style(atlas): add comment for async d2d

GitOrigin-RevId: 606a56ac4e

dnn/src/atlas/megcore/computing_context.cpp

@@ -51,6 +51,7 @@ void AtlasComputingContext::memcpy(void* dst, const void* src,
                         ACL_MEMCPY_HOST_TO_DEVICE));
             break;
         case megcoreMemcpyDeviceToDevice:
+            // async d2d is always faster than sync d2d because of SDMA
             acl_check(aclrtMemcpyAsync(dst, size_in_bytes, src, size_in_bytes,
                 ACL_MEMCPY_DEVICE_TO_DEVICE, m_ctx.stream));
             break;

src/core/impl/comp_node/atlas/comp_node.cpp

@@ -230,14 +230,10 @@ void AtlasCompNodeImpl::peer_copy_to(Impl* dest_impl, void* dest,
         auto&& src_env = m_env.atlas_env();
         activate();
         if (dst_env.device == src_env.device) {
-#if 1
+            // async d2d use SDMA which is faster than sync ctrl cpu d2d 
             MGB_ATLAS_CHECK(aclrtMemcpyAsync(dest, size, src, size,
                                              ACL_MEMCPY_DEVICE_TO_DEVICE,
                                              dst_env.stream));
-#else
-            MGB_ATLAS_CHECK(aclrtMemcpy(dest, size, src, size,
-                                             ACL_MEMCPY_DEVICE_TO_DEVICE));
-#endif
         } else {
             mgb_throw(MegBrainError,
                       "Atlas does not support peer copy between differents "

src/opr/impl/atlas_runtime_op.cpp

@@ -361,7 +361,6 @@ void AtlasRuntimeOpr::scn_do_execute() {
                        i, output(i)->cname());
             aclmdlAddDatasetBuffer(model_outputs, output_db);
         }
-
         MGB_ATLAS_CHECK(aclmdlExecute(m_model_id, model_inputs, model_outputs));
 
         for (size_t i = 0; i < nr_inputs; ++i) {