 000000000..6fb3e9515
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/write_post.cpp
@@ -0,0 +1,101 @@
+// Kernel Function: write_post
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void write_post(__gm__ float* v1, __gm__ float* v2, int32_t v3) {
+  const int32_t v4 = 3;
+  const int32_t v5 = 2;
+  const int32_t v6 = 4;
+  const int32_t v7 = 8;
+  const int32_t v8 = 16;
+  const int32_t v9 = 0;
+  const int32_t v10 = 1;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  for (size_t v11 = (size_t) v9; v11 < ((size_t) v8); v11 += (size_t) v10) {
+    int32_t v12 = (int32_t) ((uint32_t) ((int32_t) (uint32_t) v3 * (uint32_t) v8) + (uint32_t) ((int32_t) v11));
+    if (v12 < v8) {
+      int32_t v13 = (int32_t) ((uint32_t) v12 * (uint32_t) v7);
+      float v14 = v1[v13];
+      int32_t v15 = (int32_t) ((uint32_t) v12 * (uint32_t) v6);
+      v2[v15] = v14;
+      float v16 = v1[(int32_t) ((uint32_t) v13 + (uint32_t) v10)];
+      int32_t v17 = (int32_t) ((uint32_t) v15 + (uint32_t) v10);
+      v2[v17] = v16;
+      float v18 = v1[(int32_t) ((uint32_t) v13 + (uint32_t) v5)];
+      int32_t v19 = (int32_t) ((uint32_t) v15 + (uint32_t) v5);
+      v2[v19] = v18;
+      float v20 = v1[(int32_t) ((uint32_t) v13 + (uint32_t) v4)];
+      int32_t v21 = (int32_t) ((uint32_t) v15 + (uint32_t) v4);
+      v2[v21] = v20;
+    };
+  }
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: post_pad_flat_inline73__ssa_v0
+    __gm__ Tensor* post_pad_flat_inline73__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ float* post_pad_flat_inline73__ssa_v0 = reinterpret_cast<__gm__ float*>(post_pad_flat_inline73__ssa_v0_tensor->buffer.addr) + post_pad_flat_inline73__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: post_flat_inline51__ssa_v0
+    __gm__ Tensor* post_flat_inline51__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ float* post_flat_inline51__ssa_v0 = reinterpret_cast<__gm__ float*>(post_flat_inline51__ssa_v0_tensor->buffer.addr) + post_flat_inline51__ssa_v0_tensor->start_offset;
+
+    // Unpack scalar: t_inline7__co_idx_v0
+    union { uint64_t u64; int64_t val; } t_inline7__co_idx_v0_conv;
+    t_inline7__co_idx_v0_conv.u64 = args[2];
+    int64_t t_inline7__co_idx_v0 = t_inline7__co_idx_v0_conv.val;
+
+    // Forward to ptoas-generated function
+    write_post(post_pad_flat_inline73__ssa_v0, post_flat_inline51__ssa_v0, t_inline7__co_idx_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/write_route_outputs.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/write_route_outputs.cpp
new file mode 100644
index 000000000..af070e9b6
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/write_route_outputs.cpp
@@ -0,0 +1,281 @@
+// Kernel Function: write_route_outputs
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+#include "aicore/aicore.h"  // dcci — flush scalar writes out of L1 so the
+                            // downstream dispatch kernel sees them via scalar
+                            // reads. Without this, indices/weights stay in
+                            // this core's data cache and dispatch races
+                            // ~60% of runs in onboard mode.
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void write_route_outputs(__gm__ int32_t* v1, __gm__ int32_t* v2, __gm__ float* v3, __gm__ float* v4) {
+  const int32_t v5 = 31;
+  const int32_t v6 = 481;
+  const int32_t v7 = 30;
+  const int32_t v8 = 480;
+  const int32_t v9 = 29;
+  const int32_t v10 = 449;
+  const int32_t v11 = 28;
+  const int32_t v12 = 448;
+  const int32_t v13 = 27;
+  const int32_t v14 = 417;
+  const int32_t v15 = 26;
+  const int32_t v16 = 416;
+  const int32_t v17 = 25;
+  const int32_t v18 = 385;
+  const int32_t v19 = 24;
+  const int32_t v20 = 384;
+  const int32_t v21 = 23;
+  const int32_t v22 = 353;
+  const int32_t v23 = 22;
+  const int32_t v24 = 352;
+  const int32_t v25 = 21;
+  const int32_t v26 = 321;
+  const int32_t v27 = 20;
+  const int32_t v28 = 320;
+  const int32_t v29 = 19;
+  const int32_t v30 = 289;
+  const int32_t v31 = 18;
+  const int32_t v32 = 288;
+  const int32_t v33 = 17;
+  const int32_t v34 = 257;
+  const int32_t v35 = 16;
+  const int32_t v36 = 256;
+  const int32_t v37 = 15;
+  const int32_t v38 = 225;
+  const int32_t v39 = 14;
+  const int32_t v40 = 224;
+  const int32_t v41 = 13;
+  const int32_t v42 = 193;
+  const int32_t v43 = 12;
+  const int32_t v44 = 192;
+  const int32_t v45 = 11;
+  const int32_t v46 = 161;
+  const int32_t v47 = 10;
+  const int32_t v48 = 160;
+  const int32_t v49 = 9;
+  const int32_t v50 = 129;
+  const int32_t v51 = 8;
+  const int32_t v52 = 128;
+  const int32_t v53 = 7;
+  const int32_t v54 = 97;
+  const int32_t v55 = 6;
+  const int32_t v56 = 96;
+  const int32_t v57 = 5;
+  const int32_t v58 = 65;
+  const int32_t v59 = 4;
+  const int32_t v60 = 64;
+  const int32_t v61 = 3;
+  const int32_t v62 = 33;
+  const int32_t v63 = 2;
+  const int32_t v64 = 0;
+  const int32_t v65 = 32;
+  const int32_t v66 = 1;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  int32_t v67 = v1[v64];
+  v2[v64] = v67;
+  float v68 = v3[v64];
+  v4[v64] = v68;
+  int32_t v69 = v1[v66];
+  v2[v66] = v69;
+  float v70 = v3[v66];
+  v4[v66] = v70;
+  int32_t v71 = v1[v65];
+  v2[v63] = v71;
+  float v72 = v3[v65];
+  v4[v63] = v72;
+  int32_t v73 = v1[v62];
+  v2[v61] = v73;
+  float v74 = v3[v62];
+  v4[v61] = v74;
+  int32_t v75 = v1[v60];
+  v2[v59] = v75;
+  float v76 = v3[v60];
+  v4[v59] = v76;
+  int32_t v77 = v1[v58];
+  v2[v57] = v77;
+  float v78 = v3[v58];
+  v4[v57] = v78;
+  int32_t v79 = v1[v56];
+  v2[v55] = v79;
+  float v80 = v3[v56];
+  v4[v55] = v80;
+  int32_t v81 = v1[v54];
+  v2[v53] = v81;
+  float v82 = v3[v54];
+  v4[v53] = v82;
+  int32_t v83 = v1[v52];
+  v2[v51] = v83;
+  float v84 = v3[v52];
+  v4[v51] = v84;
+  int32_t v85 = v1[v50];
+  v2[v49] = v85;
+  float v86 = v3[v50];
+  v4[v49] = v86;
+  int32_t v87 = v1[v48];
+  v2[v47] = v87;
+  float v88 = v3[v48];
+  v4[v47] = v88;
+  int32_t v89 = v1[v46];
+  v2[v45] = v89;
+  float v90 = v3[v46];
+  v4[v45] = v90;
+  int32_t v91 = v1[v44];
+  v2[v43] = v91;
+  float v92 = v3[v44];
+  v4[v43] = v92;
+  int32_t v93 = v1[v42];
+  v2[v41] = v93;
+  float v94 = v3[v42];
+  v4[v41] = v94;
+  int32_t v95 = v1[v40];
+  v2[v39] = v95;
+  float v96 = v3[v40];
+  v4[v39] = v96;
+  int32_t v97 = v1[v38];
+  v2[v37] = v97;
+  float v98 = v3[v38];
+  v4[v37] = v98;
+  int32_t v99 = v1[v36];
+  v2[v35] = v99;
+  float v100 = v3[v36];
+  v4[v35] = v100;
+  int32_t v101 = v1[v34];
+  v2[v33] = v101;
+  float v102 = v3[v34];
+  v4[v33] = v102;
+  int32_t v103 = v1[v32];
+  v2[v31] = v103;
+  float v104 = v3[v32];
+  v4[v31] = v104;
+  int32_t v105 = v1[v30];
+  v2[v29] = v105;
+  float v106 = v3[v30];
+  v4[v29] = v106;
+  int32_t v107 = v1[v28];
+  v2[v27] = v107;
+  float v108 = v3[v28];
+  v4[v27] = v108;
+  int32_t v109 = v1[v26];
+  v2[v25] = v109;
+  float v110 = v3[v26];
+  v4[v25] = v110;
+  int32_t v111 = v1[v24];
+  v2[v23] = v111;
+  float v112 = v3[v24];
+  v4[v23] = v112;
+  int32_t v113 = v1[v22];
+  v2[v21] = v113;
+  float v114 = v3[v22];
+  v4[v21] = v114;
+  int32_t v115 = v1[v20];
+  v2[v19] = v115;
+  float v116 = v3[v20];
+  v4[v19] = v116;
+  int32_t v117 = v1[v18];
+  v2[v17] = v117;
+  float v118 = v3[v18];
+  v4[v17] = v118;
+  int32_t v119 = v1[v16];
+  v2[v15] = v119;
+  float v120 = v3[v16];
+  v4[v15] = v120;
+  int32_t v121 = v1[v14];
+  v2[v13] = v121;
+  float v122 = v3[v14];
+  v4[v13] = v122;
+  int32_t v123 = v1[v12];
+  v2[v11] = v123;
+  float v124 = v3[v12];
+  v4[v11] = v124;
+  int32_t v125 = v1[v10];
+  v2[v9] = v125;
+  float v126 = v3[v10];
+  v4[v9] = v126;
+  int32_t v127 = v1[v8];
+  v2[v7] = v127;
+  float v128 = v3[v8];
+  v4[v7] = v128;
+  int32_t v129 = v1[v6];
+  v2[v5] = v129;
+  float v130 = v3[v6];
+  v4[v5] = v130;
+  // Flush scalar writes (indices, weights) out of this AICore's L1 so the
+  // downstream dispatch task sees them when it scalar-reads `indices`. The
+  // PyPTO-emitted body has no tail sync, and dispatch's reader-side dcci on
+  // its own only invalidates the reader cache — both sides must participate.
+  pipe_barrier(PIPE_ALL);
+  dcci(v2, ENTIRE_DATA_CACHE, CACHELINE_OUT);
+  dcci(v4, ENTIRE_DATA_CACHE, CACHELINE_OUT);
+  dsb((mem_dsb_t)0);
+  #endif // __DAV_VEC__
+
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: topk_idx_flat_inline87__ssa_v0
+    __gm__ Tensor* topk_idx_flat_inline87__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ int32_t* topk_idx_flat_inline87__ssa_v0 = reinterpret_cast<__gm__ int32_t*>(topk_idx_flat_inline87__ssa_v0_tensor->buffer.addr) + topk_idx_flat_inline87__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: indices_flat_inline89__ssa_v0
+    __gm__ Tensor* indices_flat_inline89__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int32_t* indices_flat_inline89__ssa_v0 = reinterpret_cast<__gm__ int32_t*>(indices_flat_inline89__ssa_v0_tensor->buffer.addr) + indices_flat_inline89__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: weight_out_flat_inline86__ssa_v0
+    __gm__ Tensor* weight_out_flat_inline86__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* weight_out_flat_inline86__ssa_v0 = reinterpret_cast<__gm__ float*>(weight_out_flat_inline86__ssa_v0_tensor->buffer.addr) + weight_out_flat_inline86__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: weights_flat_inline88__ssa_v0
+    __gm__ Tensor* weights_flat_inline88__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[3]);
+    __gm__ float* weights_flat_inline88__ssa_v0 = reinterpret_cast<__gm__ float*>(weights_flat_inline88__ssa_v0_tensor->buffer.addr) + weights_flat_inline88__ssa_v0_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    write_route_outputs(topk_idx_flat_inline87__ssa_v0, indices_flat_inline89__ssa_v0, weight_out_flat_inline86__ssa_v0, weights_flat_inline88__ssa_v0);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/aiv/x_local_q.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/x_local_q.cpp
new file mode 100644
index 000000000..9e4b38721
--- /dev/null
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/aiv/x_local_q.cpp
@@ -0,0 +1,224 @@
+// Kernel Function: x_local_q
+// Generated by PyPTO IR Compiler (PTO backend)
+
+#include <cstdint>
+
+#ifndef __gm__
+#define __gm__
+#endif
+
+#ifndef __aicore__
+#if defined(__CPU_SIM)
+#define __aicore__
+#else
+#define __aicore__ [aicore]
+#endif
+#endif
+
+#include <pto/pto-inst.hpp>
+#include "tensor.h"
+
+
+using namespace pto;
+
+
+// --- ptoas-generated code ---
+
+enum class PTOAutoSyncTailMode : int {
+  kBarrierAll = 0,
+  kSetWaitMte3ToSEvent0 = 1,
+};
+
+static __aicore__ inline void ptoas_auto_sync_tail(
+    PTOAutoSyncTailMode mode = PTOAutoSyncTailMode::kBarrierAll) {
+  switch (mode) {
+  case PTOAutoSyncTailMode::kSetWaitMte3ToSEvent0:
+    set_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    wait_flag(PIPE_MTE3, PIPE_S, EVENT_ID0);
+    break;
+  case PTOAutoSyncTailMode::kBarrierAll:
+  default:
+    pipe_barrier(PIPE_ALL);
+    break;
+  }
+}
+
+static __aicore__ void x_local_q(__gm__ bfloat16_t* v1, __gm__ int8_t* v2, __gm__ float* v3) {
+  RoundMode v4 = RoundMode::CAST_TRUNC;
+  RoundMode v5 = RoundMode::CAST_ROUND;
+  unsigned v6 = 0;
+  const float v7 = 127.0f;
+  const int32_t v8 = 256;
+  const int32_t v9 = 0;
+  const float v10 = 9.99999974E-5f;
+  const int32_t v11 = 1;
+  const int32_t v12 = 4096;
+  const int32_t v13 = 16;
+  const int64_t v14 = 24640;
+  const int64_t v15 = 16448;
+  const int64_t v16 = 64;
+  const int64_t v17 = 0;
+  const int64_t v18 = 69760;
+  const int64_t v19 = 53376;
+  const int64_t v20 = 36992;
+  const int64_t v21 = 28800;
+  const int64_t v22 = 28736;
+  using T = float;
+
+  #if defined(__DAV_VEC__)
+  set_mask_norm();
+  set_vector_mask(-1, -1);
+  size_t v23 = (size_t) v12;
+  size_t v24 = (size_t) v9;
+  size_t v25 = (size_t) v8;
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v26 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v27 = (uint64_t) v22;
+  TASSIGN(v26, v27);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  TEXPANDS(v26, v10);
+  for (size_t v28 = v24; v28 < v23; v28 += v25) {
+    Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v29 = Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v30 = (uint64_t) v21;
+    TASSIGN(v29, v30);
+    pto::Shape<1, 1, 1, 16, 256> v31 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v32 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v33 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v12 + (unsigned) ((int32_t) v28) * (unsigned) v11), v31, v32);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    TLOAD(v29, v33);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v34 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v35 = (uint64_t) v20;
+    TASSIGN(v34, v35);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID0);
+    TCVT(v34, v29, v5);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v36 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v37 = (uint64_t) v19;
+    TASSIGN(v36, v37);
+    pipe_barrier(PIPE_V);
+    TNEG(v36, v34);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v38 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v39 = (uint64_t) v20;
+    TASSIGN(v38, v39);
+    pipe_barrier(PIPE_V);
+    TMAX(v38, v34, v36);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v40 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v41 = (uint64_t) v19;
+    TASSIGN(v40, v41);
+    Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v42 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+    uint64_t v43 = (uint64_t) v18;
+    TASSIGN(v42, v43);
+    pipe_barrier(PIPE_V);
+    TROWMAX(v42, v38, v40);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v44 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v45 = (uint64_t) v18;
+    TASSIGN(v44, v45);
+    Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v46 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+    uint64_t v47 = (uint64_t) v22;
+    TASSIGN(v46, v47);
+    pipe_barrier(PIPE_V);
+    TMAX(v46, v26, v44);
+  }
+  set_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v48 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v49 = (uint64_t) v17;
+  TASSIGN(v48, v49);
+  TEXPANDS(v48, v7);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v50 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v51 = (uint64_t) v22;
+  TASSIGN(v50, v51);
+  pipe_barrier(PIPE_V);
+  TDIV(v50, v48, v26);
+  Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v52 = Tile<TileType::Vec, float, 1, 16, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v11, v13);
+  uint64_t v53 = (uint64_t) v17;
+  TASSIGN(v52, v53);
+  pipe_barrier(PIPE_V);
+  TRECIP(v52, v50);
+  set_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v54 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v55 = (uint64_t) v17;
+  TASSIGN(v54, v55);
+  Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v56 = Tile<TileType::Vec, float, 16, 1, BLayout::ColMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v11);
+  uint64_t v57 = (uint64_t) v22;
+  TASSIGN(v56, v57);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID1);
+  for (size_t v58 = v24; v58 < v23; v58 += v25) {
+    int32_t v59 = (int32_t) v58;
+    Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v60 = Tile<TileType::Vec, bfloat16_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v61 = (uint64_t) v21;
+    TASSIGN(v60, v61);
+    pto::Shape<1, 1, 1, 16, 256> v62 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v63 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v64 = GlobalTensor<bfloat16_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v1 + (v6 + v6 * (unsigned) v12 + (unsigned) v59 * (unsigned) v11), v62, v63);
+    wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    TLOAD(v60, v64);
+    set_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v65 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v66 = (uint64_t) v20;
+    TASSIGN(v65, v66);
+    wait_flag(PIPE_MTE2, PIPE_V, EVENT_ID1);
+    TCVT(v65, v60, v5);
+    set_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+    Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v67 = Tile<TileType::Vec, float, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v68 = (uint64_t) v20;
+    TASSIGN(v67, v68);
+    pipe_barrier(PIPE_V);
+    TROWEXPANDMUL(v67, v65, v56);
+    Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v69 = Tile<TileType::Vec, int32_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v70 = (uint64_t) v16;
+    TASSIGN(v69, v70);
+    pipe_barrier(PIPE_V);
+    TCVT(v69, v67, v5);
+    Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v71 = Tile<TileType::Vec, half, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v72 = (uint64_t) v15;
+    TASSIGN(v71, v72);
+    pipe_barrier(PIPE_V);
+    TCVT(v71, v69, v5);
+    Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null> v73 = Tile<TileType::Vec, int8_t, 16, 256, BLayout::RowMajor, -1, -1, SLayout::NoneBox, 512, PadValue::Null, CompactMode::Null>(v13, v8);
+    uint64_t v74 = (uint64_t) v14;
+    TASSIGN(v73, v74);
+    pipe_barrier(PIPE_V);
+    wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+    TCVT(v73, v71, v4);
+    set_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    pto::Shape<1, 1, 1, 16, 256> v75 = pto::Shape<1, 1, 1, 16, 256>();
+    pto::Stride<65536, 65536, 65536, 4096, 1> v76 = pto::Stride<65536, 65536, 65536, 4096, 1>();
+    GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND> v77 = GlobalTensor<int8_t, pto::Shape<1, 1, 1, 16, 256>, pto::Stride<65536, 65536, 65536, 4096, 1>, pto::Layout::ND>(v2 + (v6 + v6 * (unsigned) v12 + (unsigned) v59 * (unsigned) v11), v75, v76);
+    wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID0);
+    TSTORE(v77, v73);
+    set_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  }
+  pto::Shape<1, 1, 1, 16, 1> v78 = pto::Shape<1, 1, 1, 16, 1>();
+  pto::Stride<16, 16, 16, 1, 16> v79 = pto::Stride<16, 16, 16, 1, 16>();
+  GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN> v80 = GlobalTensor<float, pto::Shape<1, 1, 1, 16, 1>, pto::Stride<16, 16, 16, 1, 16>, pto::Layout::DN>(v3 + (v6 + v6 * (unsigned) v11 + v6 * (unsigned) v13), v78, v79);
+  wait_flag(PIPE_V, PIPE_MTE3, EVENT_ID1);
+  TSTORE(v80, v54);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID0);
+  wait_flag(PIPE_V, PIPE_MTE2, EVENT_ID2);
+  wait_flag(PIPE_MTE3, PIPE_V, EVENT_ID0);
+  #endif // __DAV_VEC__
+
+  ptoas_auto_sync_tail(PTOAutoSyncTailMode::kBarrierAll);
+  return;
+}
+
+// --- Kernel entry point ---
+extern "C" __aicore__ __attribute__((always_inline)) void kernel_entry(__gm__ int64_t* args)
+{
+    // Unpack tensor: x_local__ssa_v0
+    __gm__ Tensor* x_local__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[0]);
+    __gm__ bfloat16_t* x_local__ssa_v0 = reinterpret_cast<__gm__ bfloat16_t*>(x_local__ssa_v0_tensor->buffer.addr) + x_local__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: x_local_i8_inline43__ssa_v0
+    __gm__ Tensor* x_local_i8_inline43__ssa_v0_tensor = reinterpret_cast<__gm__ Tensor*>(args[1]);
+    __gm__ int8_t* x_local_i8_inline43__ssa_v0 = reinterpret_cast<__gm__ int8_t*>(x_local_i8_inline43__ssa_v0_tensor->buffer.addr) + x_local_i8_inline43__ssa_v0_tensor->start_offset;
+
+    // Unpack tensor: ret0__out
+    __gm__ Tensor* ret0__out_tensor = reinterpret_cast<__gm__ Tensor*>(args[2]);
+    __gm__ float* ret0__out = reinterpret_cast<__gm__ float*>(ret0__out_tensor->buffer.addr) + ret0__out_tensor->start_offset;
+
+    // Forward to ptoas-generated function
+    x_local_q(x_local__ssa_v0, x_local_i8_inline43__ssa_v0, ret0__out);
+}
diff --git a/examples/workers/l3/ep_dispatch_combine/kernels/orchestration/ep_dispatch_combine_orch.cpp b/examples/workers/l3/ep_dispatch_combine/kernels/orchestration/ep_dispatch_combine_orch.cpp
index cf2f10a1e..5e00419ce 100644
--- a/examples/workers/l3/ep_dispatch_combine/kernels/orchestration/ep_dispatch_combine_orch.cpp
+++ b/examples/workers/l3/ep_dispatch_combine/kernels/orchestration/ep_dispatch_combine_orch.cpp
@@ -9,42 +9,74 @@
  * -----------------------------------------------------------------------------------------------------------
  */
 /**
- * EP dispatch + local-expert + combine orchestration.
+ * moe_router + EP dispatch + moe_expert + combine orchestration.
  *
- * Three child AIV kernels chained via the runtime; cross-kernel data flows
- * through host-backed device tensors (recv_*_out / recv_y) rather than the
- * HCCL window — only routed_y_buf and combine_done need cross-rank
- * visibility, both still in the shared scratch.
+ *   func_id 0..17   moe_router kernels  ffn half-compress pre-mix (hc_pre) +
+ *                                       RMSNorm + learned-score gate + top-k +
+ *                                       weight normalize. Produces x_norm,
+ *                                       indices, weights (plus post_ffn /
+ *                                       comb_ffn for hc_post).
+ *   func_id 18      dispatch.cpp        EP count exchange + 3-channel push;
+ *                                       reads chip-produced x_norm + indices,
+ *                                       plus host-packed w_padded / idx_padded.
+ *   func_id 19..35  moe_expert kernels  routed local experts + shared expert
+ *                                       (task ids +19; recv_expert_count read
+ *                                       from host-known tensor(11)).
+ *   func_id 36      combine.cpp         TPUT recv_y rows -> routed_y_buf;
+ *                                       reduce_sum -> routed_y.
+ *   func_id 37      ffn_add.cpp         ffn_out = routed_y + sh
+ *                                       (single-layer spec, model.py:644-645).
+ *   func_id 38      hc_post.cpp         hc_post(ffn_out, x_hc, post_ffn,
+ *                                       comb_ffn) -> y  (next-layer x_hc).
  *
- *   func_id 0  dispatch.cpp        count exchange + 3-channel push
- *                                  + stage-out + recv_count emission
- *   func_id 1  local_expert.cpp    placeholder for moe_expert:
- *                                  recv_y[e, s, :] = recv_x[e, s, :] * recv_w[e, s]
- *   func_id 2  combine.cpp         TPUT recv_y rows by recv_idx_out into
- *                                  routed_y_buf (relies on HCCL window
- *                                  zero-init), barrier, reduce_sum along
- *                                  TOPK -> routed_y FP32
- *
- *   tensor(0)  indices            INPUT             [T, TOPK]            INT32
- *   tensor(1)  x_norm             INPUT             [T, D]               BF16
- *   tensor(2)  w_padded           INPUT             [T*TOPK, W_PAD=8]    FP32
- *   tensor(3)  idx_padded         INPUT             [T*TOPK, IDX_PAD=8]  INT32
- *   tensor(4)  recv_x_out         OUTPUT_EXISTING   [L, R, D]            BF16
- *   tensor(5)  recv_w_out         OUTPUT_EXISTING   [L, R]               FP32
- *   tensor(6)  recv_idx_out       OUTPUT_EXISTING   [L, R]               INT32
- *   tensor(7)  recv_count_out     OUTPUT_EXISTING   [L, 1]               INT32
- *   tensor(8)  recv_y             OUTPUT_EXISTING   [L, R, D]            BF16
- *   tensor(9)  routed_y           OUTPUT_EXISTING   [T, D]               FP32
- *   tensor(10) scratch            INOUT             HCCL window slot
- *   scalar(0)  nranks
- *   scalar(1)  CommContext device pointer
- *
- * Tasks run sequentially because rt_submit_aiv_task dispatches in order.
- * Cross-rank synchronization: dispatch's data_done barrier ends the dispatch
- * step; combine's combine_done barrier (inside combine.cpp) ends combine.
+ *   tensor(0)   x_hc              INPUT  (host) [B, S, HC_MULT, D]    BF16
+ *   tensor(1)   hc_ffn_fn         INPUT  (host) [MIX_HC, HC_DIM]      FP32
+ *   tensor(2)   hc_ffn_scale      INPUT  (host) [3]                   FP32  (orch reads .data_as)
+ *   tensor(3)   hc_ffn_base       INPUT  (host) [MIX_HC]              FP32
+ *   tensor(4)   norm_w            INPUT  (host) [D]                   FP32
+ *   tensor(5)   gate_w            INPUT  (host) [N_EXPERTS, D]        FP32
+ *   tensor(6)   gate_bias         INPUT  (host) [N_EXPERTS]           FP32
+ *   tensor(7)   tid2eid           INPUT  (host) [VOCAB, TOPK]         INT32  (unused at LAYER_ID >= N_HASH_LAYERS)
+ *   tensor(8)   input_ids         INPUT  (host) [B, S]                INT64  (ditto)
+ *   tensor(9)   w_padded          INPUT  (host) [T*TOPK, W_PAD=8]     FP32   (packed from golden weights)
+ *   tensor(10)  idx_padded        INPUT  (host) [T*TOPK, IDX_PAD=8]   INT32  (packed from golden indices)
+ *   tensor(11)  recv_count_host   INPUT  (host) [L, 1]                INT32  (orch reads .data_as for moe_expert loop bounds)
+ *   tensor(12)  expert_w1         INPUT  (host) [L, MOE_INTER, D]     INT8
+ *   tensor(13)  expert_w1_scale   INPUT  (host) [L, MOE_INTER]        FP32
+ *   tensor(14)  expert_w3         INPUT  (host) [L, MOE_INTER, D]     INT8
+ *   tensor(15)  expert_w3_scale   INPUT  (host) [L, MOE_INTER]        FP32
+ *   tensor(16)  expert_w2         INPUT  (host) [L, D, MOE_INTER]     INT8
+ *   tensor(17)  expert_w2_scale   INPUT  (host) [L, D]                FP32
+ *   tensor(18)  shared_w1         INPUT  (host) [MOE_INTER, D]        INT8
+ *   tensor(19)  shared_w1_scale   INPUT  (host) [MOE_INTER]           FP32
+ *   tensor(20)  shared_w3         INPUT  (host) [MOE_INTER, D]        INT8
+ *   tensor(21)  shared_w3_scale   INPUT  (host) [MOE_INTER]           FP32
+ *   tensor(22)  shared_w2         INPUT  (host) [D, MOE_INTER]        INT8
+ *   tensor(23)  shared_w2_scale   INPUT  (host) [D]                   FP32
+ *   tensor(24)  x_norm            OUTPUT_EXISTING [T, D]              BF16   (router out -> dispatch + moe_expert)
+ *   tensor(25)  indices           OUTPUT_EXISTING [T, TOPK]           INT32  (router out -> dispatch)
+ *   tensor(26)  weights           OUTPUT_EXISTING [T, TOPK]           FP32   (router out; verification only)
+ *   tensor(27)  post_ffn          OUTPUT_EXISTING [B, S, HC_MULT]     FP32   (router out; unused)
+ *   tensor(28)  comb_ffn          OUTPUT_EXISTING [B, S, HC_MULT, HC_MULT] FP32 (router out; unused)
+ *   tensor(29)  recv_x_out        OUTPUT_EXISTING [L, R, D]           BF16
+ *   tensor(30)  recv_w_out        OUTPUT_EXISTING [L, R]              FP32
+ *   tensor(31)  recv_idx_out      OUTPUT_EXISTING [L, R]              INT32
+ *   tensor(32)  recv_count_out    OUTPUT_EXISTING [L, 1]              INT32
+ *   tensor(33)  recv_y            OUTPUT_EXISTING [L, R, D]           BF16
+ *   tensor(34)  sh                OUTPUT_EXISTING [T, D]              BF16
+ *   tensor(35)  routed_y          OUTPUT_EXISTING [T, D]              FP32
+ *   tensor(36)  ffn_out           OUTPUT_EXISTING [T, D]              BF16   (routed_y + sh)
+ *   tensor(37)  y                 OUTPUT_EXISTING [B, S, HC_MULT, D]  BF16   (hc_post out)
+ *   tensor(38)  scratch           INOUT  HCCL window slot
+ *   scalar(0)   nranks
+ *   scalar(1)   CommContext device pointer
  */
 
+#include "runtime.h"
+
+#include <stddef.h>
 #include <stdint.h>
+#include <stdio.h>
 
 #include "pto_orchestration_api.h"
 
@@ -54,24 +86,362 @@ __attribute__((visibility("default"))) PTO2OrchestrationConfig
 ep_dispatch_combine_orchestration_config(const ChipStorageTaskArgs &orch_args) {
     (void)orch_args;
     return PTO2OrchestrationConfig{
-        .expected_arg_count = 13,  // 11 tensors + 2 scalars
+        .expected_arg_count = 41,  // 39 tensors + 2 scalars
     };
 }
 
 __attribute__((visibility("default"))) void ep_dispatch_combine_orchestration(const ChipStorageTaskArgs &orch_args) {
-    Tensor indices = from_tensor_arg(orch_args.tensor(0));
-    Tensor x_norm = from_tensor_arg(orch_args.tensor(1));
-    Tensor w_padded = from_tensor_arg(orch_args.tensor(2));
-    Tensor idx_padded = from_tensor_arg(orch_args.tensor(3));
-    Tensor recv_x_out = from_tensor_arg(orch_args.tensor(4));
-    Tensor recv_w_out = from_tensor_arg(orch_args.tensor(5));
-    Tensor recv_idx_out = from_tensor_arg(orch_args.tensor(6));
-    Tensor recv_count_out = from_tensor_arg(orch_args.tensor(7));
-    Tensor recv_y = from_tensor_arg(orch_args.tensor(8));
-    Tensor routed_y = from_tensor_arg(orch_args.tensor(9));
-    Tensor scratch = from_tensor_arg(orch_args.tensor(10));
-
-    // child 0: dispatch
+    // ---- bind external tensors (kept under the orch's view) ----
+    Tensor ext_x_hc = from_tensor_arg(orch_args.tensor(0));
+    Tensor ext_hc_ffn_fn = from_tensor_arg(orch_args.tensor(1));
+    Tensor ext_hc_ffn_scale = from_tensor_arg(orch_args.tensor(2));
+    Tensor ext_hc_ffn_base = from_tensor_arg(orch_args.tensor(3));
+    Tensor ext_norm_w = from_tensor_arg(orch_args.tensor(4));
+    Tensor ext_gate_w = from_tensor_arg(orch_args.tensor(5));
+    Tensor ext_gate_bias = from_tensor_arg(orch_args.tensor(6));
+    Tensor ext_tid2eid = from_tensor_arg(orch_args.tensor(7));
+    Tensor ext_input_ids = from_tensor_arg(orch_args.tensor(8));
+    Tensor w_padded = from_tensor_arg(orch_args.tensor(9));
+    Tensor idx_padded = from_tensor_arg(orch_args.tensor(10));
+    // tensor(11) recv_count_host is read directly via .data_as inside the moe_expert loops.
+    Tensor x_norm = from_tensor_arg(orch_args.tensor(24));
+    Tensor indices = from_tensor_arg(orch_args.tensor(25));
+    Tensor recv_x_out = from_tensor_arg(orch_args.tensor(29));
+    Tensor recv_w_out = from_tensor_arg(orch_args.tensor(30));
+    Tensor recv_idx_out = from_tensor_arg(orch_args.tensor(31));
+    Tensor recv_count_out = from_tensor_arg(orch_args.tensor(32));
+    Tensor recv_y = from_tensor_arg(orch_args.tensor(33));
+    Tensor sh = from_tensor_arg(orch_args.tensor(34));
+    Tensor routed_y = from_tensor_arg(orch_args.tensor(35));
+    Tensor ffn_out = from_tensor_arg(orch_args.tensor(36));
+    Tensor y = from_tensor_arg(orch_args.tensor(37));
+    Tensor scratch = from_tensor_arg(orch_args.tensor(38));
+
+    // Router output names (PyPTO body refs them as ext_*).
+    Tensor ext_x_norm = x_norm;
+    Tensor ext_indices = indices;
+    Tensor ext_weights = from_tensor_arg(orch_args.tensor(26));
+    Tensor ext_post_ffn = from_tensor_arg(orch_args.tensor(27));
+    Tensor ext_comb_ffn = from_tensor_arg(orch_args.tensor(28));
+
+    // moe_expert external tensors (names preserved from the generated orch).
+    Tensor ext_recv_x = recv_x_out;
+    Tensor ext_recv_weights = recv_w_out;
+    Tensor ext_x_local = x_norm;
+    Tensor ext_expert_w1 = from_tensor_arg(orch_args.tensor(12));
+    Tensor ext_expert_w1_scale = from_tensor_arg(orch_args.tensor(13));
+    Tensor ext_expert_w3 = from_tensor_arg(orch_args.tensor(14));
+    Tensor ext_expert_w3_scale = from_tensor_arg(orch_args.tensor(15));
+    Tensor ext_expert_w2 = from_tensor_arg(orch_args.tensor(16));
+    Tensor ext_expert_w2_scale = from_tensor_arg(orch_args.tensor(17));
+    Tensor ext_shared_w1 = from_tensor_arg(orch_args.tensor(18));
+    Tensor ext_shared_w1_scale = from_tensor_arg(orch_args.tensor(19));
+    Tensor ext_shared_w3 = from_tensor_arg(orch_args.tensor(20));
+    Tensor ext_shared_w3_scale = from_tensor_arg(orch_args.tensor(21));
+    Tensor ext_shared_w2 = from_tensor_arg(orch_args.tensor(22));
+    Tensor ext_shared_w2_scale = from_tensor_arg(orch_args.tensor(23));
+    Tensor ext_recv_y = recv_y;
+    Tensor ext_sh = sh;
+
+    // ---- func_id 0..17: moe_router (transplanted PyPTO-generated body) ----
+        PTO2_SCOPE() {
+            uint32_t x_mixed_ci_shapes[3] = {16, 1, 4096};
+            TensorCreateInfo x_mixed_ci(x_mixed_ci_shapes, 3, DataType::BFLOAT16);
+            uint32_t x_flat_fp32_inline49_ci_shapes[2] = {16, 16384};
+            TensorCreateInfo x_flat_fp32_inline49_ci(x_flat_fp32_inline49_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t inv_rms_inline75_ci_shapes[2] = {1, 16};
+            TensorCreateInfo inv_rms_inline75_ci(inv_rms_inline75_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t mixes_inline34_ci_shapes[2] = {16, 32};
+            TensorCreateInfo mixes_inline34_ci(mixes_inline34_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t comb_logits_inline19_ci_shapes[2] = {16, 16};
+            TensorCreateInfo comb_logits_inline19_ci(comb_logits_inline19_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t ret0__out_ci_shapes[2] = {16, 8};
+            TensorCreateInfo ret0__out_ci(ret0__out_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t ret0__out_1_ci_shapes[2] = {16, 8};
+            TensorCreateInfo ret0__out_1_ci(ret0__out_1_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t ret0__out_2_ci_shapes[2] = {16, 8};
+            TensorCreateInfo ret0__out_2_ci(ret0__out_2_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t ret1__out_ci_shapes[2] = {16, 16};
+            TensorCreateInfo ret1__out_ci(ret1__out_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t inv_rms_inline126_ci_shapes[2] = {1, 16};
+            TensorCreateInfo inv_rms_inline126_ci(inv_rms_inline126_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t x_norm_bf16_inline111_ci_shapes[2] = {16, 4096};
+            TensorCreateInfo x_norm_bf16_inline111_ci(x_norm_bf16_inline111_ci_shapes, 2, DataType::BFLOAT16);
+            uint32_t biased_scores_inline131_ci_shapes[2] = {16, 32};
+            TensorCreateInfo biased_scores_inline131_ci(biased_scores_inline131_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t score_acc_buf_inline117_ci_shapes[2] = {1, 16};
+            TensorCreateInfo score_acc_buf_inline117_ci(score_acc_buf_inline117_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t sorted_rows_inline97_ci_shapes[2] = {16, 64};
+            TensorCreateInfo sorted_rows_inline97_ci(sorted_rows_inline97_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t topk_vals_pad_inline110_ci_shapes[2] = {16, 32};
+            TensorCreateInfo topk_vals_pad_inline110_ci(topk_vals_pad_inline110_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t topk_idx_pad_inline104_ci_shapes[2] = {16, 32};
+            TensorCreateInfo topk_idx_pad_inline104_ci(topk_idx_pad_inline104_ci_shapes, 2, DataType::INT32);
+            TaskOutputTensors alloc_0 = alloc_tensors(x_mixed_ci, x_flat_fp32_inline49_ci, inv_rms_inline75_ci, mixes_inline34_ci, comb_logits_inline19_ci, ret0__out_ci, ret0__out_1_ci, ret0__out_2_ci, ret1__out_ci, inv_rms_inline126_ci, x_norm_bf16_inline111_ci, biased_scores_inline131_ci, score_acc_buf_inline117_ci, sorted_rows_inline97_ci, topk_vals_pad_inline110_ci, topk_idx_pad_inline104_ci);
+            const Tensor& x_mixed = alloc_0.get_ref(0);
+            const Tensor& x_flat_fp32_inline49 = alloc_0.get_ref(1);
+            const Tensor& inv_rms_inline75 = alloc_0.get_ref(2);
+            const Tensor& mixes_inline34 = alloc_0.get_ref(3);
+            const Tensor& comb_logits_inline19 = alloc_0.get_ref(4);
+            const Tensor& ret0__out = alloc_0.get_ref(5);
+            const Tensor& ret0__out_1 = alloc_0.get_ref(6);
+            const Tensor& ret0__out_2 = alloc_0.get_ref(7);
+            const Tensor& ret1__out = alloc_0.get_ref(8);
+            const Tensor& inv_rms_inline126 = alloc_0.get_ref(9);
+            const Tensor& x_norm_bf16_inline111 = alloc_0.get_ref(10);
+            const Tensor& biased_scores_inline131 = alloc_0.get_ref(11);
+            const Tensor& score_acc_buf_inline117 = alloc_0.get_ref(12);
+            const Tensor& sorted_rows_inline97 = alloc_0.get_ref(13);
+            const Tensor& topk_vals_pad_inline110 = alloc_0.get_ref(14);
+            const Tensor& topk_idx_pad_inline104 = alloc_0.get_ref(15);
+            uint32_t weight_out_pad_inline108_ci_shapes[2] = {16, 32};
+            TensorCreateInfo weight_out_pad_inline108_ci(weight_out_pad_inline108_ci_shapes, 2, DataType::FLOAT32);
+            TaskOutputTensors alloc_1 = alloc_tensors(weight_out_pad_inline108_ci);
+            const Tensor& weight_out_pad_inline108 = alloc_1.get_ref(0);
+            uint32_t x_flat_inline36_shapes[2] = {16, 16384};
+            Tensor x_flat_inline36 = ext_x_hc.reshape(x_flat_inline36_shapes, 2);
+            uint32_t post_flat_inline51_shapes[1] = {64};
+            Tensor post_flat_inline51 = ext_post_ffn.reshape(post_flat_inline51_shapes, 1);
+            uint32_t comb_flat_inline37_shapes[1] = {256};
+            Tensor comb_flat_inline37 = ext_comb_ffn.reshape(comb_flat_inline37_shapes, 1);
+            for (int64_t kb_inline48 = 0; kb_inline48 < 32; kb_inline48 += 1) {
+                PTO2_SCOPE() {
+                    int64_t k0_inline59 = (kb_inline48 * 512);
+
+                    // Task 0: cast_x
+                    Arg params_t0;
+                    params_t0.add_input(x_flat_inline36);
+                    params_t0.add_output(x_flat_fp32_inline49);
+                    params_t0.add_scalar(k0_inline59);
+                    rt_submit_aiv_task(0, params_t0);
+                    const Tensor& x_flat_fp32_inline49__ssa_v3 = x_flat_fp32_inline49;
+                }
+            }
+            uint32_t ret0__out_3_shapes[2] = {1, 16};
+            uint32_t ret0__out_3_offsets[2] = {0, 0};
+            Tensor ret0__out_3 = inv_rms_inline75.view(ret0__out_3_shapes, ret0__out_3_offsets);
+
+            // Task 1: rms
+            Arg params_t1;
+            params_t1.add_input(x_flat_fp32_inline49);
+            params_t1.add_output(ret0__out_3);
+            rt_submit_aiv_task(1, params_t1);
+            uint32_t mixes_flat_inline45_shapes[1] = {512};
+            Tensor mixes_flat_inline45 = mixes_inline34.reshape(mixes_flat_inline45_shapes, 1);
+
+            // Task 2: linear
+            Arg params_t2;
+            params_t2.add_input(x_flat_fp32_inline49);
+            params_t2.add_input(ext_hc_ffn_fn);
+            params_t2.add_input(inv_rms_inline75);
+            params_t2.add_output(mixes_flat_inline45);
+            rt_submit_aiv_task(2, params_t2);
+            uint32_t mixes_v1_inline58_shapes[2] = {16, 32};
+            Tensor mixes_v1_inline58 = mixes_flat_inline45.reshape(mixes_v1_inline58_shapes, 2);
+            float scale0_inline21 = static_cast<float*>(orch_args.tensor(2).data_as<void>())[0];
+            float scale1_inline17 = static_cast<float*>(orch_args.tensor(2).data_as<void>())[1];
+            float scale2_inline26 = static_cast<float*>(orch_args.tensor(2).data_as<void>())[2];
+
+            // Task 3: split_pre_post
+            Arg params_t3;
+            params_t3.add_output(ret0__out);
+            rt_submit_aiv_task(3, params_t3);
+            const Tensor& ones_hc_inline79 = ret0__out;
+            uint32_t t_shapes[1] = {8};
+            uint32_t t_offsets[1] = {0};
+            Tensor t = ext_hc_ffn_base.view(t_shapes, t_offsets);
+            uint32_t pre_base_inline76_shapes[2] = {1, 8};
+            Tensor pre_base_inline76 = t.reshape(pre_base_inline76_shapes, 2);
+            uint32_t t__tmp_v26_shapes[2] = {16, 8};
+            uint32_t t__tmp_v26_offsets[2] = {0, 0};
+            Tensor t__tmp_v26 = mixes_v1_inline58.view(t__tmp_v26_shapes, t__tmp_v26_offsets);
+
+            // Task 4: split_pre_post_0
+            Arg params_t4;
+            params_t4.add_input(t__tmp_v26);
+            params_t4.add_input(ones_hc_inline79);
+            params_t4.add_input(pre_base_inline76);
+            params_t4.add_output(ret0__out_1);
+            params_t4.add_scalar(to_u64(scale0_inline21));
+            rt_submit_aiv_task(4, params_t4);
+            const Tensor& pre_val_inline62 = ret0__out_1;
+            uint32_t t__tmp_v33_shapes[1] = {8};
+            uint32_t t__tmp_v33_offsets[1] = {4};
+            Tensor t__tmp_v33 = ext_hc_ffn_base.view(t__tmp_v33_shapes, t__tmp_v33_offsets);
+            uint32_t post_base_inline50_shapes[2] = {1, 8};
+            Tensor post_base_inline50 = t__tmp_v33.reshape(post_base_inline50_shapes, 2);
+            uint32_t t__tmp_v34_shapes[2] = {16, 8};
+            uint32_t t__tmp_v34_offsets[2] = {0, 4};
+            Tensor t__tmp_v34 = mixes_v1_inline58.view(t__tmp_v34_shapes, t__tmp_v34_offsets);
+
+            // Task 5: split_pre_post_1
+            Arg params_t5;
+            params_t5.add_input(t__tmp_v34);
+            params_t5.add_input(ones_hc_inline79);
+            params_t5.add_input(post_base_inline50);
+            params_t5.add_output(ret0__out_2);
+            params_t5.add_output(ret1__out);
+            params_t5.add_scalar(to_u64(scale1_inline17));
+            rt_submit_aiv_task(5, params_t5);
+            const Tensor& post_pad_inline66 = ret0__out_2;
+            const Tensor& ones_comb_inline67 = ret1__out;
+            uint32_t t__tmp_v41_shapes[1] = {16};
+            uint32_t t__tmp_v41_offsets[1] = {8};
+            Tensor t__tmp_v41 = ext_hc_ffn_base.view(t__tmp_v41_shapes, t__tmp_v41_offsets);
+            uint32_t comb_base_inline70_shapes[2] = {1, 16};
+            Tensor comb_base_inline70 = t__tmp_v41.reshape(comb_base_inline70_shapes, 2);
+            uint32_t comb_mix_inline41_shapes[2] = {16, 16};
+            uint32_t comb_mix_inline41_offsets[2] = {0, 8};
+            Tensor comb_mix_inline41 = mixes_v1_inline58.view(comb_mix_inline41_shapes, comb_mix_inline41_offsets);
+            uint32_t ret0__out_4_shapes[2] = {16, 16};
+            uint32_t ret0__out_4_offsets[2] = {0, 0};
+            Tensor ret0__out_4 = comb_logits_inline19.view(ret0__out_4_shapes, ret0__out_4_offsets);
+
+            // Task 6: split_pre_post_2
+            Arg params_t6;
+            params_t6.add_input(comb_mix_inline41);
+            params_t6.add_input(ones_comb_inline67);
+            params_t6.add_input(comb_base_inline70);
+            params_t6.add_output(ret0__out_4);
+            params_t6.add_scalar(to_u64(scale2_inline26));
+            rt_submit_aiv_task(6, params_t6);
+            uint32_t post_pad_flat_inline73_shapes[1] = {128};
+            Tensor post_pad_flat_inline73 = post_pad_inline66.reshape(post_pad_flat_inline73_shapes, 1);
+
+            // Task 7: comb_sinkhorn
+            Arg params_t7;
+            params_t7.add_input(comb_logits_inline19);
+            params_t7.add_output(comb_flat_inline37);
+            rt_submit_aiv_task(7, params_t7);
+            for (int64_t t_inline7 = 0; t_inline7 < 1; t_inline7 += 1) {
+                PTO2_SCOPE() {
+
+                    // Task 8: write_post
+                    Arg params_t8;
+                    params_t8.add_input(post_pad_flat_inline73);
+                    params_t8.add_output(post_flat_inline51);
+                    params_t8.add_scalar((uint64_t)0);
+                    rt_submit_aiv_task(8, params_t8);
+                }
+            }
+            uint32_t pre_val_flat_inline44_shapes[1] = {128};
+            Tensor pre_val_flat_inline44 = pre_val_inline62.reshape(pre_val_flat_inline44_shapes, 1);
+            uint32_t x_mixed_view_inline55_shapes[2] = {16, 4096};
+            Tensor x_mixed_view_inline55 = x_mixed.reshape(x_mixed_view_inline55_shapes, 2);
+            for (int64_t t_inline5 = 0; t_inline5 < 1; t_inline5 += 1) {
+                PTO2_SCOPE() {
+
+                    // Task 9: mix_x
+                    Arg params_t9;
+                    params_t9.add_output(x_mixed_view_inline55);
+                    params_t9.add_input(pre_val_flat_inline44);
+                    params_t9.add_input(x_flat_fp32_inline49);
+                    params_t9.add_scalar((uint64_t)0);
+                    rt_submit_aiv_task(9, params_t9);
+                    const Tensor& x_mixed_view_inline55__co_l1_rv_v1 = x_mixed_view_inline55;
+                }
+            }
+            uint32_t x_mixed_v1_inline0_shapes[3] = {16, 1, 4096};
+            Tensor x_mixed_v1_inline0 = x_mixed_view_inline55.reshape(x_mixed_v1_inline0_shapes, 3);
+            uint32_t x_mixed_flat_inline127_shapes[2] = {16, 4096};
+            Tensor x_mixed_flat_inline127 = x_mixed.reshape(x_mixed_flat_inline127_shapes, 2);
+            uint32_t ret0__out_5_shapes[2] = {1, 16};
+            uint32_t ret0__out_5_offsets[2] = {0, 0};
+            Tensor ret0__out_5 = inv_rms_inline126.view(ret0__out_5_shapes, ret0__out_5_offsets);
+
+            // Task 10: ffn_norm_rms
+            Arg params_t10;
+            params_t10.add_input(x_mixed_flat_inline127);
+            params_t10.add_output(ret0__out_5);
+            rt_submit_aiv_task(10, params_t10);
+            for (int64_t db_inline125 = 0; db_inline125 < 8; db_inline125 += 1) {
+                PTO2_SCOPE() {
+
+                    // Task 11: ffn_norm_apply
+                    Arg params_t11;
+                    params_t11.add_input(inv_rms_inline126);
+                    params_t11.add_input(x_mixed_flat_inline127);
+                    params_t11.add_input(ext_norm_w);
+                    params_t11.add_output(x_norm_bf16_inline111);
+                    params_t11.add_output(ext_x_norm);
+                    params_t11.add_scalar(db_inline125);
+                    rt_submit_aiv_task(11, params_t11);
+                    const Tensor& x_norm_bf16_inline111__ssa_v3 = x_norm_bf16_inline111;
+                    const Tensor& x_norm__ssa_v3 = ext_x_norm;
+                }
+            }
+            uint32_t biased_flat_inline133_shapes[1] = {512};
+            Tensor biased_flat_inline133 = biased_scores_inline131.reshape(biased_flat_inline133_shapes, 1);
+            uint32_t ret0__out_6_shapes[2] = {16, 32};
+            uint32_t ret0__out_6_offsets[2] = {0, 0};
+            Tensor ret0__out_6 = biased_scores_inline131.view(ret0__out_6_shapes, ret0__out_6_offsets);
+
+            // Task 12: gate_dot
+            Arg params_t12;
+            params_t12.add_output(ret0__out_6);
+            rt_submit_aiv_task(12, params_t12);
+
+            // Task 13: gate_dot_0
+            Arg params_t13;
+            params_t13.add_inout(score_acc_buf_inline117);
+            params_t13.add_input(x_norm_bf16_inline111);
+            params_t13.add_input(ext_gate_w);
+            params_t13.add_input(ext_gate_bias);
+            params_t13.add_output(biased_flat_inline133);
+            rt_submit_aiv_task(13, params_t13);
+            uint32_t biased_scores_v1_inline136_shapes[2] = {16, 32};
+            Tensor biased_scores_v1_inline136 = biased_flat_inline133.reshape(biased_scores_v1_inline136_shapes, 2);
+
+            // Task 14: route_sort_top2
+            Arg params_t14;
+            params_t14.add_input(biased_scores_v1_inline136);
+            params_t14.add_inout(sorted_rows_inline97);
+            rt_submit_aiv_task(14, params_t14);
+            const Tensor& sorted_rows_inline97__ssa_v16 = sorted_rows_inline97;
+
+            // Task 15: route_extract_top2
+            Arg params_t15;
+            params_t15.add_inout(topk_vals_pad_inline110);
+            params_t15.add_input(sorted_rows_inline97__ssa_v16);
+            params_t15.add_inout(topk_idx_pad_inline104);
+            rt_submit_aiv_task(15, params_t15);
+            const Tensor& topk_vals_pad_inline110__ssa_v17 = topk_vals_pad_inline110;
+            const Tensor& topk_idx_pad_inline104__ssa_v16 = topk_idx_pad_inline104;
+
+            // Task 16: route_normalize_weights
+            Arg params_t16;
+            params_t16.add_input(topk_vals_pad_inline110__ssa_v17);
+            params_t16.add_inout(weight_out_pad_inline108);
+            rt_submit_aiv_task(16, params_t16);
+            const Tensor& weight_out_pad_inline108__ssa_v1 = weight_out_pad_inline108;
+            uint32_t indices_flat_inline89_shapes[1] = {32};
+            Tensor indices_flat_inline89 = ext_indices.reshape(indices_flat_inline89_shapes, 1);
+            uint32_t weights_flat_inline88_shapes[1] = {32};
+            Tensor weights_flat_inline88 = ext_weights.reshape(weights_flat_inline88_shapes, 1);
+            uint32_t topk_idx_flat_inline87_shapes[1] = {512};
+            Tensor topk_idx_flat_inline87 = topk_idx_pad_inline104__ssa_v16.reshape(topk_idx_flat_inline87_shapes, 1);
+            uint32_t weight_out_flat_inline86_shapes[1] = {512};
+            Tensor weight_out_flat_inline86 = weight_out_pad_inline108__ssa_v1.reshape(weight_out_flat_inline86_shapes, 1);
+
+            // Task 17: write_route_outputs.
+            // NOTE: declare the writes against the *same* C++ Tensor variables
+            // that dispatch reads (``indices`` / ``ext_weights``'s storage),
+            // not their flat reshape views. The L3 runtime tracks deps by
+            // Tensor-object identity, so writes to the reshape view don't
+            // establish a happens-before edge from this task to dispatch's
+            // later ``add_input(indices)`` — without this dispatch races with
+            // the router and reads stale (mostly-zero) indices on a fraction
+            // of runs. Buffer addr / offset are unchanged so the generated
+            // kernel still indexes the same memory.
+            Arg params_t17;
+            params_t17.add_input(topk_idx_flat_inline87);
+            params_t17.add_output(indices);
+            params_t17.add_input(weight_out_flat_inline86);
+            params_t17.add_output(ext_weights);
+            rt_submit_aiv_task(17, params_t17);
+        }
+
+    // ---- func_id 18: dispatch ----
     {
         Arg p;
         p.add_input(indices);
@@ -85,30 +455,366 @@ __attribute__((visibility("default"))) void ep_dispatch_combine_orchestration(co
         p.add_inout(scratch);
         p.add_scalar(orch_args.scalar(0));  // nranks
         p.add_scalar(orch_args.scalar(1));  // CommContext
-        rt_submit_aiv_task(0, p);
+        rt_submit_aiv_task(18, p);
     }
 
-    // child 1: local_expert (pure local, host-backed I/O only — no scratch)
-    {
-        Arg p;
-        p.add_input(recv_x_out);
-        p.add_input(recv_w_out);
-        p.add_input(recv_count_out);
-        p.add_output(recv_y);
-        p.add_scalar(orch_args.scalar(1));  // CommContext (only for ABI symmetry)
-        rt_submit_aiv_task(1, p);
-    }
+    // ---- func_id 19..35: moe_expert (transplanted PyPTO-generated body) ----
+        PTO2_SCOPE() {
+            uint32_t x_local_i8_inline43_ci_shapes[2] = {16, 4096};
+            TensorCreateInfo x_local_i8_inline43_ci(x_local_i8_inline43_ci_shapes, 2, DataType::INT8);
+            uint32_t ret0__out_ci_shapes[2] = {16, 1};
+            TensorCreateInfo ret0__out_ci(ret0__out_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t sh_tile_fp32_inline88_ci_shapes[2] = {16, 4096};
+            TensorCreateInfo sh_tile_fp32_inline88_ci(sh_tile_fp32_inline88_ci_shapes, 2, DataType::FLOAT32);
+            uint32_t sh_tile_i8_inline126_ci_shapes[2] = {16, 4096};
+            TensorCreateInfo sh_tile_i8_inline126_ci(sh_tile_i8_inline126_ci_shapes, 2, DataType::INT8);
+            uint32_t ret0__out_1_ci_shapes[2] = {16, 1};
+            TensorCreateInfo ret0__out_1_ci(ret0__out_1_ci_shapes, 2, DataType::FLOAT32);
+            TaskOutputTensors alloc_0 = alloc_tensors(x_local_i8_inline43_ci, ret0__out_ci, sh_tile_fp32_inline88_ci, sh_tile_i8_inline126_ci, ret0__out_1_ci);
+            const Tensor& x_local_i8_inline43 = alloc_0.get_ref(0);
+            const Tensor& ret0__out = alloc_0.get_ref(1);
+            const Tensor& sh_tile_fp32_inline88 = alloc_0.get_ref(2);
+            const Tensor& sh_tile_i8_inline126 = alloc_0.get_ref(3);
+            const Tensor& ret0__out_1 = alloc_0.get_ref(4);
+            uint32_t recv_y_flat_inline53_shapes[2] = {256, 4096};
+            Tensor recv_y_flat_inline53 = ext_recv_y.reshape(recv_y_flat_inline53_shapes, 2);
+            uint32_t recv_weights_flat_inline62_shapes[2] = {256, 1};
+            Tensor recv_weights_flat_inline62 = ext_recv_weights.reshape(recv_weights_flat_inline62_shapes, 2);
 
-    // child 2: combine (push to routed_y_buf in scratch, barrier, reduce_sum)
+            // Task 19: x_local_q
+            Arg params_t0;
+            params_t0.add_input(ext_x_local);
+            params_t0.add_output(x_local_i8_inline43);
+            params_t0.add_output(ret0__out);
+            rt_submit_aiv_task(19, params_t0);
+            const Tensor& x_local_i8_inline43__rv_v2 = x_local_i8_inline43;
+            const Tensor& x_local_scale_dq_inline32 = ret0__out;
+            for (int64_t local_i_inline67 = 0; local_i_inline67 < 8; local_i_inline67 += 1) {
+                PTO2_SCOPE() {
+                    size_t idx_n_rows_inline68 = local_i_inline67 * 1 + 0;
+                    int32_t n_rows_inline68 = static_cast<int32_t*>(orch_args.tensor(11).data_as<void>())[idx_n_rows_inline68];
+                    int64_t n_tiles_inline91 = ((static_cast<int64_t>(n_rows_inline68) + 15) / 16);
+                    int64_t flat_base_inline30 = (local_i_inline67 * 32);
+                    for (int64_t t_inline108 = 0; t_inline108 < n_tiles_inline91; t_inline108 += 1) {
+                        PTO2_SCOPE() {
+                            uint32_t recv_x_tile_i8_inline75_ci_shapes[2] = {16, 4096};
+                            TensorCreateInfo recv_x_tile_i8_inline75_ci(recv_x_tile_i8_inline75_ci_shapes, 2, DataType::INT8);
+                            uint32_t ret0__out_2_ci_shapes[2] = {16, 1};
+                            TensorCreateInfo ret0__out_2_ci(ret0__out_2_ci_shapes, 2, DataType::FLOAT32);
+                            uint32_t h_tile_fp32_inline18_ci_shapes[2] = {16, 4096};
+                            TensorCreateInfo h_tile_fp32_inline18_ci(h_tile_fp32_inline18_ci_shapes, 2, DataType::FLOAT32);
+                            uint32_t h_tile_i8_inline92_ci_shapes[2] = {16, 4096};
+                            TensorCreateInfo h_tile_i8_inline92_ci(h_tile_i8_inline92_ci_shapes, 2, DataType::INT8);
+                            uint32_t ret0__out_3_ci_shapes[2] = {16, 1};
+                            TensorCreateInfo ret0__out_3_ci(ret0__out_3_ci_shapes, 2, DataType::FLOAT32);
+                            TaskOutputTensors alloc_1 = alloc_tensors(recv_x_tile_i8_inline75_ci, ret0__out_2_ci, h_tile_fp32_inline18_ci, h_tile_i8_inline92_ci, ret0__out_3_ci);
+                            const Tensor& recv_x_tile_i8_inline75 = alloc_1.get_ref(0);
+                            const Tensor& ret0__out_2 = alloc_1.get_ref(1);
+                            const Tensor& h_tile_fp32_inline18 = alloc_1.get_ref(2);
+                            const Tensor& h_tile_i8_inline92 = alloc_1.get_ref(3);
+                            const Tensor& ret0__out_3 = alloc_1.get_ref(4);
+                            int64_t t0_inline47 = (t_inline108 * 16);
+                            int64_t flat_t0_inline40 = (flat_base_inline30 + t0_inline47);
+                            int64_t valid_rows_inline73 = std::min<int64_t>((static_cast<int64_t>(n_rows_inline68) - t0_inline47), 16);
+
+                            // Task 20: recv_x_q
+                            Arg params_t1;
+                            params_t1.add_input(ext_recv_x);
+                            params_t1.add_output(recv_x_tile_i8_inline75);
+                            params_t1.add_output(ret0__out_2);
+                            params_t1.add_scalar(local_i_inline67);
+                            params_t1.add_scalar(t0_inline47);
+                            rt_submit_aiv_task(20, params_t1);
+                            const Tensor& recv_x_tile_i8_inline75__rv_v2 = recv_x_tile_i8_inline75;
+                            const Tensor& recv_x_scale_dq_inline29 = ret0__out_2;
+                            for (int64_t n0_inline72 = 0; n0_inline72 < 4096; n0_inline72 += 256) {
+                                PTO2_SCOPE() {
+                                    uint32_t ret0__out_4_ci_shapes[3] = {1, 16, 256};
+                                    TensorCreateInfo ret0__out_4_ci(ret0__out_4_ci_shapes, 3, DataType::INT32);
+                                    uint32_t ret1__out_ci_shapes[3] = {1, 16, 256};
+                                    TensorCreateInfo ret1__out_ci(ret1__out_ci_shapes, 3, DataType::INT32);
+                                    uint32_t ret0__out_5_ci_shapes[2] = {16, 256};
+                                    TensorCreateInfo ret0__out_5_ci(ret0__out_5_ci_shapes, 2, DataType::FLOAT32);
+                                    uint32_t ret1__out_1_ci_shapes[2] = {16, 256};
+                                    TensorCreateInfo ret1__out_1_ci(ret1__out_1_ci_shapes, 2, DataType::FLOAT32);
+                                    uint32_t ret0__out_6_ci_shapes[2] = {16, 256};
+                                    TensorCreateInfo ret0__out_6_ci(ret0__out_6_ci_shapes, 2, DataType::FLOAT32);
+                                    TaskOutputTensors alloc_2 = alloc_tensors(ret0__out_4_ci, ret1__out_ci, ret0__out_5_ci, ret1__out_1_ci, ret0__out_6_ci);
+                                    const Tensor& ret0__out_4 = alloc_2.get_ref(0);
+                                    const Tensor& ret1__out = alloc_2.get_ref(1);
+                                    const Tensor& ret0__out_5 = alloc_2.get_ref(2);
+                                    const Tensor& ret1__out_1 = alloc_2.get_ref(3);
+                                    const Tensor& ret0__out_6 = alloc_2.get_ref(4);
+
+                                    // Task 21: exp_gate_up_matmul
+                                    Arg params_t2;
+                                    params_t2.add_input(recv_x_tile_i8_inline75__rv_v2);
+                                    params_t2.add_input(ext_expert_w1);
+                                    params_t2.add_input(ext_expert_w3);
+                                    params_t2.add_output(ret0__out_4);
+                                    params_t2.add_output(ret1__out);
+                                    params_t2.add_scalar(local_i_inline67);
+                                    params_t2.add_scalar(n0_inline72);
+                                    rt_submit_aic_task(21, params_t2);
+                                    const Tensor& gate_acc_inline69 = ret0__out_4;
+                                    const Tensor& up_acc_inline46 = ret1__out;
+
+                                    // Task 22: exp_gate_up_dequant
+                                    Arg params_t3;
+                                    params_t3.add_input(gate_acc_inline69);
+                                    params_t3.add_input(up_acc_inline46);
+                                    params_t3.add_input(ext_expert_w1_scale);
+                                    params_t3.add_input(ext_expert_w3_scale);
+                                    params_t3.add_input(recv_x_scale_dq_inline29);
+                                    params_t3.add_output(ret0__out_5);
+                                    params_t3.add_output(ret1__out_1);
+                                    params_t3.add_scalar(local_i_inline67);
+                                    params_t3.add_scalar(n0_inline72);
+                                    rt_submit_aiv_task(22, params_t3);
+                                    const Tensor& gate_2d_v1_inline7 = ret0__out_5;
+                                    const Tensor& up_2d_v1_inline78 = ret1__out_1;
+
+                                    // Task 23: exp_swiglu
+                                    Arg params_t4;
+                                    params_t4.add_input(gate_2d_v1_inline7);
+                                    params_t4.add_input(up_2d_v1_inline78);
+                                    params_t4.add_input(recv_weights_flat_inline62);
+                                    params_t4.add_output(ret0__out_6);
+                                    params_t4.add_scalar(flat_t0_inline40);
+                                    rt_submit_aiv_task(23, params_t4);
+                                    const Tensor& h_chunk_inline89 = ret0__out_6;
+
+                                    // Task 24: exp_swiglu_mask
+                                    Arg params_t5;
+                                    params_t5.add_input(h_chunk_inline89);
+                                    params_t5.add_output(h_tile_fp32_inline18);
+                                    params_t5.add_scalar(valid_rows_inline73);
+                                    params_t5.add_scalar(n0_inline72);
+                                    rt_submit_aiv_task(24, params_t5);
+                                    const Tensor& h_tile_fp32_inline18__ssa_v3 = h_tile_fp32_inline18;
+                                }
+                            }
+
+                            // Task 25: exp_h_q
+                            Arg params_t6;
+                            params_t6.add_input(h_tile_fp32_inline18);
+                            params_t6.add_output(h_tile_i8_inline92);
+                            params_t6.add_output(ret0__out_3);
+                            rt_submit_aiv_task(25, params_t6);
+                            const Tensor& h_tile_i8_inline92__rv_v2 = h_tile_i8_inline92;
+                            const Tensor& h_tile_scale_dq_inline63 = ret0__out_3;
+                            for (int64_t d0_inline49 = 0; d0_inline49 < 4096; d0_inline49 += 512) {
+                                PTO2_SCOPE() {
+                                    uint32_t ret0__out_7_ci_shapes[3] = {1, 16, 512};
+                                    TensorCreateInfo ret0__out_7_ci(ret0__out_7_ci_shapes, 3, DataType::INT32);
+                                    uint32_t ret0__out_8_ci_shapes[2] = {16, 512};
+                                    TensorCreateInfo ret0__out_8_ci(ret0__out_8_ci_shapes, 2, DataType::FLOAT32);
+                                    TaskOutputTensors alloc_3 = alloc_tensors(ret0__out_7_ci, ret0__out_8_ci);
+                                    const Tensor& ret0__out_7 = alloc_3.get_ref(0);
+                                    const Tensor& ret0__out_8 = alloc_3.get_ref(1);
+
+                                    // Task 26: exp_w2_matmul
+                                    Arg params_t7;
+                                    params_t7.add_input(h_tile_i8_inline92__rv_v2);
+                                    params_t7.add_input(ext_expert_w2);
+                                    params_t7.add_output(ret0__out_7);
+                                    params_t7.add_scalar(local_i_inline67);
+                                    params_t7.add_scalar(d0_inline49);
+                                    rt_submit_aic_task(26, params_t7);
+                                    const Tensor& y_acc_inline109 = ret0__out_7;
+
+                                    // Task 27: exp_w2_dequant
+                                    Arg params_t8;
+                                    params_t8.add_input(y_acc_inline109);
+                                    params_t8.add_input(ext_expert_w2_scale);
+                                    params_t8.add_input(h_tile_scale_dq_inline63);
+                                    params_t8.add_output(ret0__out_8);
+                                    params_t8.add_scalar(local_i_inline67);
+                                    params_t8.add_scalar(d0_inline49);
+                                    rt_submit_aiv_task(27, params_t8);
+                                    const Tensor& y_2d_v1_inline17 = ret0__out_8;
+
+                                    // Task 28: exp_recv_y_write
+                                    Arg params_t9;
+                                    params_t9.add_input(y_2d_v1_inline17);
+                                    params_t9.add_output(recv_y_flat_inline53);
+                                    params_t9.add_scalar(flat_t0_inline40);
+                                    params_t9.add_scalar(d0_inline49);
+                                    rt_submit_aiv_task(28, params_t9);
+                                    const Tensor& recv_y_flat_inline53__ssa_v7 = recv_y_flat_inline53;
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            for (int64_t n0_inline113 = 0; n0_inline113 < 4096; n0_inline113 += 256) {
+                PTO2_SCOPE() {
+                    uint32_t ret0__out_9_ci_shapes[2] = {16, 256};
+                    TensorCreateInfo ret0__out_9_ci(ret0__out_9_ci_shapes, 2, DataType::INT32);
+                    uint32_t ret1__out_2_ci_shapes[2] = {16, 256};
+                    TensorCreateInfo ret1__out_2_ci(ret1__out_2_ci_shapes, 2, DataType::INT32);
+                    uint32_t ret0__out_10_ci_shapes[2] = {16, 256};
+                    TensorCreateInfo ret0__out_10_ci(ret0__out_10_ci_shapes, 2, DataType::FLOAT32);
+                    uint32_t ret1__out_3_ci_shapes[2] = {16, 256};
+                    TensorCreateInfo ret1__out_3_ci(ret1__out_3_ci_shapes, 2, DataType::FLOAT32);
+                    TaskOutputTensors alloc_4 = alloc_tensors(ret0__out_9_ci, ret1__out_2_ci, ret0__out_10_ci, ret1__out_3_ci);
+                    const Tensor& ret0__out_9 = alloc_4.get_ref(0);
+                    const Tensor& ret1__out_2 = alloc_4.get_ref(1);
+                    const Tensor& ret0__out_10 = alloc_4.get_ref(2);
+                    const Tensor& ret1__out_3 = alloc_4.get_ref(3);
+
+                    // Task 29: sh_gate_up_matmul
+                    Arg params_t10;
+                    params_t10.add_input(x_local_i8_inline43__rv_v2);
+                    params_t10.add_input(ext_shared_w1);
+                    params_t10.add_input(ext_shared_w3);
+                    params_t10.add_output(ret0__out_9);
+                    params_t10.add_output(ret1__out_2);
+                    params_t10.add_scalar(n0_inline113);
+                    rt_submit_aic_task(29, params_t10);
+                    const Tensor& sh_gate_acc_inline95 = ret0__out_9;
+                    const Tensor& sh_up_acc_inline118 = ret1__out_2;
+
+                    // Task 30: sh_gate_up_dequant
+                    Arg params_t11;
+                    params_t11.add_input(ext_shared_w1_scale);
+                    params_t11.add_input(ext_shared_w3_scale);
+                    params_t11.add_input(sh_gate_acc_inline95);
+                    params_t11.add_input(sh_up_acc_inline118);
+                    params_t11.add_input(x_local_scale_dq_inline32);
+                    params_t11.add_output(ret0__out_10);
+                    params_t11.add_output(ret1__out_3);
+                    params_t11.add_scalar(n0_inline113);
+                    rt_submit_aiv_task(30, params_t11);
+                    const Tensor& sh_gate_v1_inline50 = ret0__out_10;
+                    const Tensor& sh_up_v1_inline9 = ret1__out_3;
+
+                    // Task 31: sh_swiglu
+                    Arg params_t12;
+                    params_t12.add_input(sh_gate_v1_inline50);
+                    params_t12.add_input(sh_up_v1_inline9);
+                    params_t12.add_output(sh_tile_fp32_inline88);
+                    params_t12.add_scalar(n0_inline113);
+                    rt_submit_aiv_task(31, params_t12);
+                    const Tensor& sh_tile_fp32_inline88__ssa_v3 = sh_tile_fp32_inline88;
+                }
+            }
+
+            // Task 32: sh_h_q
+            Arg params_t13;
+            params_t13.add_input(sh_tile_fp32_inline88);
+            params_t13.add_output(sh_tile_i8_inline126);
+            params_t13.add_output(ret0__out_1);
+            rt_submit_aiv_task(32, params_t13);
+            const Tensor& sh_tile_i8_inline126__rv_v2 = sh_tile_i8_inline126;
+            const Tensor& sh_tile_scale_dq_inline99 = ret0__out_1;
+            for (int64_t d0_inline64 = 0; d0_inline64 < 4096; d0_inline64 += 512) {
+                PTO2_SCOPE() {
+                    uint32_t ret0__out_11_ci_shapes[2] = {16, 512};
+                    TensorCreateInfo ret0__out_11_ci(ret0__out_11_ci_shapes, 2, DataType::INT32);
+                    uint32_t ret0__out_12_ci_shapes[2] = {16, 512};
+                    TensorCreateInfo ret0__out_12_ci(ret0__out_12_ci_shapes, 2, DataType::FLOAT32);
+                    TaskOutputTensors alloc_5 = alloc_tensors(ret0__out_11_ci, ret0__out_12_ci);
+                    const Tensor& ret0__out_11 = alloc_5.get_ref(0);
+                    const Tensor& ret0__out_12 = alloc_5.get_ref(1);
+
+                    // Task 33: sh_w2_matmul
+                    Arg params_t14;
+                    params_t14.add_input(sh_tile_i8_inline126__rv_v2);
+                    params_t14.add_input(ext_shared_w2);
+                    params_t14.add_output(ret0__out_11);
+                    params_t14.add_scalar(d0_inline64);
+                    rt_submit_aic_task(33, params_t14);
+                    const Tensor& sh_y_acc_inline4 = ret0__out_11;
+
+                    // Task 34: sh_w2_dequant
+                    Arg params_t15;
+                    params_t15.add_input(ext_shared_w2_scale);
+                    params_t15.add_input(sh_y_acc_inline4);
+                    params_t15.add_input(sh_tile_scale_dq_inline99);
+                    params_t15.add_output(ret0__out_12);
+                    params_t15.add_scalar(d0_inline64);
+                    rt_submit_aiv_task(34, params_t15);
+                    const Tensor& sh_y_v1_inline114 = ret0__out_12;
+
+                    // Task 35: sh_write
+                    Arg params_t16;
+                    params_t16.add_input(sh_y_v1_inline114);
+                    params_t16.add_output(ext_sh);
+                    params_t16.add_scalar(d0_inline64);
+                    rt_submit_aiv_task(35, params_t16);
+                    const Tensor& sh = ext_sh;
+                }
+            }
+        }
+
+    // ---- func_id 36: combine ----
     {
         Arg p;
         p.add_input(recv_y);
         p.add_input(recv_idx_out);
         p.add_output(routed_y);
         p.add_inout(scratch);
-        p.add_scalar(orch_args.scalar(0));
-        p.add_scalar(orch_args.scalar(1));
-        rt_submit_aiv_task(2, p);
+        p.add_scalar(orch_args.scalar(0));  // nranks
+        p.add_scalar(orch_args.scalar(1));  // CommContext
+        rt_submit_aiv_task(36, p);
+    }
+
+    // ---- func_id 37: ffn_add (ffn_out = routed_y + sh) ----
+    {
+        Arg p;
+        p.add_input(routed_y);
+        p.add_input(sh);
+        p.add_output(ffn_out);
+        rt_submit_aiv_task(37, p);
+    }
+
+    // ---- func_id 38: hc_post (transplanted PyPTO-generated body) ----
+    // Local aliases for the names the generated body refers to (`ext_x` etc.):
+    //   x        ← ffn_out (the MoE FFN result)
+    //   residual ← ext_x_hc (the router's input x_hc carries through as
+    //              residual per the single-layer spec line 99)
+    //   post / comb ← post_ffn / comb_ffn from the router
+    //   y        ← chip-allocated output tensor (next-layer x_hc)
+    {
+        Tensor ext_x = ffn_out;
+        Tensor ext_residual = ext_x_hc;
+        Tensor ext_post = ext_post_ffn;
+        Tensor ext_comb = ext_comb_ffn;
+        Tensor ext_y = y;
+
+        PTO2_SCOPE() {
+            uint32_t x_flat_inline2_shapes[2] = {16, 4096};
+            Tensor x_flat_inline2 = ext_x.reshape(x_flat_inline2_shapes, 2);
+            uint32_t residual_flat_inline6_shapes[2] = {16, 16384};
+            Tensor residual_flat_inline6 = ext_residual.reshape(residual_flat_inline6_shapes, 2);
+            uint32_t post_flat_inline8_shapes[1] = {64};
+            Tensor post_flat_inline8 = ext_post.reshape(post_flat_inline8_shapes, 1);
+            uint32_t comb_flat_inline9_shapes[1] = {256};
+            Tensor comb_flat_inline9 = ext_comb.reshape(comb_flat_inline9_shapes, 1);
+            uint32_t y_flat_inline7_shapes[2] = {16, 16384};
+            Tensor y_flat_inline7 = ext_y.reshape(y_flat_inline7_shapes, 2);
+            for (int64_t out_h_inline3 = 0; out_h_inline3 < 4; out_h_inline3 += 1) {
+                PTO2_SCOPE() {
+                    for (int64_t t_inline10 = 0; t_inline10 < 1; t_inline10 += 1) {
+                        PTO2_SCOPE() {
+
+                            // Task 38: hc_post
+                            Arg params_t0;
+                            params_t0.add_output(y_flat_inline7);
+                            params_t0.add_input(post_flat_inline8);
+                            params_t0.add_input(x_flat_inline2);
+                            params_t0.add_input(comb_flat_inline9);
+                            params_t0.add_input(residual_flat_inline6);
+                            params_t0.add_scalar((uint64_t)0);
+                            params_t0.add_scalar(out_h_inline3);
+                            rt_submit_aiv_task(38, params_t0);
+                        }
+                    }
+                }
+            }
+        }
     }
 }
 
diff --git a/examples/workers/l3/ep_dispatch_combine/main.py b/examples/workers/l3/ep_dispatch_combine/main.py
index 3ebc60f70..0ec10068f 100644
--- a/examples/workers/l3/ep_dispatch_combine/main.py
+++ b/examples/workers/l3/ep_dispatch_combine/main.py
@@ -7,51 +7,37 @@
 # INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
 # See LICENSE in the root of the software repository for the full text of the License.
 # -----------------------------------------------------------------------------------------------------------
-"""End-to-end 2-card EP dispatch + local_expert + combine demo.
-
-A single orchestration runs three child AIV kernels back-to-back over a
-shared HCCL window scratch:
-
-  dispatch.cpp      count exchange + 3-channel push + per-channel stage-out
-  local_expert.cpp  recv_y[e, s, :] = recv_x[e, s, :] * recv_w[e, s]
-                    (placeholder for the production moe_expert)
-  combine.cpp       TPUT recv_y rows by recv_idx_out into
-                    routed_y_buf[t, k, :] (relies on HCCL window zero-init,
-                    no per-call clear), barrier, reduce_sum along
-                    TOPK -> routed_y FP32
-
-Each rank drives the dispatch kernel through these phases (0..4):
-
-  histogram     scalar histogram + (dst, loc_e)-sorted route table from indices
-  publish       publish full send_counts table to peers via TNOTIFY(AtomicAdd)
-                + count_done barrier
-  prefix_sum    local prefix sums over global pub_counts (no comm)
-  payload_push  for each route: TPUT three independent payload tiles —
-                  x      [BF16, 1xD]        x_norm[t, :]
-                  weight [FP32, 1xW_PAD]    w_padded[r, :]   = [weight, 0, …, 0]
-                  idx    [INT32, 1xIDX_PAD] idx_padded[r, :] = [r, 0, …, 0]
-                                            where r = t * TOPK + k
-                to peer's recv_x[loc_e][slot, :] / recv_w[…] / recv_idx[…]
-                + data_done barrier
-  stage_out     stage out recv_x / recv_w / recv_idx windows -> host-backed outputs
-
-Type/shape contract:
-  - ``x_norm`` and ``recv_x_out`` are **BF16**. Test inputs use small
-    integer values (≤ 256) that fit BF16 exactly.
-  - Weight uses a 1xW_PAD=8 FP32 tile per route (the minimum vector tile
-    granularity = 32 B = one MTE burst). The host pre-packs each row as
-    [weight, 0, 0, …, 0]; receiver writes recv_w[loc_e][slot, :W_PAD]
-    and the kernel TROWSUM-compacts to a [L, R] FP32 host output.
-  - Idx uses the same minimum-tile rationale: 1xIDX_PAD=8 INT32 per
-    route, actual r=t*TOPK+k at slot [0]; compacted via scalar copy to
-    [L, R] INT32 host output. Combine reads it to address
-    routed_y_buf[t, k, :] without a host-built origin_map.
-  - ``recv_count_out`` is [L, 1] INT32 emitted by dispatch's prefix_sum
-    phase.
+"""End-to-end 2-card moe_router + EP dispatch + moe_expert + combine demo.
+
+A single orchestration runs four back-to-back stages over a shared HCCL window
+scratch:
+
+  moe_router kernels  FFN half-compress pre-mix (hc_pre) + RMSNorm + learned-
+                      score gate + top-k + weight normalize. Produces x_norm,
+                      indices, weights (plus post_ffn / comb_ffn for hc_post).
+                      18 PyPTO-generated AIV kernels.
+  dispatch.cpp        EP count exchange + 3-channel push (x BF16 / weight FP32 /
+                      idx INT32) + per-channel stage-out + recv_count emission.
+                      Now reads the *chip-produced* x_norm and indices, plus
+                      host-packed w_padded / idx_padded.
+  moe_expert kernels  DeepSeek-V4 decode MoE block — routed local experts
+                      (per-tile A8 gate/up matmul → SwiGLU → routing-weight mul
+                      → A8 requant → w2 matmul → recv_y) + shared expert
+                      (x_local A8 → gate/up → SwiGLU → A8 → w2 → sh). 17
+                      PyPTO-generated incore kernels (4 AIC matmuls + 13 AIV).
+  combine.cpp         TPUT recv_y rows by recv_idx_out into routed_y_buf,
+                      barrier, reduce_sum along TOPK -> routed_y FP32.
+
+Dimensions mirror the ``DEMO`` decode config: D = hidden_size = 4096,
+MOE_INTER = 4096, L = N_LOCAL_EXPERTS = 8, T = 16, TOPK = 2, R = RECV_MAX = 32,
+HC_MULT = 4. INT8 expert weight banks + FP32 router fixtures are generated
+randomly on host (shared across the two ranks). The host golden is
+``golden_moe_router`` (ported from ``models/deepseek/v4/moe_router.py``) →
+dispatch protocol replay → ``golden_moe_expert`` → combine reduce.
 
 Run:
 
-    python examples/workers/l3/ep_dispatch_combine/main.py -p a2a3sim -d 0-1
+    python examples/workers/l3/ep_dispatch_combine/main.py -p a2a3 -d 0-1
 """
 
 from __future__ import annotations
@@ -59,10 +45,12 @@
 import argparse
 import os
 import sys
+from concurrent.futures import ThreadPoolExecutor
 
 os.environ.setdefault("KMP_DUPLICATE_LIB_OK", "TRUE")
 
 import torch  # noqa: E402
+import torch.nn.functional as F  # noqa: E402
 from simpler.task_interface import (  # noqa: E402
     ArgDirection,
     CallConfig,
@@ -85,37 +73,96 @@
 from simpler_setup.torch_interop import make_tensor_arg  # noqa: E402
 
 HERE = os.path.dirname(os.path.abspath(__file__))
+RUNTIME = "tensormap_and_ringbuffer"
 
-# Demo dimensions — must mirror constants at the top of the kernel.
+# Demo dimensions — mirror the ``DEMO`` decode config and the constants at the
+# top of dispatch.cpp / combine.cpp.
 N_RANKS = 2
-T = 8
+B = 16  # DECODE_BATCH
+S = 1  # DECODE_SEQ
+T = B * S  # tokens per rank
 TOPK = 2
-D = 64
-L = 4  # N_LOCAL_EXPERTS per rank
-R = 32  # RECV_MAX (single-expert receive upper bound)
-W_PAD = 8  # weight tile width — minimum vector tile (1x8 FP32 = 32 B)
-IDX_PAD = 8  # idx tile width   — minimum vector tile (1x8 INT32 = 32 B)
-E_GLOBAL = N_RANKS * L
+D = 4096  # hidden_size
+L = 8  # N_LOCAL_EXPERTS per rank
+N_EXPERTS = L  # global experts in moe_expert's view (its JIT was compiled with EP_WORLD_SIZE=1)
+R = 32  # RECV_MAX
+MOE_INTER = 4096
+SWIGLU_LIMIT = 0.0
+INT8_SCALE_MAX = 127.0
+INT8_AMAX_EPS = 1e-4
+W_PAD = 8
+IDX_PAD = 8
+E_GLOBAL = N_RANKS * L  # global EP expert count
 N_ROUTES = T * TOPK
 
-# Window region byte sizes — mirror k*Bytes / kOff* in the kernels.
-PUB_COUNTS_BYTES = N_RANKS * N_RANKS * L * 4  # N*N*L INT32
-SIGNAL_BYTES = 64  # padded slot per signal area
-RECV_X_BYTES = L * R * D * 2  # 16 KB (BF16)
-RECV_W_BYTES = L * R * W_PAD * 4  # 4  KB (FP32; weight at slot 0)
-RECV_IDX_BYTES = L * R * IDX_PAD * 4  # 4  KB (INT32; r at slot 0)
-ROUTED_Y_BUF_BYTES = T * TOPK * D * 2  # 2  KB (BF16; combine push dest)
+# Router constants — must match the ``DEMO`` config the router was JIT'd against.
+HC_MULT = 4
+HC_DIM = HC_MULT * D  # 16384
+MIX_HC = (2 + HC_MULT) * HC_MULT  # 24
+HC_SINKHORN_ITER = 20
+HC_EPS = 1e-6
+NORM_EPS = 1e-6
+VOCAB = 129280  # vocab_size (tid2eid is unused at LAYER_ID >= N_HASH_LAYERS; still must allocate)
+ROUTE_SCALE = 1.0
+LAYER_ID = 1
+N_HASH_LAYERS = 0  # LAYER_ID >= N_HASH_LAYERS -> learned-score path
+
+# Window region byte sizes — mirror k*Bytes / kOff* in dispatch.cpp / combine.cpp.
+PUB_COUNTS_BYTES = N_RANKS * N_RANKS * L * 4
+SIGNAL_BYTES = 64
+RECV_X_BYTES = L * R * D * 2
+RECV_W_BYTES = L * R * W_PAD * 4
+RECV_IDX_BYTES = L * R * IDX_PAD * 4
+ROUTED_Y_BUF_BYTES = T * TOPK * D * 2
 SCRATCH_NBYTES = (
-    PUB_COUNTS_BYTES
-    + SIGNAL_BYTES  # count_done_sig
-    + RECV_X_BYTES
-    + RECV_W_BYTES
-    + RECV_IDX_BYTES
-    + SIGNAL_BYTES  # data_done_sig
-    + ROUTED_Y_BUF_BYTES  # combine push destination
-    + SIGNAL_BYTES  # combine_done_sig
+    PUB_COUNTS_BYTES + SIGNAL_BYTES + RECV_X_BYTES + RECV_W_BYTES + RECV_IDX_BYTES
+    + SIGNAL_BYTES + ROUTED_Y_BUF_BYTES + SIGNAL_BYTES
 )
 
+# Kernel set: 18 router (0..17), dispatch (18), 17 moe_expert (19..35), combine (36).
+# func_id matches rt_submit_*_task calls in ep_dispatch_combine_orch.cpp.
+KERNELS: list[tuple[int, str, str, str]] = [
+    (0, "cast_x", "kernels/aiv/cast_x.cpp", "aiv"),
+    (1, "rms", "kernels/aiv/rms.cpp", "aiv"),
+    (2, "linear", "kernels/aiv/linear.cpp", "aiv"),
+    (3, "split_pre_post", "kernels/aiv/split_pre_post.cpp", "aiv"),
+    (4, "split_pre_post_0", "kernels/aiv/split_pre_post_0.cpp", "aiv"),
+    (5, "split_pre_post_1", "kernels/aiv/split_pre_post_1.cpp", "aiv"),
+    (6, "split_pre_post_2", "kernels/aiv/split_pre_post_2.cpp", "aiv"),
+    (7, "comb_sinkhorn", "kernels/aiv/comb_sinkhorn.cpp", "aiv"),
+    (8, "write_post", "kernels/aiv/write_post.cpp", "aiv"),
+    (9, "mix_x", "kernels/aiv/mix_x.cpp", "aiv"),
+    (10, "ffn_norm_rms", "kernels/aiv/ffn_norm_rms.cpp", "aiv"),
+    (11, "ffn_norm_apply", "kernels/aiv/ffn_norm_apply.cpp", "aiv"),
+    (12, "gate_dot", "kernels/aiv/gate_dot.cpp", "aiv"),
+    (13, "gate_dot_0", "kernels/aiv/gate_dot_0.cpp", "aiv"),
+    (14, "route_sort_top2", "kernels/aiv/route_sort_top2.cpp", "aiv"),
+    (15, "route_extract_top2", "kernels/aiv/route_extract_top2.cpp", "aiv"),
+    (16, "route_normalize_weights", "kernels/aiv/route_normalize_weights.cpp", "aiv"),
+    (17, "write_route_outputs", "kernels/aiv/write_route_outputs.cpp", "aiv"),
+    (18, "dispatch", "kernels/aiv/dispatch.cpp", "aiv"),
+    (19, "x_local_q", "kernels/aiv/x_local_q.cpp", "aiv"),
+    (20, "recv_x_q", "kernels/aiv/recv_x_q.cpp", "aiv"),
+    (21, "exp_gate_up_matmul", "kernels/aic/exp_gate_up_matmul.cpp", "aic"),
+    (22, "exp_gate_up_dequant", "kernels/aiv/exp_gate_up_dequant.cpp", "aiv"),
+    (23, "exp_swiglu", "kernels/aiv/exp_swiglu.cpp", "aiv"),
+    (24, "exp_swiglu_mask", "kernels/aiv/exp_swiglu_mask.cpp", "aiv"),
+    (25, "exp_h_q", "kernels/aiv/exp_h_q.cpp", "aiv"),
+    (26, "exp_w2_matmul", "kernels/aic/exp_w2_matmul.cpp", "aic"),
+    (27, "exp_w2_dequant", "kernels/aiv/exp_w2_dequant.cpp", "aiv"),
+    (28, "exp_recv_y_write", "kernels/aiv/exp_recv_y_write.cpp", "aiv"),
+    (29, "sh_gate_up_matmul", "kernels/aic/sh_gate_up_matmul.cpp", "aic"),
+    (30, "sh_gate_up_dequant", "kernels/aiv/sh_gate_up_dequant.cpp", "aiv"),
+    (31, "sh_swiglu", "kernels/aiv/sh_swiglu.cpp", "aiv"),
+    (32, "sh_h_q", "kernels/aiv/sh_h_q.cpp", "aiv"),
+    (33, "sh_w2_matmul", "kernels/aic/sh_w2_matmul.cpp", "aic"),
+    (34, "sh_w2_dequant", "kernels/aiv/sh_w2_dequant.cpp", "aiv"),
+    (35, "sh_write", "kernels/aiv/sh_write.cpp", "aiv"),
+    (36, "combine", "kernels/aiv/combine.cpp", "aiv"),
+    (37, "ffn_add", "kernels/aiv/ffn_add.cpp", "aiv"),
+    (38, "hc_post", "kernels/aiv/hc_post.cpp", "aiv"),
+]
+
 
 def parse_device_range(spec: str) -> list[int]:
     if "," in spec:
@@ -131,146 +178,202 @@ def parse_device_range(spec: str) -> list[int]:
 
 
 def build_chip_callable(platform: str, pto_isa_commit: str | None) -> ChipCallable:
-    """Compile the dispatch / local_expert / combine AIV kernels + their
-    shared C++ orchestration shim into a single ChipCallable with three
-    child callables.
-    """
+    """Compile the 18 router + dispatch + 17 moe_expert + combine kernels and
+    the merged C++ orchestration into a single ChipCallable. ccec runs in a
+    thread pool because the AIC matmuls alone take real time."""
     kc = KernelCompiler(platform=platform)
-    runtime = "tensormap_and_ringbuffer"
     pto_isa_root = ensure_pto_isa_root(commit=pto_isa_commit, clone_protocol="https")
-    include_dirs = kc.get_orchestration_include_dirs(runtime)
-    kernel_include_dirs = list(include_dirs) + [str(kc.project_root / "src" / "common")]
+    include_dirs = kc.get_orchestration_include_dirs(RUNTIME)
+    is_sim = platform.endswith("sim")
+    # dispatch.cpp uses `dcci(...)` (via aicore/aicore.h) to invalidate the
+    # D-cache before its scalar read of the router-written `indices`. aicore.h
+    # transitively includes "inner_kernel.h" which lives under the platform-
+    # specific aicore subdir, so plumb that in.
+    arch = "a2a3" if platform.startswith("a2a3") else "a5"
+    inner_kernel_subdir = "sim" if is_sim else "onboard"
+    kernel_include_dirs = list(include_dirs) + [
+        str(kc.project_root / "src" / "common"),
+        str(kc.project_root / "src" / arch / "platform" / inner_kernel_subdir / "aicore"),
+    ]
 
-    def compile_aiv(name: str) -> bytes:
+    def compile_one(rel_src: str, core_type: str) -> bytes:
         b = kc.compile_incore(
-            source_path=os.path.join(HERE, "kernels/aiv", name),
-            core_type="aiv",
+            source_path=os.path.join(HERE, rel_src),
+            core_type=core_type,
             pto_isa_root=pto_isa_root,
             extra_include_dirs=kernel_include_dirs,
         )
-        if not platform.endswith("sim"):
+        if not is_sim:
             b = extract_text_section(b)
         return b
 
-    dispatch_bin = compile_aiv("dispatch.cpp")
-    local_expert_bin = compile_aiv("local_expert.cpp")
-    combine_bin = compile_aiv("combine.cpp")
+    with ThreadPoolExecutor(max_workers=min(8, len(KERNELS))) as ex:
+        futs = {fid: ex.submit(compile_one, src, ct) for (fid, _name, src, ct) in KERNELS}
+        bins = {fid: f.result() for fid, f in futs.items()}
 
     orch_bytes = kc.compile_orchestration(
-        runtime_name=runtime,
+        runtime_name=RUNTIME,
         source_path=os.path.join(HERE, "kernels/orchestration/ep_dispatch_combine_orch.cpp"),
     )
 
-    # Per-child signatures — each kernel sees only the args it actually
-    # consumes (matching the orch's `Arg` packing for that submit).
+    # dispatch / combine keep explicit per-child sigs (they're hand-written
+    # comm kernels). PyPTO-generated incore kernels get an empty signature like
+    # the other generated examples (dependency tracking lives in the orch's
+    # add_input / add_output / add_inout calls).
     sig_dispatch = [
-        ArgDirection.IN,  # indices
-        ArgDirection.IN,  # x_norm
-        ArgDirection.IN,  # w_padded
-        ArgDirection.IN,  # idx_padded
-        ArgDirection.OUT,  # recv_x_out
-        ArgDirection.OUT,  # recv_w_out
-        ArgDirection.OUT,  # recv_idx_out
-        ArgDirection.OUT,  # recv_count_out
-        ArgDirection.INOUT,  # scratch
-    ]
-    sig_local_expert = [
-        ArgDirection.IN,  # recv_x_out (reused as INPUT)
-        ArgDirection.IN,  # recv_w_out (reused as INPUT)
-        ArgDirection.IN,  # recv_count_out (reused as INPUT)
-        ArgDirection.OUT,  # recv_y
-    ]
-    sig_combine = [
-        ArgDirection.IN,  # recv_y (reused as INPUT)
-        ArgDirection.IN,  # recv_idx_out (reused as INPUT)
-        ArgDirection.OUT,  # routed_y
-        ArgDirection.INOUT,  # scratch
-    ]
-
-    # The orch's view is the union of every child's tensor footprint.
-    sig_orch = [
-        ArgDirection.IN,  # indices
-        ArgDirection.IN,  # x_norm
-        ArgDirection.IN,  # w_padded
-        ArgDirection.IN,  # idx_padded
-        ArgDirection.OUT,  # recv_x_out
-        ArgDirection.OUT,  # recv_w_out
-        ArgDirection.OUT,  # recv_idx_out
-        ArgDirection.OUT,  # recv_count_out
-        ArgDirection.OUT,  # recv_y
-        ArgDirection.OUT,  # routed_y
-        ArgDirection.INOUT,  # scratch
+        ArgDirection.IN, ArgDirection.IN, ArgDirection.IN, ArgDirection.IN,
+        ArgDirection.OUT, ArgDirection.OUT, ArgDirection.OUT, ArgDirection.OUT,
+        ArgDirection.INOUT,
     ]
+    sig_combine = [ArgDirection.IN, ArgDirection.IN, ArgDirection.OUT, ArgDirection.INOUT]
+    sig_ffn_add = [ArgDirection.IN, ArgDirection.IN, ArgDirection.OUT]
+    children: list[tuple[int, CoreCallable]] = []
+    for fid, name, _src, _ct in KERNELS:
+        if name == "dispatch":
+            sig = sig_dispatch
+        elif name == "combine":
+            sig = sig_combine
+        elif name == "ffn_add":
+            sig = sig_ffn_add
+        else:
+            sig = []
+        children.append((fid, CoreCallable.build(signature=sig, binary=bins[fid])))
+
+    # Orchestration arg view: 24 INs (x_hc..shared_w2_scale), 14 OUTPUT_EXISTING
+    # (x_norm..routed_y, ffn_out, y), 1 INOUT scratch — 39 tensors + 2 scalars.
+    sig_orch = [ArgDirection.IN] * 24 + [ArgDirection.OUT] * 14 + [ArgDirection.INOUT]
 
     return ChipCallable.build(
         signature=sig_orch,
         func_name="ep_dispatch_combine_orchestration",
         config_name="ep_dispatch_combine_orchestration_config",
         binary=orch_bytes,
-        children=[
-            (0, CoreCallable.build(signature=sig_dispatch, binary=dispatch_bin)),
-            (1, CoreCallable.build(signature=sig_local_expert, binary=local_expert_bin)),
-            (2, CoreCallable.build(signature=sig_combine, binary=combine_bin)),
-        ],
+        children=children,
     )
 
 
-def generate_routing_indices(seed: int) -> torch.Tensor:
-    """Generate `indices[N_RANKS][T, TOPK]` so no expert exceeds RECV_MAX.
-
-    Each (t, k) is a global expert id in [0, E_GLOBAL). Top-k entries within
-    a single token are forced unique. Reseed if any per-expert receive count
-    would overflow R.
-    """
-    rng = torch.Generator().manual_seed(seed)
-    while True:
-        indices = torch.zeros(N_RANKS, T, TOPK, dtype=torch.int32)
-        for r in range(N_RANKS):
-            for t in range(T):
-                perm = torch.randperm(E_GLOBAL, generator=rng)[:TOPK]
-                indices[r, t, :] = perm.to(torch.int32)
-
-        per_expert = torch.zeros(N_RANKS, L, dtype=torch.int32)
-        for r in range(N_RANKS):
-            for t in range(T):
-                for k in range(TOPK):
-                    eid = int(indices[r, t, k].item())
-                    dst = eid // L
-                    loc_e = eid % L
-                    per_expert[dst, loc_e] += 1
-        if int(per_expert.max().item()) <= R:
-            return indices
-        seed += 1
-        rng.manual_seed(seed)
-
-
-def compute_golden(
-    x_norms: list[torch.Tensor],  # [N_RANKS] of [T, D] BF16
-    indices: torch.Tensor,  # [N_RANKS, T, TOPK] INT32
-    weights: torch.Tensor,  # [N_RANKS, T, TOPK] FP32
-):
-    """Replay dispatch protocol on host -> per-rank
-    expected_recv_x[L, R, D]   BF16  (x payload)
-    expected_recv_w[L, R]      FP32  (weight payload)
-    expected_recv_idx[L, R]    INT32 (r = t*TOPK+k for each delivered row)
-    expected_count[L]          INT32
-    """
+# --------------------------------------------------------------------------- #
+# Host golden: moe_router (hc_pre + RMSNorm + learned-score gate + top-k)
+# Ported from models/deepseek/v4/{hc_pre,moe_router}.py to keep this example
+# self-contained (the pypto-lib model code isn't installed alongside the wheel).
+# --------------------------------------------------------------------------- #
+def _golden_hc_pre(x_hc, hc_ffn_fn, hc_ffn_scale, hc_ffn_base):
+    """Half-compress pre-mix: produces (x_mixed BF16 [B,S,D], post_ffn FP32
+    [B,S,HC_MULT], comb_ffn FP32 [B,S,HC_MULT,HC_MULT])."""
+    x = x_hc.float()
+    hc_fn = hc_ffn_fn.float()
+    hc_scale = hc_ffn_scale.float()
+    hc_base = hc_ffn_base.float()
+
+    x_flat = x.flatten(2).reshape(T, HC_DIM)  # [T, HC_MULT*D]
+    sq_sum = (x_flat * x_flat).sum(dim=1, keepdim=True)
+    rsqrt = torch.rsqrt(sq_sum / HC_DIM + NORM_EPS)
+    mixes = (x_flat @ hc_fn.T) * rsqrt  # [T, MIX_HC]
+    mixes = mixes.reshape(B, S, MIX_HC)
+
+    pre = torch.sigmoid(mixes[..., :HC_MULT] * hc_scale[0] + hc_base[:HC_MULT]) + HC_EPS
+    post_t = 2 * torch.sigmoid(
+        mixes[..., HC_MULT : HC_MULT * 2] * hc_scale[1] + hc_base[HC_MULT : HC_MULT * 2]
+    )
+    comb_t = (mixes[..., HC_MULT * 2 :] * hc_scale[2] + hc_base[HC_MULT * 2 :]).view(B, S, HC_MULT, HC_MULT)
+    comb_t = torch.softmax(comb_t, dim=-1) + HC_EPS
+    comb_t = comb_t / (comb_t.sum(-2, keepdim=True) + HC_EPS)
+    for _ in range(HC_SINKHORN_ITER - 1):
+        comb_t = comb_t / (comb_t.sum(-1, keepdim=True) + HC_EPS)
+        comb_t = comb_t / (comb_t.sum(-2, keepdim=True) + HC_EPS)
+
+    y = torch.zeros(B, S, D, dtype=torch.float32)
+    for h in range(HC_MULT):
+        y += x[:, :, h, :] * pre[:, :, h : h + 1]
+    return y.to(torch.bfloat16), post_t, comb_t
+
+
+def golden_moe_router(x_hc, hc_ffn_fn, hc_ffn_scale, hc_ffn_base, norm_w, gate_w, gate_bias):
+    """Router golden — returns x_norm [T,D] BF16, indices [T,TOPK] INT32,
+    weights [T,TOPK] FP32, post_ffn [B,S,HC_MULT] FP32, comb_ffn [B,S,HC_MULT,HC_MULT] FP32."""
+    x_mixed, post_ffn, comb_ffn = _golden_hc_pre(x_hc, hc_ffn_fn, hc_ffn_scale, hc_ffn_base)
+
+    # FFN RMSNorm (returns BF16 to match the chip's cast-to-BF16 store).
+    norm_w_f = norm_w.float()
+    x_f = x_mixed.float()
+    var = x_f.square().mean(-1, keepdim=True)
+    x_n = x_f * torch.rsqrt(var + NORM_EPS)
+    x_norm = (norm_w_f * x_n).to(torch.bfloat16).view(T, D)
+
+    # Learned routing scores + top-k.
+    scores = F.softplus(x_norm.float() @ gate_w.float().T).sqrt()  # [T, N_EXPERTS]
+    biased = scores + gate_bias.float()
+    indices = biased.topk(TOPK, dim=-1).indices.to(torch.int32)  # [T, TOPK]
+    weights = scores.gather(1, indices.long())  # [T, TOPK] FP32
+    weights = weights / weights.sum(dim=-1, keepdim=True) * ROUTE_SCALE
+
+    # NOTE: the router writes `indices` modulo N_EXPERTS = L (its JIT was built
+    # with EP_WORLD_SIZE=1 so it only addresses its own 8 local experts). The
+    # dispatch protocol below treats them as *global* IDs in [0, E_GLOBAL), so
+    # rebroadcast across ranks by mixing in the rank id at host fixture time.
+    return x_norm, indices, weights.float(), post_ffn, comb_ffn
+
+
+def build_router_inputs(seed: int):
+    """Per-rank x_hc / input_ids plus shared FFN-norm + gate + hc_pre fixtures."""
+    gen = torch.Generator().manual_seed(seed)
+    x_hcs = [(torch.randn(B, S, HC_MULT, D, generator=gen) * 0.1).to(torch.bfloat16).share_memory_()
+             for _ in range(N_RANKS)]
+    hc_ffn_fn = (torch.randn(MIX_HC, HC_DIM, generator=gen) / HC_DIM**0.5).contiguous().share_memory_()
+    hc_ffn_scale = (torch.ones(3, dtype=torch.float32) * 0.5).contiguous().share_memory_()
+    hc_ffn_base = torch.zeros(MIX_HC, dtype=torch.float32).contiguous().share_memory_()
+    norm_w = torch.ones(D, dtype=torch.float32).contiguous().share_memory_()
+    # gate_w shaped [N_EXPERTS, D] — N_EXPERTS = L = 8 (the moe_expert JIT was
+    # compiled with EP_WORLD_SIZE=1, so the router only addresses local experts).
+    gate_w = (torch.randn(L, D, generator=gen) / D**0.5).contiguous().share_memory_()
+    gate_bias = torch.zeros(L, dtype=torch.float32).contiguous().share_memory_()
+    # tid2eid / input_ids are unused at LAYER_ID >= N_HASH_LAYERS=0 but must
+    # still be passed (the orch binds the slots).
+    tid2eid = torch.randint(0, L, (VOCAB, TOPK), generator=gen, dtype=torch.int32).contiguous().share_memory_()
+    input_ids_list = [torch.randint(0, VOCAB, (B, S), generator=gen, dtype=torch.int64).contiguous().share_memory_()
+                      for _ in range(N_RANKS)]
+    return {
+        "x_hcs": x_hcs,
+        "hc_ffn_fn": hc_ffn_fn,
+        "hc_ffn_scale": hc_ffn_scale,
+        "hc_ffn_base": hc_ffn_base,
+        "norm_w": norm_w,
+        "gate_w": gate_w,
+        "gate_bias": gate_bias,
+        "tid2eid": tid2eid,
+        "input_ids_list": input_ids_list,
+    }
+
+
+# --------------------------------------------------------------------------- #
+# Routing / dispatch host model
+# --------------------------------------------------------------------------- #
+def _route_dst(src_rank: int, k: int) -> int:
+    """EP routing policy mirrored from dispatch.cpp: route slot k from rank
+    src_rank goes to peer (src_rank + k) % N_RANKS. The router's chip-produced
+    `indices` are local expert IDs in [0, L); the rank component is layered on
+    here so the demo spreads tokens across both ranks."""
+    return (src_rank + k) % N_RANKS
+
+
+def compute_dispatch_golden(x_norms, indices_local, weights):
+    """Replay the dispatch protocol on host. ``indices_local`` is per-rank
+    [N_RANKS, T, TOPK] of local expert IDs in [0, L); destination rank is
+    derived from (src, k) per the EP policy above."""
     expected_recv_x = [torch.zeros(L, R, D, dtype=torch.bfloat16) for _ in range(N_RANKS)]
     expected_recv_w = [torch.zeros(L, R, dtype=torch.float32) for _ in range(N_RANKS)]
     expected_recv_idx = [torch.zeros(L, R, dtype=torch.int32) for _ in range(N_RANKS)]
     expected_count = [torch.zeros(L, dtype=torch.int32) for _ in range(N_RANKS)]
+    route_dest = [[[None] * TOPK for _ in range(T)] for _ in range(N_RANKS)]
 
     send_counts = torch.zeros(N_RANKS, N_RANKS, L, dtype=torch.int32)
     for src in range(N_RANKS):
         for t in range(T):
             for k in range(TOPK):
-                eid = int(indices[src, t, k].item())
-                dst = eid // L
-                loc_e = eid % L
-                send_counts[src, dst, loc_e] += 1
+                loc_e = int(indices_local[src][t, k].item())
+                send_counts[src, _route_dst(src, k), loc_e] += 1
 
     for dst in range(N_RANKS):
-        # Per-destination slot_offset[src][e] = sum_{s < src} send_counts[s, dst, e].
         slot_offset = torch.zeros(N_RANKS, L, dtype=torch.int32)
         running = torch.zeros(L, dtype=torch.int32)
         for src in range(N_RANKS):
@@ -281,70 +384,194 @@ def compute_golden(
             cursor = torch.zeros(L, dtype=torch.int32)
             for t in range(T):
                 for k in range(TOPK):
-                    eid = int(indices[src, t, k].item())
-                    if eid // L != dst:
+                    if _route_dst(src, k) != dst:
                         continue
-                    loc_e = eid % L
+                    loc_e = int(indices_local[src][t, k].item())
                     slot = int(slot_offset[src, loc_e].item() + cursor[loc_e].item())
                     cursor[loc_e] += 1
                     expected_recv_x[dst][loc_e, slot, :] = x_norms[src][t, :]
-                    expected_recv_w[dst][loc_e, slot] = weights[src, t, k]
+                    expected_recv_w[dst][loc_e, slot] = weights[src][t, k]
                     expected_recv_idx[dst][loc_e, slot] = t * TOPK + k
+                    route_dest[src][t][k] = (dst, loc_e, slot)
 
         for e in range(L):
             expected_count[dst][e] = int(running[e].item())
 
-    return expected_recv_x, expected_recv_w, expected_recv_idx, expected_count
+    return expected_recv_x, expected_recv_w, expected_recv_idx, expected_count, route_dest
 
 
 def pack_weights_padded(weights_row: torch.Tensor) -> torch.Tensor:
-    """Build [N_ROUTES, W_PAD] FP32 where row r = (weight_value, 0, …, 0).
-
-    The kernel TPUTs row r as a 1xW_PAD tile to the receiver's
-    recv_w[loc_e][slot, :], so the actual weight ends up at recv_w[..., 0].
-    Slots [1, W_PAD) are zero — bandwidth waste vs. a true [L, R] FP32
-    output, but W_PAD=8 is the minimum vector tile size in PTO ISA.
-    """
+    """[T*TOPK, W_PAD] FP32 where row r = (weight_value, 0, …, 0)."""
     out = torch.zeros(N_ROUTES, W_PAD, dtype=torch.float32)
     for t in range(T):
         for k in range(TOPK):
-            r = t * TOPK + k
-            out[r, 0] = weights_row[t, k]
+            out[t * TOPK + k, 0] = weights_row[t, k]
     return out
 
 
 def pack_idx_padded() -> torch.Tensor:
-    """Build [N_ROUTES, IDX_PAD] INT32 where row r = (r, 0, …, 0).
-
-    Identical layout for every rank — `r = t*TOPK + k` is an intrinsic
-    label, not rank-specific. Receiver picks slot [0] in the combine
-    kernel to address routed_y_buf[t, k, :].
-    """
+    """[T*TOPK, IDX_PAD] INT32 where row r = (r, 0, …, 0)."""
     out = torch.zeros(N_ROUTES, IDX_PAD, dtype=torch.int32)
     for t in range(T):
         for k in range(TOPK):
-            r = t * TOPK + k
-            out[r, 0] = r
+            out[t * TOPK + k, 0] = t * TOPK + k
     return out
 
 
+# --------------------------------------------------------------------------- #
+# moe_expert host golden (ported from models/deepseek/v4/moe_expert.py)
+# --------------------------------------------------------------------------- #
+def _round_half_away_from_zero(x: torch.Tensor) -> torch.Tensor:
+    return torch.sign(x) * torch.floor(torch.abs(x) + 0.5)
+
+
+def _int8_quant_per_row(x: torch.Tensor):
+    rows = x.float().reshape(-1, x.shape[-1])
+    amax = rows.abs().amax(dim=-1, keepdim=True).clamp_min(INT8_AMAX_EPS)
+    scale_quant = INT8_SCALE_MAX / amax
+    out_i8 = _round_half_away_from_zero(rows * scale_quant).to(torch.int32).to(torch.float16).to(torch.int8)
+    return out_i8.reshape_as(x), (1.0 / scale_quant).reshape(*x.shape[:-1], 1)
+
+
+def _quant_w_per_channel(w: torch.Tensor):
+    amax = w.float().abs().amax(dim=-1).clamp_min(INT8_AMAX_EPS)
+    scale_quant = INT8_SCALE_MAX / amax
+    w_i8 = (_round_half_away_from_zero(w.float() * scale_quant.unsqueeze(-1))
+            .to(torch.int32).to(torch.float16).to(torch.int8))
+    return w_i8, (1.0 / scale_quant).float()
+
+
+def build_expert_weights(seed: int):
+    gen = torch.Generator().manual_seed(seed)
+    out: dict[str, torch.Tensor] = {}
+
+    def _store(name: str, w_bf16: torch.Tensor) -> None:
+        w_i8, w_s = _quant_w_per_channel(w_bf16)
+        out[name] = w_i8.contiguous().share_memory_()
+        out[name + "_scale"] = w_s.contiguous().share_memory_()
+
+    _store("expert_w1", (torch.randn(L, MOE_INTER, D, generator=gen) / D**0.5).to(torch.bfloat16))
+    _store("expert_w3", (torch.randn(L, MOE_INTER, D, generator=gen) / D**0.5).to(torch.bfloat16))
+    _store("expert_w2", (torch.randn(L, D, MOE_INTER, generator=gen) / MOE_INTER**0.5).to(torch.bfloat16))
+    _store("shared_w1", (torch.randn(MOE_INTER, D, generator=gen) / D**0.5).to(torch.bfloat16))
+    _store("shared_w3", (torch.randn(MOE_INTER, D, generator=gen) / D**0.5).to(torch.bfloat16))
+    _store("shared_w2", (torch.randn(D, MOE_INTER, generator=gen) / MOE_INTER**0.5).to(torch.bfloat16))
+    return out
+
+
+def _dequant_w(w_i8: torch.Tensor, w_scale: torch.Tensor) -> torch.Tensor:
+    return w_i8.to(torch.float32) * w_scale.unsqueeze(-1)
+
+
+def golden_moe_expert(recv_x, recv_weights, recv_count, x_local, w):
+    """Torch reference for one rank's moe_expert call (see moe_expert.py)."""
+    recv_x = recv_x.float()
+    recv_weights = recv_weights.float()
+    x_local = x_local.float()
+    w1 = _dequant_w(w["expert_w1"], w["expert_w1_scale"].float())
+    w3 = _dequant_w(w["expert_w3"], w["expert_w3_scale"].float())
+    w2 = _dequant_w(w["expert_w2"], w["expert_w2_scale"].float())
+    sw1 = _dequant_w(w["shared_w1"], w["shared_w1_scale"].float())
+    sw3 = _dequant_w(w["shared_w3"], w["shared_w3_scale"].float())
+    sw2 = _dequant_w(w["shared_w2"], w["shared_w2_scale"].float())
+
+    x_local_i8, x_local_sd = _int8_quant_per_row(x_local)
+    x_local_q = x_local_i8.float() * x_local_sd
+
+    recv_y = torch.zeros(L, R, D, dtype=torch.float32)
+    for e in range(L):
+        n_rows = int(recv_count[e].item())
+        if n_rows == 0:
+            continue
+        x_sub = recv_x[e, :n_rows, :]
+        w_sub = recv_weights[e, :n_rows]
+
+        x_sub_i8, x_sub_sd = _int8_quant_per_row(x_sub)
+        x_sub_q = x_sub_i8.float() * x_sub_sd
+
+        gate = x_sub_q @ w1[e].T
+        up = x_sub_q @ w3[e].T
+        if SWIGLU_LIMIT > 0:
+            gate = gate.clamp(max=SWIGLU_LIMIT)
+            up = up.clamp(-SWIGLU_LIMIT, SWIGLU_LIMIT)
+        h = F.silu(gate) * up
+        h = h * w_sub.unsqueeze(-1)
+        h_i8, h_sd = _int8_quant_per_row(h)
+        h = h_i8.float() * h_sd
+        recv_y[e, :n_rows, :] = h @ w2[e].T
+
+    sh_gate = x_local_q @ sw1.T
+    sh_up = x_local_q @ sw3.T
+    if SWIGLU_LIMIT > 0:
+        sh_gate = sh_gate.clamp(max=SWIGLU_LIMIT)
+        sh_up = sh_up.clamp(-SWIGLU_LIMIT, SWIGLU_LIMIT)
+    sh_h = F.silu(sh_gate) * sh_up
+    sh_h_i8, sh_h_sd = _int8_quant_per_row(sh_h)
+    sh_h = sh_h_i8.float() * sh_h_sd
+    sh = sh_h @ sw2.T
+
+    return recv_y.to(torch.bfloat16), sh.to(torch.bfloat16)
+
+
+# --------------------------------------------------------------------------- #
+# Verification
+# --------------------------------------------------------------------------- #
+def _verify_router_outputs(
+    nranks, x_norm_goldens, indices_goldens, weights_goldens,
+    x_norm_outs, indices_outs, weights_outs,
+) -> bool:
+    ok = True
+    for r in range(nranks):
+        got_xn = x_norm_outs[r].float()
+        exp_xn = x_norm_goldens[r].float()
+        d_xn = (got_xn - exp_xn).abs()
+        print(f"[ep_dispatch] chip {r}: router x_norm max|diff|={float(d_xn.max()):.3e} (BF16 tol 1e-2)")
+        if not torch.allclose(got_xn, exp_xn, rtol=1e-2, atol=1e-2):
+            ok = False
+            print(f"[ep_dispatch] chip {r}: x_norm mismatch")
+
+        # indices: rerank locally so chip[T,TOPK]'s local IDs in [0,L) align
+        # with the golden (the host extended them by rank for dispatch; the
+        # chip writes raw local IDs).
+        got_idx = indices_outs[r]
+        exp_idx_local = indices_goldens[r]  # already local IDs (pre-rerank)
+        # Compare the (sorted) set of (index, weight) pairs per row to absorb
+        # the tie-break order differences moe_router.py's topk_pair_compare
+        # handles (sort32 vs torch.topk).
+        got_pairs = sorted(zip(got_idx[r_t].tolist(), weights_outs[r][r_t].tolist())
+                           for r_t in range(0))  # placeholder; we'll do it per row below
+        idx_mismatches = 0
+        for t in range(T):
+            got_set = sorted((int(got_idx[t, k].item()), round(float(weights_outs[r][t, k].item()), 4)) for k in range(TOPK))
+            exp_set = sorted((int(exp_idx_local[t, k].item()), round(float(weights_goldens[r][t, k].item()), 4)) for k in range(TOPK))
+            # match expert set; weights within fp32 tol
+            if [s[0] for s in got_set] != [s[0] for s in exp_set]:
+                idx_mismatches += 1
+                if idx_mismatches <= 3:
+                    print(f"[ep_dispatch] chip {r} token {t}: indices got={[s[0] for s in got_set]} exp={[s[0] for s in exp_set]}")
+        if idx_mismatches > 0:
+            ok = False
+
+        got_w = weights_outs[r]
+        exp_w = weights_goldens[r]
+        d_w = (got_w - exp_w).abs()
+        print(f"[ep_dispatch] chip {r}: router weights max|diff|={float(d_w.max()):.3e}")
+        if not torch.allclose(got_w, exp_w, rtol=1e-3, atol=1e-3):
+            ok = False
+            print(f"[ep_dispatch] chip {r}: weights mismatch")
+    return ok
+
+
 def _verify_recv_outputs(
-    nranks: int,
-    expected_count: list[torch.Tensor],
-    expected_recv_x: list[torch.Tensor],
-    expected_recv_w: list[torch.Tensor],
-    expected_recv_idx: list[torch.Tensor],
-    recv_count_outs: list[torch.Tensor],
-    recv_x_outs: list[torch.Tensor],
-    recv_w_outs: list[torch.Tensor],
-    recv_idx_outs: list[torch.Tensor],
+    nranks, expected_count, expected_recv_x, expected_recv_w, expected_recv_idx,
+    recv_count_outs, recv_x_outs, recv_w_outs, recv_idx_outs,
 ) -> bool:
-    """Compare dispatch outputs against the host golden, per rank and per expert."""
+    """dispatch outputs vs the protocol replay. recv_x now compares with BF16
+    tolerance — its source x_norm goes through router rounding too."""
     ok = True
     for r in range(nranks):
         cnt = expected_count[r]
         print(f"[ep_dispatch] chip {r}: expected counts per expert = {cnt.tolist()}")
-        # recv_count_out is [L, 1] INT32 per the protocol.
         got_count = recv_count_outs[r].squeeze(-1)
         if (got_count - cnt).abs().max().item() != 0:
             ok = False
@@ -353,226 +580,349 @@ def _verify_recv_outputs(
             n = int(cnt[e].item())
             if n == 0:
                 continue
-            # Cast BF16 → FP32 for diff math; values are integer ≤ 256 so
-            # the comparison is bit-exact.
-            got_x = recv_x_outs[r][e, :n, :].to(torch.float32)
-            exp_x = expected_recv_x[r][e, :n, :].to(torch.float32)
-            got_w = recv_w_outs[r][e, :n]
-            exp_w = expected_recv_w[r][e, :n]
-            got_idx = recv_idx_outs[r][e, :n]
-            exp_idx = expected_recv_idx[r][e, :n]
+            got_x = recv_x_outs[r][e, :n, :].float()
+            exp_x = expected_recv_x[r][e, :n, :].float()
             x_diff = (got_x - exp_x).abs().max().item()
-            w_diff = (got_w - exp_w).abs().max().item()
-            idx_diff = (got_idx - exp_idx).abs().max().item()
-            if x_diff > 0 or w_diff > 1e-5 or idx_diff != 0:
+            w_diff = (recv_w_outs[r][e, :n] - expected_recv_w[r][e, :n]).abs().max().item()
+            idx_diff = (recv_idx_outs[r][e, :n] - expected_recv_idx[r][e, :n]).abs().max().item()
+            if (x_diff > 5e-2) or (w_diff > 1e-3) or (idx_diff != 0):
                 ok = False
-                print(
-                    f"[ep_dispatch] chip {r} expert {e}: cnt={n} "
-                    f"x_diff={x_diff:.3e} w_diff={w_diff:.3e} idx_diff={idx_diff}"
-                )
-                if x_diff > 0:
-                    for s in range(min(n, 3)):
-                        print(f"  slot {s}: got x[0]={float(got_x[s, 0])} expected={float(exp_x[s, 0])}")
-                if idx_diff != 0:
-                    for s in range(min(n, 3)):
-                        print(f"  slot {s}: got idx={int(got_idx[s])} expected={int(exp_idx[s])}")
+                print(f"[ep_dispatch] chip {r} expert {e}: cnt={n} x_diff={x_diff:.3e} w_diff={w_diff:.3e} idx_diff={idx_diff}")
     return ok
 
 
-def _verify_routed_y(
-    nranks: int,
-    x_norms: list[torch.Tensor],
-    weights: torch.Tensor,
-    routed_y_outs: list[torch.Tensor],
-) -> bool:
-    """Compare combine output routed_y[t, :] against the local-only golden.
-
-    routed_y[t, :] should equal sum_k weights[me, t, k] * x_norms[me][t, :]
-    since local_expert is elementwise x*weight and combine reduces along
-    TOPK. The only BF16 round-trip is local_expert's `cast(x*w, bf16)`;
-    combine's accumulator stays FP32 — mirror that exactly so the model
-    captures every cast and we can assert ~0 diff in steady state.
-    """
+def _verify_expert_outputs(nranks, recv_y_goldens, sh_goldens, expected_count, recv_y_outs, sh_outs) -> bool:
     ok = True
+    # Loose tolerance: chip x_norm differs from golden by up to 1 BF16 ulp
+    # (~1.5e-2 in this fixture), then the INT8 per-row quant amax shifts,
+    # compounded across the gate/up matmul → SwiGLU → A8 requant → w2 chain.
+    rtol, atol = 1e-2, 5e-2
     for r in range(nranks):
-        expected = torch.zeros(T, D, dtype=torch.float32)
-        for t in range(T):
-            for k in range(TOPK):
-                weighted_fp32 = weights[r, t, k] * x_norms[r][t, :].to(torch.float32)
-                expected[t, :] += weighted_fp32.to(torch.bfloat16).to(torch.float32)
-        diff = (routed_y_outs[r] - expected).abs().max().item()
-        rel_diff = diff / (expected.abs().max().item() + 1e-9)
-        print(f"[ep_dispatch] chip {r}: routed_y max|diff|={diff:.3e} (rel={rel_diff:.3e})")
-        # Allow 1e-3 abs as headroom for any fp32 reorder we missed.
-        if diff > 1e-3:
+        for e in range(L):
+            n = int(expected_count[r][e].item())
+            if n == 0:
+                continue
+            got = recv_y_outs[r][e, :n, :].float()
+            exp = recv_y_goldens[r][e, :n, :].float()
+            if not torch.allclose(got, exp, rtol=rtol, atol=atol):
+                ok = False
+                diff = (got - exp).abs()
+                print(f"[ep_dispatch] chip {r} expert {e}: recv_y mismatch n={n} max|diff|={float(diff.max()):.3e}")
+        got_sh, exp_sh = sh_outs[r].float(), sh_goldens[r].float()
+        d_sh = (got_sh - exp_sh).abs()
+        print(f"[ep_dispatch] chip {r}: sh max|diff|={float(d_sh.max()):.3e}")
+        if not torch.allclose(got_sh, exp_sh, rtol=rtol, atol=atol):
             ok = False
-            print(f"[ep_dispatch] chip {r}: routed_y mismatch (tol=1e-3)")
-            per_token_diff = (routed_y_outs[r] - expected).abs().max(dim=1).values
-            for t in range(T):
-                if per_token_diff[t] > 1e-3:
-                    print(
-                        f"  token {t}: got[0]={float(routed_y_outs[r][t, 0]):.4f} "
-                        f"expected[0]={float(expected[t, 0]):.4f}"
-                    )
+            print(f"[ep_dispatch] chip {r}: sh mismatch")
+    return ok
+
+
+def _routed_y_golden(route_dest, recv_y_goldens, rank: int) -> torch.Tensor:
+    """Reconstruct the golden routed_y[rank] = sum_k recv_y_golden[holder][loc_e, slot, :]
+    in FP32 (mirrors combine's reduce_sum FP32 accumulator)."""
+    out = torch.zeros(T, D, dtype=torch.float32)
+    for t in range(T):
+        for k in range(TOPK):
+            dst, loc_e, slot = route_dest[rank][t][k]
+            out[t, :] += recv_y_goldens[dst][loc_e, slot, :].float()
+    return out
+
+
+def _verify_routed_y(nranks, route_dest, recv_y_goldens, routed_y_outs) -> bool:
+    ok = True
+    # routed_y is the FP32 sum of TOPK BF16 recv_y rows; each row carries the
+    # same kind of compounded noise we tolerate in _verify_expert_outputs, so
+    # propagate that bound (TOPK * 5e-2 plus a small FP32 reorder buffer).
+    atol = TOPK * 5e-2 + 1e-2
+    rtol = 1e-2
+    for me in range(nranks):
+        expected = _routed_y_golden(route_dest, recv_y_goldens, me)
+        got = routed_y_outs[me]
+        diff = (got - expected).abs()
+        print(f"[ep_dispatch] chip {me}: routed_y max|diff|={float(diff.max()):.3e}")
+        if not torch.allclose(got, expected, rtol=rtol, atol=atol):
+            ok = False
+            print(f"[ep_dispatch] chip {me}: routed_y mismatch (rtol={rtol}, atol={atol})")
+    return ok
+
+
+def _golden_hc_post(x_bf16, residual_bf16, post_fp32, comb_fp32):
+    """Torch port of ``models/deepseek/v4/hc_post.py::golden_hc_post``.
+
+    Inputs:
+      x        [B, S, D]               BF16 (= ffn_out)
+      residual [B, S, HC_MULT, D]      BF16 (= moe_router input x_hc)
+      post     [B, S, HC_MULT]         FP32
+      comb     [B, S, HC_MULT, HC_MULT] FP32
+    Returns y [B, S, HC_MULT, D] BF16."""
+    x = x_bf16.float()
+    residual = residual_bf16.float()
+    post = post_fp32.float()
+    comb = comb_fp32.float()
+    y = torch.zeros(B, S, HC_MULT, D, dtype=torch.float32)
+    for out_h in range(HC_MULT):
+        y_row = x * post[:, :, out_h : out_h + 1]
+        for in_h in range(HC_MULT):
+            y_row = y_row + residual[:, :, in_h, :] * comb[:, :, in_h, out_h : out_h + 1]
+        y[:, :, out_h, :] = y_row
+    return y.to(torch.bfloat16)
+
+
+def _verify_hc_post(nranks, x_hcs, post_ffn_goldens, comb_ffn_goldens,
+                    route_dest, recv_y_goldens, sh_goldens, y_outs) -> bool:
+    """Verify hc_post output. Feed the host-side ffn_out_golden (= routed_y +
+    sh, both golden) through golden_hc_post and compare against chip y."""
+    ok = True
+    # hc_post is a 4×4 linear combination of (x, residual) with FP32 fma then a
+    # single BF16 cast. Noise budget propagates from ffn_out (~5e-2) scaled by
+    # post (~O(1)) plus the residual contribution; cap at ~1e-1.
+    atol = 1e-1
+    rtol = 1e-2
+    for me in range(nranks):
+        routed_golden = _routed_y_golden(route_dest, recv_y_goldens, me)
+        ffn_out_golden = (routed_golden + sh_goldens[me].float()).to(torch.bfloat16)
+        x_bs = ffn_out_golden.view(B, S, D)
+        expected = _golden_hc_post(x_bs, x_hcs[me], post_ffn_goldens[me], comb_ffn_goldens[me]).float()
+        got = y_outs[me].float()
+        diff = (got - expected).abs()
+        print(f"[ep_dispatch] chip {me}: y (hc_post) max|diff|={float(diff.max()):.3e}")
+        if not torch.allclose(got, expected, rtol=rtol, atol=atol):
+            ok = False
+            print(f"[ep_dispatch] chip {me}: hc_post mismatch (rtol={rtol}, atol={atol})")
+    return ok
+
+
+def _verify_ffn_out(nranks, route_dest, recv_y_goldens, sh_goldens, ffn_out_outs) -> bool:
+    """ffn_out = (routed_y + sh).to(bf16). chip uses an FP32 add then one
+    BF16 cast (CAST_RINT); golden mirrors exactly."""
+    ok = True
+    # Same noise budget as routed_y plus one BF16 ulp from the final cast — keep
+    # the routed_y bound and let the trailing 1e-2 absorb the cast.
+    atol = TOPK * 5e-2 + 2e-2
+    rtol = 1e-2
+    for me in range(nranks):
+        routed_golden = _routed_y_golden(route_dest, recv_y_goldens, me)
+        expected = (routed_golden + sh_goldens[me].float()).to(torch.bfloat16).float()
+        got = ffn_out_outs[me].float()
+        diff = (got - expected).abs()
+        print(f"[ep_dispatch] chip {me}: ffn_out max|diff|={float(diff.max()):.3e}")
+        if not torch.allclose(got, expected, rtol=rtol, atol=atol):
+            ok = False
+            print(f"[ep_dispatch] chip {me}: ffn_out mismatch (rtol={rtol}, atol={atol})")
     return ok
 
 
+# --------------------------------------------------------------------------- #
+# Driver
+# --------------------------------------------------------------------------- #
 def run(
     device_ids: list[int],
     platform: str = "a2a3",
     pto_isa_commit: str | None = None,
     build: bool = False,
+    seed: int = 20260513,
 ) -> int:
-    """Core logic — callable from CLI and pytest."""
     nranks = len(device_ids)
     assert nranks == N_RANKS
 
     window_size = max(SCRATCH_NBYTES, 128 * 1024)
-
     rootinfo_path = f"/tmp/pto_ep_dispatch_rootinfo_{os.getpid()}.bin"
     try:
         os.unlink(rootinfo_path)
     except FileNotFoundError:
         pass
 
-    print(f"[ep_dispatch] platform={platform} devices={device_ids} nranks={nranks}")
-
-    # x_norm[r, t, d] = r*100 + t*10 + d  →  max value = 1*100 + 7*10 + 63 = 233.
-    # All values are integers ≤ 256, so they fit BF16 exactly (8-bit mantissa
-    # with hidden bit = exact integers up to 2^8). The host can compare BF16
-    # outputs bit-for-bit against this golden.
-    x_norms = [
-        torch.tensor(
-            [[r * 100 + t * 10 + d for d in range(D)] for t in range(T)],
-            dtype=torch.bfloat16,
-        ).share_memory_()
-        for r in range(nranks)
-    ]
-    weights = torch.tensor(
-        [[[(r + 1) * 0.01 + t * 0.1 + k * 0.001 for k in range(TOPK)] for t in range(T)] for r in range(nranks)],
-        dtype=torch.float32,
-    )
+    print(f"[ep_dispatch] platform={platform} devices={device_ids} nranks={nranks} seed={seed}")
 
-    indices = generate_routing_indices(seed=20260510)
-    print(f"[ep_dispatch] indices shape={tuple(indices.shape)} (rank,t,k -> global expert id)")
+    # ---- host fixtures: router inputs (per-rank x_hc / input_ids + shared FFN/gate banks) ----
+    R_in = build_router_inputs(seed)
+    # ---- moe_expert weight banks (shared across both ranks) ----
+    weight_banks = build_expert_weights(seed=seed)
 
-    indices_per_rank = [indices[r].clone().contiguous().share_memory_() for r in range(nranks)]
-    w_padded_list = [pack_weights_padded(weights[r]).share_memory_() for r in range(nranks)]
-    # idx_padded is rank-independent (r = t*TOPK + k is intrinsic), but each
-    # rank gets its own shared-memory copy so the framework can plumb it as a
-    # per-rank input tensor.
+    # ---- host golden: run the router golden per rank to get x_norm / indices / weights ----
+    print("[ep_dispatch] computing host golden (router -> dispatch replay -> moe_expert -> combine)...")
+    x_norm_goldens: list[torch.Tensor] = []
+    indices_goldens: list[torch.Tensor] = []  # local IDs in [0, L)
+    weights_goldens: list[torch.Tensor] = []
+    post_ffn_goldens: list[torch.Tensor] = []
+    comb_ffn_goldens: list[torch.Tensor] = []
+    for r in range(nranks):
+        xn, idx_local, w, pf, cf = golden_moe_router(
+            R_in["x_hcs"][r], R_in["hc_ffn_fn"], R_in["hc_ffn_scale"], R_in["hc_ffn_base"],
+            R_in["norm_w"], R_in["gate_w"], R_in["gate_bias"],
+        )
+        x_norm_goldens.append(xn)
+        indices_goldens.append(idx_local)
+        weights_goldens.append(w)
+        post_ffn_goldens.append(pf)
+        comb_ffn_goldens.append(cf)
+
+    # The router's `indices` are local IDs in [0, L); destination rank is
+    # derived from (src_rank, k) per the EP routing policy that dispatch.cpp
+    # and compute_dispatch_golden share. No host-side rebroadcast needed.
+    (expected_recv_x, expected_recv_w, expected_recv_idx, expected_count, route_dest) = \
+        compute_dispatch_golden(x_norm_goldens, indices_goldens, weights_goldens)
+
+    # ---- moe_expert per-rank golden (recv_y, sh) ----
+    recv_y_goldens, sh_goldens = [], []
+    for r in range(nranks):
+        ry, shg = golden_moe_expert(
+            expected_recv_x[r], expected_recv_w[r], expected_count[r], x_norm_goldens[r], weight_banks,
+        )
+        recv_y_goldens.append(ry)
+        sh_goldens.append(shg)
+
+    # ---- pack dispatch host inputs (w_padded / idx_padded) from the GOLDEN weights/indices ----
+    # The chip's router produces near-identical weights modulo fp32 rounding; we
+    # feed the host-known versions through dispatch so recv_w / recv_idx stay
+    # deterministic.
+    w_padded_list = [pack_weights_padded(weights_goldens[r]).share_memory_() for r in range(nranks)]
     idx_padded_list = [pack_idx_padded().share_memory_() for _ in range(nranks)]
-
-    # Outputs — recv_x_out is BF16 (matches kernel TPUT/TSTORE element type);
-    # recv_w_out / recv_idx_out are compacted to [L, R] inside the kernel.
-    # recv_count_out is [L, 1] INT32 — dispatch's prefix_sum phase fills it
-    # from pub_counts so local_expert can iterate `recv_count[e]` rows per expert.
+    recv_count_host_list = [expected_count[r].reshape(L, 1).clone().contiguous().share_memory_()
+                            for r in range(nranks)]
+
+    # ---- chip output / cross-stage tensors (zero-init, OUTPUT_EXISTING) ----
+    x_norm_outs = [torch.zeros(T, D, dtype=torch.bfloat16).share_memory_() for _ in range(nranks)]
+    indices_outs = [torch.zeros(T, TOPK, dtype=torch.int32).share_memory_() for _ in range(nranks)]
+    weights_outs = [torch.zeros(T, TOPK, dtype=torch.float32).share_memory_() for _ in range(nranks)]
+    post_ffn_outs = [torch.zeros(B, S, HC_MULT, dtype=torch.float32).share_memory_() for _ in range(nranks)]
+    comb_ffn_outs = [torch.zeros(B, S, HC_MULT, HC_MULT, dtype=torch.float32).share_memory_() for _ in range(nranks)]
     recv_x_outs = [torch.zeros(L, R, D, dtype=torch.bfloat16).share_memory_() for _ in range(nranks)]
     recv_w_outs = [torch.zeros(L, R, dtype=torch.float32).share_memory_() for _ in range(nranks)]
     recv_idx_outs = [torch.zeros(L, R, dtype=torch.int32).share_memory_() for _ in range(nranks)]
     recv_count_outs = [torch.zeros(L, 1, dtype=torch.int32).share_memory_() for _ in range(nranks)]
-    # Cross-kernel host-backed tensors:
-    #   recv_y    [L, R, D]  BF16 — local_expert output, also visible to host
-    #                                for debug.
-    #   routed_y  [T, D]     FP32 — combine output; the FP32 reduce accumulator
-    #                                is written out directly without a final
-    #                                cast back to BF16 (mirrors how the
-    #                                production block keeps the FP32 accumulator
-    #                                live until exit).
     recv_y_outs = [torch.zeros(L, R, D, dtype=torch.bfloat16).share_memory_() for _ in range(nranks)]
+    sh_outs = [torch.zeros(T, D, dtype=torch.bfloat16).share_memory_() for _ in range(nranks)]
     routed_y_outs = [torch.zeros(T, D, dtype=torch.float32).share_memory_() for _ in range(nranks)]
-
-    print("[ep_dispatch] computing host golden...")
-    expected_recv_x, expected_recv_w, expected_recv_idx, expected_count = compute_golden(x_norms, indices, weights)
+    # ffn_out = routed_y + sh — final post-MoE add (single-layer spec).
+    ffn_out_outs = [torch.zeros(T, D, dtype=torch.bfloat16).share_memory_() for _ in range(nranks)]
+    # hc_post output — next-layer x_hc [B, S, HC_MULT, D] BF16.
+    y_outs = [torch.zeros(B, S, HC_MULT, D, dtype=torch.bfloat16).share_memory_() for _ in range(nranks)]
 
     cfgs = [
         ChipBootstrapConfig(
             comm=ChipCommBootstrapConfig(
-                rank=rank,
-                nranks=nranks,
-                rootinfo_path=rootinfo_path,
-                window_size=window_size,
+                rank=rank, nranks=nranks, rootinfo_path=rootinfo_path, window_size=window_size,
             ),
-            buffers=[
-                ChipBufferSpec(
-                    name="scratch",
-                    dtype="float32",
-                    count=SCRATCH_NBYTES // 4,
-                    nbytes=SCRATCH_NBYTES,
-                ),
-            ],
+            buffers=[ChipBufferSpec(name="scratch", dtype="float32", count=SCRATCH_NBYTES // 4, nbytes=SCRATCH_NBYTES)],
         )
         for rank in range(nranks)
     ]
 
-    print("[ep_dispatch] compiling kernels...")
+    print(f"[ep_dispatch] compiling orchestration + {len(KERNELS)} kernels for {platform}...")
     chip_callable = build_chip_callable(platform, pto_isa_commit)
 
     worker = Worker(
-        level=3,
-        platform=platform,
-        runtime="tensormap_and_ringbuffer",
-        device_ids=device_ids,
-        num_sub_workers=0,
-        chip_bootstrap_configs=cfgs,
-        build=build,
+        level=3, platform=platform, runtime=RUNTIME, device_ids=device_ids,
+        num_sub_workers=0, chip_bootstrap_configs=cfgs, build=build,
     )
     chip_cid = worker.register(chip_callable)
 
+    cfg = CallConfig()
+    cfg.block_dim = 24
+    cfg.aicpu_thread_num = 4
+    # Swimlane / L2 perf trace — each chip writes <output_prefix>/l2_perf_records.json,
+    # so per-chip dirs (chip-0 / chip-1) are required to avoid the second rank
+    # overwriting the first. We also emit a func_names.json mapping
+    # func_id -> kernel name so swimlane_converter can label tasks instead of
+    # showing the default ``func_xx`` placeholders.
+    swimlane_base = None
+    if os.environ.get("EP_SWIMLANE", "") == "1":
+        swimlane_base = os.environ.get("EP_SWIMLANE_DIR", os.path.join(HERE, "outputs", "swimlane"))
+        os.makedirs(swimlane_base, exist_ok=True)
+        with open(os.path.join(swimlane_base, "func_names.json"), "w") as f:
+            import json as _json
+            _json.dump({
+                "orchestrator_name": "ep_dispatch_combine_orchestration",
+                "callable_id_to_name": {str(fid): name for (fid, name, _, _) in KERNELS},
+            }, f, indent=2)
+        print(f"[ep_dispatch] swimlane enabled, base={swimlane_base}")
+
     try:
         print("[ep_dispatch] init worker (forks chip children + bootstraps HCCL)...")
         worker.init()
-
         contexts: list[ChipContext] = worker.chip_contexts
         assert len(contexts) == nranks
         for i, ctx in enumerate(contexts):
             print(
                 f"[ep_dispatch] chip {i}: device={ctx.device_id} rank={ctx.rank}/{ctx.nranks} "
-                f"window=[0x{ctx.local_window_base:x} +{ctx.actual_window_size}B] "
-                f"scratch=0x{ctx.buffer_ptrs['scratch']:x}"
+                f"window=[0x{ctx.local_window_base:x} +{ctx.actual_window_size}B] scratch=0x{ctx.buffer_ptrs['scratch']:x}"
             )
 
-        def orch_fn(orch, _args, cfg):
+        def orch_fn(orch, _args, _cfg):
             for i, ctx in enumerate(contexts):
-                chip_args = TaskArgs()
-                chip_args.add_tensor(make_tensor_arg(indices_per_rank[i]), TensorArgType.INPUT)
-                chip_args.add_tensor(make_tensor_arg(x_norms[i]), TensorArgType.INPUT)
-                chip_args.add_tensor(make_tensor_arg(w_padded_list[i]), TensorArgType.INPUT)
-                chip_args.add_tensor(make_tensor_arg(idx_padded_list[i]), TensorArgType.INPUT)
-                chip_args.add_tensor(make_tensor_arg(recv_x_outs[i]), TensorArgType.OUTPUT_EXISTING)
-                chip_args.add_tensor(make_tensor_arg(recv_w_outs[i]), TensorArgType.OUTPUT_EXISTING)
-                chip_args.add_tensor(make_tensor_arg(recv_idx_outs[i]), TensorArgType.OUTPUT_EXISTING)
-                chip_args.add_tensor(make_tensor_arg(recv_count_outs[i]), TensorArgType.OUTPUT_EXISTING)
-                chip_args.add_tensor(make_tensor_arg(recv_y_outs[i]), TensorArgType.OUTPUT_EXISTING)
-                chip_args.add_tensor(make_tensor_arg(routed_y_outs[i]), TensorArgType.OUTPUT_EXISTING)
-                chip_args.add_tensor(
+                a = TaskArgs()
+                # 9 router host inputs (0..8)
+                a.add_tensor(make_tensor_arg(R_in["x_hcs"][i]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["hc_ffn_fn"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["hc_ffn_scale"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["hc_ffn_base"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["norm_w"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["gate_w"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["gate_bias"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["tid2eid"]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(R_in["input_ids_list"][i]), TensorArgType.INPUT)
+                # 2 dispatch host inputs (9..10)
+                a.add_tensor(make_tensor_arg(w_padded_list[i]), TensorArgType.INPUT)
+                a.add_tensor(make_tensor_arg(idx_padded_list[i]), TensorArgType.INPUT)
+                # 1 host-known recv count (11)
+                a.add_tensor(make_tensor_arg(recv_count_host_list[i]), TensorArgType.INPUT)
+                # 12 moe_expert weight tensors (12..23)
+                for name in ("expert_w1", "expert_w1_scale", "expert_w3", "expert_w3_scale",
+                             "expert_w2", "expert_w2_scale", "shared_w1", "shared_w1_scale",
+                             "shared_w3", "shared_w3_scale", "shared_w2", "shared_w2_scale"):
+                    a.add_tensor(make_tensor_arg(weight_banks[name]), TensorArgType.INPUT)
+                # 13 chip OUTPUT_EXISTING tensors (24..36)
+                a.add_tensor(make_tensor_arg(x_norm_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(indices_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(weights_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(post_ffn_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(comb_ffn_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(recv_x_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(recv_w_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(recv_idx_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(recv_count_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(recv_y_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(sh_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(routed_y_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(ffn_out_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                a.add_tensor(make_tensor_arg(y_outs[i]), TensorArgType.OUTPUT_EXISTING)
+                # scratch (38)
+                a.add_tensor(
                     ContinuousTensor.make(
-                        data=ctx.buffer_ptrs["scratch"],
-                        shapes=(SCRATCH_NBYTES // 4,),
-                        dtype=DataType.FLOAT32,
-                        child_memory=True,
+                        data=ctx.buffer_ptrs["scratch"], shapes=(SCRATCH_NBYTES // 4,),
+                        dtype=DataType.FLOAT32, child_memory=True,
                     ),
                     TensorArgType.INOUT,
                 )
-                chip_args.add_scalar(ctx.nranks)
-                chip_args.add_scalar(ctx.device_ctx)
-                orch.submit_next_level(chip_cid, chip_args, cfg, worker=i)
-
-        print("[ep_dispatch] running 2-chip dispatch DAG...")
-        worker.run(orch_fn, args=None, config=CallConfig())
-
-        ok = _verify_recv_outputs(
-            nranks,
-            expected_count,
-            expected_recv_x,
-            expected_recv_w,
-            expected_recv_idx,
-            recv_count_outs,
-            recv_x_outs,
-            recv_w_outs,
-            recv_idx_outs,
+                a.add_scalar(ctx.nranks)
+                a.add_scalar(ctx.device_ctx)
+                # Per-chip cfg so each rank's swimlane lands in its own subdir.
+                cfg_i = CallConfig()
+                cfg_i.block_dim = cfg.block_dim
+                cfg_i.aicpu_thread_num = cfg.aicpu_thread_num
+                if swimlane_base is not None:
+                    cfg_i.enable_l2_swimlane = True
+                    cfg_i.output_prefix = os.path.join(swimlane_base, f"chip-{i}")
+                    os.makedirs(cfg_i.output_prefix, exist_ok=True)
+                orch.submit_next_level(chip_cid, a, cfg_i, worker=i)
+
+        print("[ep_dispatch] running 2-chip router + dispatch + moe_expert + combine DAG...")
+        worker.run(orch_fn, args=None, config=cfg)
+
+        ok = _verify_router_outputs(
+            nranks, x_norm_goldens, indices_goldens, weights_goldens,
+            x_norm_outs, indices_outs, weights_outs,
         )
-        ok = _verify_routed_y(nranks, x_norms, weights, routed_y_outs) and ok
+        ok = _verify_recv_outputs(
+            nranks, expected_count, expected_recv_x, expected_recv_w, expected_recv_idx,
+            recv_count_outs, recv_x_outs, recv_w_outs, recv_idx_outs,
+        ) and ok
+        ok = _verify_expert_outputs(nranks, recv_y_goldens, sh_goldens, expected_count, recv_y_outs, sh_outs) and ok
+        ok = _verify_routed_y(nranks, route_dest, recv_y_goldens, routed_y_outs) and ok
+        ok = _verify_ffn_out(nranks, route_dest, recv_y_goldens, sh_goldens, ffn_out_outs) and ok
+        ok = _verify_hc_post(nranks, R_in["x_hcs"], post_ffn_goldens, comb_ffn_goldens,
+                             route_dest, recv_y_goldens, sh_goldens, y_outs) and ok
 
         if not ok:
             print("[ep_dispatch] golden check FAILED")
@@ -590,14 +940,14 @@ def orch_fn(orch, _args, cfg):
 def main() -> int:
     parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
     parser.add_argument("-d", "--device", default="0-1", help="Device range, e.g. '0-1'. Two chips required.")
-    parser.add_argument("-p", "--platform", default="a2a3", help="Platform backend, e.g. a2a3 or a2a3sim.")
-    parser.add_argument(
-        "--build", action="store_true", help="Rebuild runtime from source instead of using cached libs."
-    )
+    parser.add_argument("-p", "--platform", default="a2a3", help="Platform backend.")
+    parser.add_argument("--build", action="store_true", help="Rebuild runtime from source instead of using cached libs.")
     parser.add_argument("--pto-isa-commit", default=None, help="Optional PTO ISA commit/tag to fetch before compiling.")
+    parser.add_argument("--seed", type=int, default=20260513, help="Seed for the random input fixture.")
     cli = parser.parse_args()
     return run(
-        parse_device_range(cli.device), platform=cli.platform, pto_isa_commit=cli.pto_isa_commit, build=cli.build
+        parse_device_range(cli.device), platform=cli.platform, pto_isa_commit=cli.pto_isa_commit,
+        build=cli.build, seed=cli.seed,
     )
 
 
diff --git a/examples/workers/l3/ep_dispatch_combine/test_ep_dispatch_combine.py b/examples/workers/l3/ep_dispatch_combine/test_ep_dispatch_combine.py
index 44a1bf5e6..e1dc65f5a 100644
--- a/examples/workers/l3/ep_dispatch_combine/test_ep_dispatch_combine.py
+++ b/examples/workers/l3/ep_dispatch_combine/test_ep_dispatch_combine.py
@@ -13,7 +13,9 @@
 from .main import run
 
 
-@pytest.mark.platforms(["a2a3sim", "a2a3", "a5sim"])
+# Hardware only — the moe_expert path runs D=MOE_INTER=4096 INT8 matmuls that
+# are far too slow under simulation.
+@pytest.mark.platforms(["a2a3"])
 @pytest.mark.runtime("tensormap_and_ringbuffer")
 @pytest.mark.device_count(2)
 def test_ep_dispatch_combine(st_device_ids, st_platform):