Merge branch 'main' into buffer-copy-host-ptr

Author: zhaomaosu

.github/workflows/benchmarks_compute.yml

@@ -12,6 +12,7 @@ on:
         default: 'level_zero'
         options:
           - level_zero
+          - level_zero_v2
       unit:
         description: Test unit (cpu/gpu)
         type: choice

include/ur_api.h

@@ -2062,15 +2062,15 @@ typedef struct ur_device_native_properties_t {
 ///     - ::UR_RESULT_ERROR_DEVICE_LOST
 ///     - ::UR_RESULT_ERROR_ADAPTER_SPECIFIC
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_HANDLE
-///         + `NULL == hPlatform`
+///         + `NULL == hAdapter`
 ///     - ::UR_RESULT_ERROR_INVALID_NULL_POINTER
 ///         + `NULL == phDevice`
 ///     - ::UR_RESULT_ERROR_UNSUPPORTED_FEATURE
 ///         + If the adapter has no underlying equivalent handle.
 UR_APIEXPORT ur_result_t UR_APICALL
 urDeviceCreateWithNativeHandle(
     ur_native_handle_t hNativeDevice,                 ///< [in][nocheck] the native handle of the device.
-    ur_platform_handle_t hPlatform,                   ///< [in] handle of the platform instance
+    ur_adapter_handle_t hAdapter,                     ///< [in] handle of the adapter to which `hNativeDevice` belongs
     const ur_device_native_properties_t *pProperties, ///< [in][optional] pointer to native device properties struct.
     ur_device_handle_t *phDevice                      ///< [out] pointer to the handle of the device object created.
 );
@@ -4785,6 +4785,7 @@ urKernelSetArgValue(
     size_t argSize,                                      ///< [in] size of argument type
     const ur_kernel_arg_value_properties_t *pProperties, ///< [in][optional] pointer to value properties.
     const void *pArgValue                                ///< [in] argument value represented as matching arg type.
+                                                         ///< The data pointed to will be copied and therefore can be reused on return.
 );
 
 ///////////////////////////////////////////////////////////////////////////////
@@ -11972,7 +11973,7 @@ typedef struct ur_device_get_native_handle_params_t {
 ///     allowing the callback the ability to modify the parameter's value
 typedef struct ur_device_create_with_native_handle_params_t {
     ur_native_handle_t *phNativeDevice;
-    ur_platform_handle_t *phPlatform;
+    ur_adapter_handle_t *phAdapter;
     const ur_device_native_properties_t **ppProperties;
     ur_device_handle_t **pphDevice;
 } ur_device_create_with_native_handle_params_t;

include/ur_ddi.h

@@ -2373,7 +2373,7 @@ typedef ur_result_t(UR_APICALL *ur_pfnDeviceGetNativeHandle_t)(
 /// @brief Function-pointer for urDeviceCreateWithNativeHandle
 typedef ur_result_t(UR_APICALL *ur_pfnDeviceCreateWithNativeHandle_t)(
     ur_native_handle_t,
-    ur_platform_handle_t,
+    ur_adapter_handle_t,
     const ur_device_native_properties_t *,
     ur_device_handle_t *);
 

include/ur_print.hpp

@@ -17357,10 +17357,10 @@ inline std::ostream &operator<<(std::ostream &os, [[maybe_unused]] const struct
                                   *(params->phNativeDevice)));
 
     os << ", ";
-    os << ".hPlatform = ";
+    os << ".hAdapter = ";
 
     ur::details::printPtr(os,
-                          *(params->phPlatform));
+                          *(params->phAdapter));
 
     os << ", ";
     os << ".pProperties = ";

scripts/benchmarks/benches/compute.py

@@ -15,7 +15,7 @@ class ComputeBench:
     def __init__(self, directory):
         self.directory = directory
         self.built = False
-        self.adapter_short_name = {'level_zero' : 'L0'}
+        self.adapter_short_name = {'level_zero' : 'L0', "level_zero_v2" : 'L0_V2'}
         return
 
     def setup(self):

scripts/core/device.yml

@@ -820,9 +820,9 @@ params:
     - type: $x_native_handle_t
       name: hNativeDevice
       desc: "[in][nocheck] the native handle of the device."
-    - type: $x_platform_handle_t
-      name: hPlatform
-      desc: "[in] handle of the platform instance"
+    - type: $x_adapter_handle_t
+      name: hAdapter
+      desc: "[in] handle of the adapter to which `hNativeDevice` belongs"
     - type: const $x_device_native_properties_t*
       name: pProperties
       desc: "[in][optional] pointer to native device properties struct."

scripts/core/kernel.yml

@@ -64,7 +64,9 @@ params:
       desc: "[in][optional] pointer to value properties."
     - type: "const void*"
       name: pArgValue
-      desc: "[in] argument value represented as matching arg type."
+      desc: |
+         [in] argument value represented as matching arg type.
+         The data pointed to will be copied and therefore can be reused on return.
 returns:
     - $X_RESULT_ERROR_INVALID_KERNEL_ARGUMENT_INDEX
     - $X_RESULT_ERROR_INVALID_KERNEL_ARGUMENT_SIZE

source/adapters/cuda/device.cpp

@@ -1185,27 +1185,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetNativeHandle(
 /// \return TBD
 
 UR_APIEXPORT ur_result_t UR_APICALL urDeviceCreateWithNativeHandle(
-    ur_native_handle_t hNativeDevice, ur_platform_handle_t hPlatform,
-    const ur_device_native_properties_t *pProperties,
+    ur_native_handle_t hNativeDevice,
+    [[maybe_unused]] ur_adapter_handle_t hAdapter,
+    [[maybe_unused]] const ur_device_native_properties_t *pProperties,
     ur_device_handle_t *phDevice) {
-  std::ignore = pProperties;
-
   CUdevice CuDevice = static_cast<CUdevice>(hNativeDevice);
 
   auto IsDevice = [=](std::unique_ptr<ur_device_handle_t_> &Dev) {
     return Dev->get() == CuDevice;
   };
 
-  // If a platform is provided just check if the device is in it
-  if (hPlatform) {
-    auto SearchRes = std::find_if(begin(hPlatform->Devices),
-                                  end(hPlatform->Devices), IsDevice);
-    if (SearchRes != end(hPlatform->Devices)) {
-      *phDevice = SearchRes->get();
-      return UR_RESULT_SUCCESS;
-    }
-  }
-
   // Get list of platforms
   uint32_t NumPlatforms = 0;
   ur_adapter_handle_t AdapterHandle = &adapter;

source/adapters/cuda/image.cpp

@@ -455,32 +455,40 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImageAllocateExp(
 
   // Allocate a cuArray
   if (pImageDesc->numMipLevel == 1) {
-    CUarray ImageArray;
+    CUarray ImageArray{};
 
     try {
       UR_CHECK_ERROR(cuArray3DCreate(&ImageArray, &array_desc));
       *phImageMem = (ur_exp_image_mem_native_handle_t)ImageArray;
     } catch (ur_result_t Err) {
-      cuArrayDestroy(ImageArray);
+      if (ImageArray != CUarray{}) {
+        UR_CHECK_ERROR(cuArrayDestroy(ImageArray));
+      }
       return Err;
     } catch (...) {
-      cuArrayDestroy(ImageArray);
+      if (ImageArray != CUarray{}) {
+        UR_CHECK_ERROR(cuArrayDestroy(ImageArray));
+      }
       return UR_RESULT_ERROR_UNKNOWN;
     }
   } else // Allocate a cuMipmappedArray
   {
-    CUmipmappedArray mip_array;
+    CUmipmappedArray mip_array{};
     array_desc.Flags = CUDA_ARRAY3D_SURFACE_LDST;
 
     try {
       UR_CHECK_ERROR(cuMipmappedArrayCreate(&mip_array, &array_desc,
                                             pImageDesc->numMipLevel));
       *phImageMem = (ur_exp_image_mem_native_handle_t)mip_array;
     } catch (ur_result_t Err) {
-      cuMipmappedArrayDestroy(mip_array);
+      if (mip_array) {
+        UR_CHECK_ERROR(cuMipmappedArrayDestroy(mip_array));
+      }
       return Err;
     } catch (...) {
-      cuMipmappedArrayDestroy(mip_array);
+      if (mip_array) {
+        UR_CHECK_ERROR(cuMipmappedArrayDestroy(mip_array));
+      }
       return UR_RESULT_ERROR_UNKNOWN;
     }
   }

source/adapters/cuda/memory.cpp

@@ -439,7 +439,7 @@ ur_result_t allocateMemObjOnDeviceIfNeeded(ur_mem_handle_t Mem,
       UR_CHECK_ERROR(cuMemAlloc(&DevPtr, Buffer.Size));
     }
   } else {
-    CUarray ImageArray;
+    CUarray ImageArray{};
     CUsurfObject Surface;
     try {
       auto &Image = std::get<SurfaceMem>(Mem->Mem);
@@ -465,12 +465,12 @@ ur_result_t allocateMemObjOnDeviceIfNeeded(ur_mem_handle_t Mem,
       UR_CHECK_ERROR(cuSurfObjectCreate(&Surface, &ImageResDesc));
       Image.SurfObjs[DeviceIdx] = Surface;
     } catch (ur_result_t Err) {
-      if (ImageArray) {
+      if (ImageArray != CUarray{}) {
         UR_CHECK_ERROR(cuArrayDestroy(ImageArray));
       }
       return Err;
     } catch (...) {
-      if (ImageArray) {
+      if (ImageArray != CUarray{}) {
         UR_CHECK_ERROR(cuArrayDestroy(ImageArray));
       }
       return UR_RESULT_ERROR_UNKNOWN;