[Cuda] Implement device info ext properties for HW_THREADS_PER_EU and EU_SIMD_WIDTH on Cuda

Author: Georgi Mirazchiyski

source/adapters/cuda/device.cpp

@@ -1082,13 +1082,23 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
   case UR_DEVICE_INFO_COMPOSITE_DEVICE:
   case UR_DEVICE_INFO_MAX_READ_WRITE_IMAGE_ARGS:
   case UR_DEVICE_INFO_GPU_EU_COUNT:
-  case UR_DEVICE_INFO_GPU_EU_SIMD_WIDTH:
   case UR_DEVICE_INFO_GPU_EU_SLICES:
   case UR_DEVICE_INFO_GPU_SUBSLICES_PER_SLICE:
   case UR_DEVICE_INFO_GPU_EU_COUNT_PER_SUBSLICE:
-  case UR_DEVICE_INFO_GPU_HW_THREADS_PER_EU:
     return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
 
+  case UR_DEVICE_INFO_GPU_HW_THREADS_PER_EU: {
+    return ReturnValue(hDevice->getMaxHwThreadsPerCU());
+  }
+  case UR_DEVICE_INFO_GPU_EU_SIMD_WIDTH: {
+    int WarpSize{0};
+    UR_CHECK_ERROR(cuDeviceGetAttribute(
+        &WarpSize, CU_DEVICE_ATTRIBUTE_WARP_SIZE, hDevice->get()));
+    detail::ur::assertion(WarpSize > 0);
+    uint32_t WarpCount = hDevice->getMaxHwThreadsPerCU() / WarpSize;
+    return ReturnValue(WarpCount);
+  }
+
   case UR_DEVICE_INFO_COMMAND_BUFFER_SUPPORT_EXP:
   case UR_DEVICE_INFO_COMMAND_BUFFER_EVENT_SUPPORT_EXP:
     return ReturnValue(true);

source/adapters/cuda/device.hpp

@@ -33,6 +33,7 @@ struct ur_device_handle_t_ {
   int MaxChosenLocalMem{0};
   bool MaxLocalMemSizeChosen{false};
   uint32_t NumComputeUnits{0};
+  uint32_t MaxHwThreadsPerCU{0};
 
 public:
   ur_device_handle_t_(native_type cuDevice, CUcontext cuContext, CUevent evBase,
@@ -58,6 +59,9 @@ struct ur_device_handle_t_ {
     UR_CHECK_ERROR(cuDeviceGetAttribute(
         reinterpret_cast<int *>(&NumComputeUnits),
         CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, cuDevice));
+    UR_CHECK_ERROR(cuDeviceGetAttribute(
+        reinterpret_cast<int *>(&MaxHwThreadsPerCU),
+        CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, cuDevice));
 
     // Set local mem max size if env var is present
     static const char *LocalMemSizePtrUR =
@@ -114,6 +118,8 @@ struct ur_device_handle_t_ {
   bool maxLocalMemSizeChosen() { return MaxLocalMemSizeChosen; };
 
   uint32_t getNumComputeUnits() const noexcept { return NumComputeUnits; };
+
+  uint32_t getMaxHwThreadsPerCU() const noexcept { return MaxHwThreadsPerCU; };
 };
 
 int getAttribute(ur_device_handle_t Device, CUdevice_attribute Attribute);