[UR][L0 v2] check if copy offload is supported before requesting it (#17120)

Author: igchor

unified-runtime/source/adapters/level_zero/platform.cpp

@@ -276,6 +276,15 @@ ur_result_t ur_platform_handle_t_::initialize() {
         ZeDriverEuCountExtensionFound = true;
       }
     }
+    if (strncmp(extension.name,
+                ZEX_INTEL_QUEUE_COPY_OPERATIONS_OFFLOAD_HINT_EXP_NAME,
+                strlen(ZEX_INTEL_QUEUE_COPY_OPERATIONS_OFFLOAD_HINT_EXP_NAME) +
+                    1) == 0) {
+      if (extension.version ==
+          ZEX_INTEL_QUEUE_COPY_OPERATIONS_OFFLOAD_HINT_EXP_VERSION_1_0) {
+        ZeCopyOffloadExtensionSupported = true;
+      }
+    }
     zeDriverExtensionMap[extension.name] = extension.version;
   }
 

unified-runtime/source/adapters/level_zero/platform.hpp

@@ -62,6 +62,7 @@ struct ur_platform_handle_t_ : public _ur_platform {
   bool ZeDriverEventPoolCountingEventsExtensionFound{false};
   bool zeDriverImmediateCommandListAppendFound{false};
   bool ZeDriverEuCountExtensionFound{false};
+  bool ZeCopyOffloadExtensionSupported{false};
 
   // Cache UR devices for reuse
   std::vector<std::unique_ptr<ur_device_handle_t_>> URDevicesCache;

unified-runtime/source/adapters/level_zero/v2/command_list_cache.cpp

@@ -46,15 +46,25 @@ inline size_t command_list_descriptor_hash_t::operator()(
   }
 }
 
-command_list_cache_t::command_list_cache_t(ze_context_handle_t ZeContext)
-    : ZeContext{ZeContext} {}
+command_list_cache_t::command_list_cache_t(ze_context_handle_t ZeContext,
+                                           bool ZeCopyOffloadExtensionSupported)
+    : ZeContext{ZeContext},
+      ZeCopyOffloadExtensionSupported{ZeCopyOffloadExtensionSupported} {}
 
 raii::ze_command_list_handle_t
 command_list_cache_t::createCommandList(const command_list_descriptor_t &desc) {
   ZeStruct<zex_intel_queue_copy_operations_offload_hint_exp_desc_t> offloadDesc;
-  offloadDesc.copyOffloadEnabled =
+  auto requestedCopyOffload =
       std::visit([](auto &&arg) { return arg.CopyOffloadEnabled; }, desc);
 
+  if (!ZeCopyOffloadExtensionSupported && requestedCopyOffload) {
+    logger::info(
+        "Copy offload is requested but is not supported by the driver.");
+    offloadDesc.copyOffloadEnabled = false;
+  } else {
+    offloadDesc.copyOffloadEnabled = requestedCopyOffload;
+  }
+
   if (auto ImmCmdDesc =
           std::get_if<immediate_command_list_descriptor_t>(&desc)) {
     ze_command_list_handle_t ZeCommandList;

unified-runtime/source/adapters/level_zero/v2/command_list_cache.hpp

@@ -54,7 +54,8 @@ struct command_list_descriptor_hash_t {
 };
 
 struct command_list_cache_t {
-  command_list_cache_t(ze_context_handle_t ZeContext);
+  command_list_cache_t(ze_context_handle_t ZeContext,
+                       bool ZeCopyOffloadExtensionSupported);
 
   raii::command_list_unique_handle
   getImmediateCommandList(ze_device_handle_t ZeDevice, bool IsInOrder,
@@ -72,6 +73,7 @@ struct command_list_cache_t {
 
 private:
   ze_context_handle_t ZeContext;
+  bool ZeCopyOffloadExtensionSupported;
   std::unordered_map<command_list_descriptor_t,
                      std::stack<raii::ze_command_list_handle_t>,
                      command_list_descriptor_hash_t>

unified-runtime/source/adapters/level_zero/v2/context.cpp

@@ -49,7 +49,9 @@ ur_context_handle_t_::ur_context_handle_t_(ze_context_handle_t hContext,
                                            const ur_device_handle_t *phDevices,
                                            bool ownZeContext)
     : hContext(hContext, ownZeContext),
-      hDevices(phDevices, phDevices + numDevices), commandListCache(hContext),
+      hDevices(phDevices, phDevices + numDevices),
+      commandListCache(hContext,
+                       phDevices[0]->Platform->ZeCopyOffloadExtensionSupported),
       eventPoolCache(this, phDevices[0]->Platform->getNumDevices(),
                      [context = this, platform = phDevices[0]->Platform](
                          DeviceId deviceId, v2::event_flags_t flags)

unified-runtime/test/adapters/level_zero/v2/command_list_cache_test.cpp

@@ -24,7 +24,7 @@ struct CommandListCacheTest : public uur::urContextTest {};
 UUR_INSTANTIATE_DEVICE_TEST_SUITE(CommandListCacheTest);
 
 TEST_P(CommandListCacheTest, CanStoreAndRetriveImmediateAndRegularCmdLists) {
-  v2::command_list_cache_t cache(context->getZeHandle());
+  v2::command_list_cache_t cache(context->getZeHandle(), false);
 
   bool IsInOrder = false;
   uint32_t Ordinal = 0;
@@ -76,7 +76,7 @@ TEST_P(CommandListCacheTest, CanStoreAndRetriveImmediateAndRegularCmdLists) {
 }
 
 TEST_P(CommandListCacheTest, ImmediateCommandListsHaveProperAttributes) {
-  v2::command_list_cache_t cache(context->getZeHandle());
+  v2::command_list_cache_t cache(context->getZeHandle(), false);
 
   uint32_t numQueueGroups = 0;
   ASSERT_EQ(zeDeviceGetCommandQueueGroupProperties(device->ZeDevice,