GPU Deppart Tiling API

.gitignore

@@ -21,6 +21,7 @@ install/
 .idea/
 .vscode/
 .cursor/
+.codex
 
 # clangd LSP cache
 .cache/

cmake/deppart_tmpl.cu.in

@@ -0,0 +1,20 @@
+/*
+ * Copyright 2025 Stanford University, NVIDIA Corporation
+ * SPDX-License-Identifier: Apache-2.0
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#cmakedefine INST_N1 @INST_N1@
+#cmakedefine INST_N2 @INST_N2@
+#include "@SRCFILE@_gpu_tmpl.cu"

src/CMakeLists.txt

@@ -38,7 +38,6 @@ set(REALM_SOURCES
     nodeset.cc
     operation.cc
     proc_impl.cc
-    realm_assert.cc
     repl_heap.cc
     rsrv_impl.cc
     runtime_impl.cc
@@ -64,6 +63,7 @@ set(REALM_SOURCES
     deppart/partitions.cc
     deppart/setops.cc
     deppart/sparsity_impl.cc
+    deppart/untemplated_gpu_kernels.cu
     numa/numa_module.cc
     numa/numasysif.cc
     procset/procset_module.cc
@@ -120,7 +120,7 @@ if(REALM_USE_UCX)
 endif()
 
 if(REALM_USE_GASNETEX)
-  if (NOT REALM_ENABLE_GASNETEX_WRAPPER)
+  if(NOT REALM_ENABLE_GASNETEX_WRAPPER)
     list(APPEND REALM_SOURCES gasnet1/gasnet1_module.cc gasnet1/gasnetmsg.cc)
   endif()
   list(APPEND REALM_SOURCES gasnetex/gasnetex_module.cc gasnetex/gasnetex_internal.cc)
@@ -145,7 +145,7 @@ configure_file(
   @ONLY
 )
 
-# generate per-dimension object files for deppart stuff
+# Generate per-dimension object files for CPU deppart.
 foreach(INST_N1 RANGE 1 ${REALM_MAX_DIM})
   foreach(INST_N2 RANGE 1 ${REALM_MAX_DIM})
     foreach(SRCFILE realm/deppart/image realm/deppart/preimage realm/deppart/byfield)
@@ -157,6 +157,18 @@ foreach(INST_N1 RANGE 1 ${REALM_MAX_DIM})
   endforeach()
 endforeach()
 
+# Generate per-dimension object files for GPU deppart.
+foreach(INST_N1 RANGE 1 ${REALM_MAX_DIM})
+    foreach(INST_N2 RANGE 1 ${REALM_MAX_DIM})
+        foreach(SRCFILE realm/deppart/byfield realm/deppart/image realm/deppart/preimage)
+            set(_result_file "${CMAKE_CURRENT_BINARY_DIR}/${SRCFILE}_gpu_${INST_N1}_${INST_N2}.cu")
+            # use cmake's configure_file for a portable way of creating wrapper source files
+            configure_file("${PROJECT_SOURCE_DIR}/cmake/deppart_tmpl.cu.in" "${_result_file}")
+            list(APPEND REALM_SOURCES "${_result_file}")
+        endforeach()
+    endforeach()
+endforeach()
+
 set(REALM_SOURCES
     ${REALM_SOURCES}
     PARENT_SCOPE

src/realm/cuda/cuda_internal.h

@@ -412,6 +412,7 @@ namespace Realm {
       get_null_task_stream(void) const; // needed by librealm_kokkos.so
       GPUStream *get_next_task_stream(bool create = false);
       GPUStream *get_next_d2d_stream();
+      GPUStream *get_deppart_stream() const;
 
       void launch_batch_affine_fill_kernel(void *fill_info, size_t dim, size_t elemSize,
                                            size_t volume, GPUStream *stream);
@@ -489,6 +490,8 @@ namespace Realm {
       GPUStream *host_to_device_stream = nullptr;
       GPUStream *device_to_host_stream = nullptr;
       GPUStream *device_to_device_stream = nullptr;
+      GPUStream *deppart_stream = nullptr;
+
       std::vector<GPUStream *> device_to_device_streams;
       std::vector<GPUStream *> peer_to_peer_streams; // indexed by target
       std::vector<GPUStream *> task_streams;

src/realm/cuda/cuda_module.cc

@@ -1058,6 +1058,11 @@ namespace Realm {
       return device_to_device_streams[d2d_stream_index];
     }
 
+    GPUStream *GPU::get_deppart_stream() const
+    {
+      return deppart_stream;
+    }
+
     static void launch_kernel(const Realm::Cuda::GPU::GPUFuncInfo &func_info,
                               void *params, size_t num_elems, GPUStream *stream)
     {
@@ -2040,6 +2045,7 @@ namespace Realm {
 
       host_to_device_stream = new GPUStream(this, worker);
       device_to_host_stream = new GPUStream(this, worker);
+      deppart_stream = new GPUStream(this, worker);
 
       CUdevice dev;
       int numSMs;
@@ -2164,6 +2170,7 @@ namespace Realm {
       // destroy streams
       delete host_to_device_stream;
       delete device_to_host_stream;
+      delete deppart_stream;
 
       delete_container_contents(device_to_device_streams);