From 6a119ba4ebac43e04f7165f8456db554ec6663e7 Mon Sep 17 00:00:00 2001
From: Dmitry Ilyin <6576495+widgetii@users.noreply.github.com>
Date: Mon, 13 Apr 2026 13:58:05 +0300
Subject: [PATCH] =?UTF-8?q?openrknn:=20fix=20FP16=20output=20copy=20size?=
 =?UTF-8?q?=20=E2=80=94=20achieves=20bit-exact=20vendor=20match?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The non-4D output path copied n_elems bytes instead of ti->size bytes.
For FP16 3D tensors like SmolVLM [1,1024,768], n_elems=786432 but the
actual data size is 1,572,864 bytes (786432 × 2). This caused 50% of
the output to be zero.

Fix: use ti->size (byte count) instead of ti->n_elems (element count).

Result: SmolVLM l0_mlp output is now **bit-exact** with vendor rknnlite2
(786432/786432 FP16 values match, cosine=1.000000).

Also adds ORKNN_DUMP_ACT env var for activation BO post-run dump.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 openrknn/src/openrknn_output.c |  9 +++++----
 openrknn/src/openrknn_run.c    | 13 +++++++++++++
 2 files changed, 18 insertions(+), 4 deletions(-)

diff --git a/openrknn/src/openrknn_output.c b/openrknn/src/openrknn_output.c
index 77c1138..7226f98 100644
--- a/openrknn/src/openrknn_output.c
+++ b/openrknn/src/openrknn_output.c
@@ -159,10 +159,11 @@ int orknn_own_outputs_get(struct orknn_context *ctx, uint32_t n_outputs,
             #undef SRC_OFF
             #undef USER_OFF
         } else {
-            /* Non-4D (e.g., 2D [1,1001]): direct copy, trim padding.
-             * Native BO may be padded (e.g., 1024 for 1001 elements).
-             * Just copy the first n_elems bytes. */
-            uint32_t copy_size = ti->n_elems;
+            /* Non-4D (e.g., 2D [1,1001] or 3D [1,1024,768]): direct copy.
+             * copy_size is the tensor data size in bytes (n_elems × dtype_size).
+             * For INT8 models n_elems == byte count; for FP16 models
+             * n_elems is the element count so we need ti->size instead. */
+            uint32_t copy_size = ti->size;
             if (outputs[i].want_float) {
                 float *fdst = (float *)dst;
                 float scale = ti->scale;
diff --git a/openrknn/src/openrknn_run.c b/openrknn/src/openrknn_run.c
index 076b3ff..2098ff2 100644
--- a/openrknn/src/openrknn_run.c
+++ b/openrknn/src/openrknn_run.c
@@ -2157,6 +2157,19 @@ int orknn_own_run(struct orknn_context *ctx, rknn_run_extend *extend)
         }
     }
 
+    /* Dev: dump activation BO after run for layout analysis */
+    const char *act_dump = getenv("ORKNN_DUMP_ACT");
+    if (act_dump && ctx->activation_bo.map) {
+        orknn_bo_sync_from_device(ctx->npu_fd, &ctx->activation_bo);
+        FILE *af = fopen(act_dump, "wb");
+        if (af) {
+            fwrite(ctx->activation_bo.map, 1, ctx->activation_bo.size, af);
+            fclose(af);
+            orknn_log(1, "run: dumped act BO (%u bytes) to %s",
+                      ctx->activation_bo.size, act_dump);
+        }
+    }
+
     ctx->run_count++;
     return RKNN_SUCC;
 }