Don't attempt d2h copy if owner is the host

Signed-off-by: Joseph Schuchart <[email protected]>
devreal · Sep 16, 2024 · 7193769 · 7193769
1 parent 9e3dd92
commit 7193769
Showing 1 changed file with 8 additions and 7 deletions.
diff --git a/ttg/ttg/parsec/devicefunc.h b/ttg/ttg/parsec/devicefunc.h
@@ -130,13 +130,14 @@ namespace ttg_parsec {
       parsec_gpu_task_t *gpu_task = detail::parsec_ttg_caller->dev_ptr->gpu_task;
       parsec_gpu_exec_stream_t *stream = detail::parsec_ttg_caller->dev_ptr->stream;
 
-      /* enqueue the transfer into the compute stream to come back once the compute and transfer are complete */
-      parsec_device_gpu_module_t *device_module = detail::parsec_ttg_caller->dev_ptr->device;
-      device_module->memcpy_async(device_module, stream,
-                                  data->device_copies[0]->device_private,
-                                  data->device_copies[data->owner_device]->device_private,
-                                  data->nb_elts, parsec_device_gpu_transfer_direction_d2h);
-
+      if (data->owner_device != 0) {
+        /* enqueue the transfer into the compute stream to come back once the compute and transfer are complete */
+        parsec_device_gpu_module_t *device_module = detail::parsec_ttg_caller->dev_ptr->device;
+        device_module->memcpy_async(device_module, stream,
+                                    data->device_copies[0]->device_private,
+                                    data->device_copies[data->owner_device]->device_private,
+                                    data->nb_elts, parsec_device_gpu_transfer_direction_d2h);
+      }
       if constexpr (sizeof...(Is) > 0) {
         // recursion
         mark_device_out(views, std::index_sequence<Is...>{});