openrknn: fix FP16 output copy — bit-exact vendor match

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;

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;
 }