Merge pull request oneapi-src#1588 from steffenlarsen/steffen/fix_tim…

…estamp_merge_mistake [CUDA] Add back device info enums after merge mistake
omarahmed1111 · May 16, 2024 · aac2eeb · aac2eeb
2 parents 7ce68e0 + cb3d945
commit aac2eeb
Show file tree

Hide file tree

Showing 8 changed files with 81 additions and 10 deletions.
diff --git a/source/adapters/cuda/device.cpp b/source/adapters/cuda/device.cpp
@@ -926,6 +926,30 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     // CUDA supports cubemap seamless filtering.
     return ReturnValue(true);
   }
+  case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_1D_USM_EXP: {
+    // CUDA does support fetching 1D USM sampled image data.
+    return ReturnValue(true);
+  }
+  case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_1D_EXP: {
+    // CUDA does not support fetching 1D non-USM sampled image data.
+    return ReturnValue(false);
+  }
+  case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_2D_USM_EXP: {
+    // CUDA does support fetching 2D USM sampled image data.
+    return ReturnValue(true);
+  }
+  case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_2D_EXP: {
+    // CUDA does support fetching 2D non-USM sampled image data.
+    return ReturnValue(true);
+  }
+  case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_3D_USM_EXP: {
+    // CUDA does not support 3D USM sampled textures
+    return ReturnValue(false);
+  }
+  case UR_DEVICE_INFO_BINDLESS_SAMPLED_IMAGE_FETCH_3D_EXP: {
+    // CUDA does support fetching 3D non-USM sampled image data.
+    return ReturnValue(true);
+  }
   case UR_DEVICE_INFO_TIMESTAMP_RECORDING_SUPPORT_EXP: {
     // CUDA supports recording timestamp events.
     return ReturnValue(true);

diff --git a/source/adapters/cuda/event.cpp b/source/adapters/cuda/event.cpp
@@ -176,7 +176,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetInfo(ur_event_handle_t hEvent,
   case UR_EVENT_INFO_CONTEXT:
     return ReturnValue(hEvent->getContext());
   default:
-    detail::ur::die("Event info request not implemented");
+    break;
   }
 
   return UR_RESULT_ERROR_INVALID_ENUMERATION;
@@ -207,8 +207,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEventGetProfilingInfo(
   default:
     break;
   }
-  detail::ur::die("Event Profiling info request not implemented");
-  return {};
+  return UR_RESULT_ERROR_INVALID_ENUMERATION;
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urEventSetCallback(ur_event_handle_t,

diff --git a/test/conformance/event/event_adapter_cuda.match b/test/conformance/event/event_adapter_cuda.match
@@ -1,3 +1,4 @@
+urEventGetProfilingInfoTest.Success/NVIDIA_CUDA_BACKEND___{{.*}}_
 urEventSetCallbackTest.Success/NVIDIA_CUDA_BACKEND___{{.*}}_
 urEventSetCallbackTest.ValidateParameters/NVIDIA_CUDA_BACKEND___{{.*}}_
 urEventSetCallbackTest.AllStates/NVIDIA_CUDA_BACKEND___{{.*}}_

diff --git a/test/conformance/event/event_adapter_hip.match b/test/conformance/event/event_adapter_hip.match
@@ -1,3 +1,4 @@
+urEventGetProfilingInfoTest.Success/AMD_HIP_BACKEND___{{.*}}_
 urEventSetCallbackTest.Success/AMD_HIP_BACKEND___{{.*}}_
 urEventSetCallbackTest.ValidateParameters/AMD_HIP_BACKEND___{{.*}}_
 urEventSetCallbackTest.AllStates/AMD_HIP_BACKEND___{{.*}}_

diff --git a/test/conformance/event/event_adapter_level_zero.match b/test/conformance/event/event_adapter_level_zero.match
@@ -1,4 +1,3 @@
 {{OPT}}urEventGetInfoTest.Success/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}___UR_EVENT_INFO_COMMAND_TYPE
-{{OPT}}urEventGetProfilingInfoTest.Success/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}___UR_PROFILING_INFO_COMMAND_QUEUED
-{{OPT}}urEventGetProfilingInfoTest.Success/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}___UR_PROFILING_INFO_COMMAND_SUBMIT
+{{OPT}}urEventGetProfilingInfoTest.Success/Intel_R__oneAPI_Unified_Runtime_over_Level_Zero___{{.*}}_
 {{OPT}}{{Segmentation fault|Aborted}}
diff --git a/test/conformance/event/event_adapter_native_cpu.match b/test/conformance/event/event_adapter_native_cpu.match
@@ -9,10 +9,7 @@ urEventGetInfoNegativeTest.InvalidSizePropSize/SYCL_NATIVE_CPU___SYCL_Native_CPU
 urEventGetInfoNegativeTest.InvalidSizePropSizeSmall/SYCL_NATIVE_CPU___SYCL_Native_CPU_
 urEventGetInfoNegativeTest.InvalidNullPointerPropValue/SYCL_NATIVE_CPU___SYCL_Native_CPU_
 urEventGetInfoNegativeTest.InvalidNullPointerPropSizeRet/SYCL_NATIVE_CPU___SYCL_Native_CPU_
-urEventGetProfilingInfoTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU___UR_PROFILING_INFO_COMMAND_QUEUED
-urEventGetProfilingInfoTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU___UR_PROFILING_INFO_COMMAND_SUBMIT
-urEventGetProfilingInfoTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU___UR_PROFILING_INFO_COMMAND_START
-urEventGetProfilingInfoTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU___UR_PROFILING_INFO_COMMAND_END
+urEventGetProfilingInfoTest.Success/SYCL_NATIVE_CPU___SYCL_Native_CPU_
 urEventGetProfilingInfoNegativeTest.InvalidNullHandle/SYCL_NATIVE_CPU___SYCL_Native_CPU_
 urEventGetProfilingInfoNegativeTest.InvalidEnumeration/SYCL_NATIVE_CPU___SYCL_Native_CPU_
 urEventGetProfilingInfoNegativeTest.InvalidValue/SYCL_NATIVE_CPU___SYCL_Native_CPU_

diff --git a/test/conformance/event/event_adapter_opencl.match b/test/conformance/event/event_adapter_opencl.match
@@ -0,0 +1 @@
+ urEventGetProfilingInfoTest.Success/Intel_R__OpenCL___{{.*}}
diff --git a/test/conformance/event/urEventGetProfilingInfo.cpp b/test/conformance/event/urEventGetProfilingInfo.cpp
@@ -30,9 +30,58 @@ UUR_TEST_SUITE_P(urEventGetProfilingInfoTest,
                  ::testing::Values(UR_PROFILING_INFO_COMMAND_QUEUED,
                                    UR_PROFILING_INFO_COMMAND_SUBMIT,
                                    UR_PROFILING_INFO_COMMAND_START,
-                                   UR_PROFILING_INFO_COMMAND_END),
+                                   UR_PROFILING_INFO_COMMAND_END,
+                                   UR_PROFILING_INFO_COMMAND_COMPLETE),
                  uur::deviceTestWithParamPrinter<ur_profiling_info_t>);
 
+using urEventGetProfilingInfoWithTimingComparisonTest = uur::event::urEventTest;
+
+TEST_P(urEventGetProfilingInfoWithTimingComparisonTest, Success) {
+    uint8_t size = 8;
+
+    std::vector<uint8_t> queued_data(size);
+    ASSERT_SUCCESS(urEventGetProfilingInfo(event,
+                                           UR_PROFILING_INFO_COMMAND_QUEUED,
+                                           size, queued_data.data(), nullptr));
+    auto queued_timing = reinterpret_cast<size_t *>(queued_data.data());
+    ASSERT_NE(*queued_timing, 0);
+
+    std::vector<uint8_t> submit_data(size);
+    ASSERT_SUCCESS(urEventGetProfilingInfo(event,
+                                           UR_PROFILING_INFO_COMMAND_SUBMIT,
+                                           size, submit_data.data(), nullptr));
+    auto submit_timing = reinterpret_cast<size_t *>(submit_data.data());
+    ASSERT_NE(*submit_timing, 0);
+
+    std::vector<uint8_t> start_data(size);
+    ASSERT_SUCCESS(urEventGetProfilingInfo(event,
+                                           UR_PROFILING_INFO_COMMAND_START,
+                                           size, start_data.data(), nullptr));
+    auto start_timing = reinterpret_cast<size_t *>(start_data.data());
+    ASSERT_NE(*start_timing, 0);
+
+    std::vector<uint8_t> end_data(size);
+    ASSERT_SUCCESS(urEventGetProfilingInfo(event, UR_PROFILING_INFO_COMMAND_END,
+                                           size, end_data.data(), nullptr));
+    auto end_timing = reinterpret_cast<size_t *>(end_data.data());
+    ASSERT_NE(*end_timing, 0);
+
+    std::vector<uint8_t> complete_data(size);
+    ASSERT_SUCCESS(
+        urEventGetProfilingInfo(event, UR_PROFILING_INFO_COMMAND_COMPLETE, size,
+                                complete_data.data(), nullptr));
+    auto complete_timing = reinterpret_cast<size_t *>(complete_data.data());
+    ASSERT_NE(*complete_timing, 0);
+
+    ASSERT_LE(*queued_timing, *submit_timing);
+    ASSERT_LT(*submit_timing, *start_timing);
+    ASSERT_LT(*start_timing, *end_timing);
+    ASSERT_LE(*end_timing, *complete_timing);
+}
+
+UUR_INSTANTIATE_DEVICE_TEST_SUITE_P(
+    urEventGetProfilingInfoWithTimingComparisonTest);
+
 using urEventGetProfilingInfoNegativeTest = uur::event::urEventTest;
 
 TEST_P(urEventGetProfilingInfoNegativeTest, InvalidNullHandle) {