Merge branch 'master' into as/npuw_parallel_for_improve

openvinotoolkit · Oct 3, 2024 · d9d389e · d9d389e
2 parents b9efe9c + 890f2e1
commit d9d389e
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 8 deletions.
diff --git a/src/plugins/intel_gpu/src/graph/layout_optimizer.cpp b/src/plugins/intel_gpu/src/graph/layout_optimizer.cpp
@@ -1089,11 +1089,13 @@ format layout_optimizer::get_expected_format(quantize_node const& node) {
     auto use_onednn_impls = _optimization_attributes.use_onednn_impls;
 
     if (use_onednn_impls) {
-        auto& user = node.get_users().front();
-        if (user != nullptr && user->get_preferred_input_fmt(user->get_dependency_index(node)) != format::any) {
-            expected = user->get_preferred_input_fmt(user->get_dependency_index(node));
-        } else {
-            expected = format::any;
+        expected = format::any;
+        auto& users = node.get_users();
+        if (users.size() != 0) {
+            auto& user = users.front();
+            if (user != nullptr && user->get_preferred_input_fmt(user->get_dependency_index(node)) != format::any) {
+                expected = user->get_preferred_input_fmt(user->get_dependency_index(node));
+            }
         }
     } else if (only_gemm_users(node)) {
         // TODO: Gemm is not supporting fsv layouts

diff --git a/src/plugins/intel_gpu/src/plugin/sync_infer_request.cpp b/src/plugins/intel_gpu/src/plugin/sync_infer_request.cpp
@@ -34,9 +34,12 @@ namespace {
 inline bool can_use_usm_host(const cldnn::engine& engine) {
     auto can_use_usm = engine.use_unified_shared_memory();
 
-    if (engine.get_device_info().gfx_ver.major == 12 && engine.get_device_info().gfx_ver.minor == 60) {
-        // WA: Disable USM host memory for infer request`s tensors for PVC as
-        // it has performance issues in case of host <-> device data transfers inside kernels
+    const auto& device_info = engine.get_device_info();
+    if ((device_info.gfx_ver.major == 12 && device_info.gfx_ver.minor == 60) ||
+        (device_info.gfx_ver.major >= 20 && device_info.dev_type == cldnn::device_type::discrete_gpu)) {
+        // WA: Disable USM host memory for infer request`s tensors for PVC and subsequent dGPUs, as kernel access
+        // to system memory is slower than using an explicit memcpy (Host <-> Device) call with the copy engine
+        // Driver tickets with additional details: 6155, 10054
         GPU_DEBUG_TRACE << "Do not use usm_host for performance issue" << std::endl;
         can_use_usm = false;
     }