Merge pull request #2095 from Bensuo/ewan/update_opencl_headers

Reflect upstream OpenCL command-buffer API changes

Author: aarongreig

source/adapters/opencl/CMakeLists.txt

@@ -54,7 +54,7 @@ if(UR_OPENCL_INCLUDE_DIR)
 else()
     FetchContent_Declare(OpenCL-Headers
         GIT_REPOSITORY  "https://github.com/KhronosGroup/OpenCL-Headers.git"
-        GIT_TAG         1e193332d02e27e15812d24ff2a3a7a908eb92a3
+        GIT_TAG         542d7a8f65ecfd88b38de35d8b10aa67b36b33b2
     )
     FetchContent_MakeAvailable(OpenCL-Headers)
     FetchContent_GetProperties(OpenCL-Headers

source/adapters/opencl/command_buffer.cpp

@@ -156,15 +156,15 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendKernelLaunchExp(
   cl_mutable_command_khr *OutCommandHandle =
       hCommandBuffer->IsUpdatable ? &CommandHandle : nullptr;
 
-  cl_ndrange_kernel_command_properties_khr UpdateProperties[] = {
+  cl_command_properties_khr UpdateProperties[] = {
       CL_MUTABLE_DISPATCH_UPDATABLE_FIELDS_KHR,
       CL_MUTABLE_DISPATCH_GLOBAL_OFFSET_KHR |
           CL_MUTABLE_DISPATCH_GLOBAL_SIZE_KHR |
           CL_MUTABLE_DISPATCH_LOCAL_SIZE_KHR |
           CL_MUTABLE_DISPATCH_ARGUMENTS_KHR | CL_MUTABLE_DISPATCH_EXEC_INFO_KHR,
       0};
 
-  cl_ndrange_kernel_command_properties_khr *Properties =
+  cl_command_properties_khr *Properties =
       hCommandBuffer->IsUpdatable ? UpdateProperties : nullptr;
   CL_RETURN_ON_FAILURE(clCommandNDRangeKernelKHR(
       hCommandBuffer->CLCommandBuffer, nullptr, Properties,
@@ -176,8 +176,11 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendKernelLaunchExp(
     auto URCommandHandle =
         std::make_unique<ur_exp_command_buffer_command_handle_t_>(
             hCommandBuffer, CommandHandle, workDim, pLocalWorkSize != nullptr);
-    *phCommandHandle = URCommandHandle.release();
-    hCommandBuffer->CommandHandles.push_back(*phCommandHandle);
+    ur_exp_command_buffer_command_handle_t Handle = URCommandHandle.release();
+    hCommandBuffer->CommandHandles.push_back(Handle);
+    if (phCommandHandle) {
+      *phCommandHandle = Handle;
+    }
   } catch (...) {
     return UR_RESULT_ERROR_OUT_OF_RESOURCES;
   }
@@ -222,7 +225,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyExp(
           cl_ext::CommandCopyBufferName, &clCommandCopyBufferKHR));
 
   CL_RETURN_ON_FAILURE(clCommandCopyBufferKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr,
+      hCommandBuffer->CLCommandBuffer, nullptr, nullptr,
       cl_adapter::cast<cl_mem>(hSrcMem), cl_adapter::cast<cl_mem>(hDstMem),
       srcOffset, dstOffset, size, numSyncPointsInWaitList, pSyncPointWaitList,
       pSyncPoint, nullptr));
@@ -256,7 +259,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyRectExp(
           cl_ext::CommandCopyBufferRectName, &clCommandCopyBufferRectKHR));
 
   CL_RETURN_ON_FAILURE(clCommandCopyBufferRectKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr,
+      hCommandBuffer->CLCommandBuffer, nullptr, nullptr,
       cl_adapter::cast<cl_mem>(hSrcMem), cl_adapter::cast<cl_mem>(hDstMem),
       OpenCLOriginRect, OpenCLDstRect, OpenCLRegion, srcRowPitch, srcSlicePitch,
       dstRowPitch, dstSlicePitch, numSyncPointsInWaitList, pSyncPointWaitList,
@@ -274,9 +277,6 @@ ur_result_t UR_APICALL urCommandBufferAppendMemBufferWriteExp(
     [[maybe_unused]] const ur_exp_command_buffer_sync_point_t
         *pSyncPointWaitList,
     [[maybe_unused]] ur_exp_command_buffer_sync_point_t *pSyncPoint) {
-
-  cl_adapter::die("Experimental Command-buffer feature is not "
-                  "implemented for OpenCL adapter.");
   return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
 }
 
@@ -343,7 +343,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferFillExp(
           cl_ext::CommandFillBufferName, &clCommandFillBufferKHR));
 
   CL_RETURN_ON_FAILURE(clCommandFillBufferKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr,
+      hCommandBuffer->CLCommandBuffer, nullptr, nullptr,
       cl_adapter::cast<cl_mem>(hBuffer), pPattern, patternSize, offset, size,
       numSyncPointsInWaitList, pSyncPointWaitList, pSyncPoint, nullptr));
 
@@ -556,8 +556,6 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
   cl_mutable_command_khr command =
       cl_adapter::cast<cl_mutable_command_khr>(hCommand->CLMutableCommand);
   cl_mutable_dispatch_config_khr dispatch_config = {
-      CL_STRUCTURE_TYPE_MUTABLE_DISPATCH_CONFIG_KHR,
-      nullptr,
       command,
       static_cast<cl_uint>(CLArgs.size()),    // num_args
       static_cast<cl_uint>(CLUSMArgs.size()), // num_svm_args
@@ -570,10 +568,12 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
       CLGlobalWorkSize.data(),                // global_work_size
       CLLocalWorkSize.data(),                 // local_work_size
   };
-  cl_mutable_base_config_khr config = {
-      CL_STRUCTURE_TYPE_MUTABLE_BASE_CONFIG_KHR, nullptr, 1, &dispatch_config};
-  CL_RETURN_ON_FAILURE(
-      clUpdateMutableCommandsKHR(hCommandBuffer->CLCommandBuffer, &config));
+  cl_uint num_configs = 1;
+  cl_command_buffer_update_type_khr config_types[1] = {
+      CL_STRUCTURE_TYPE_MUTABLE_DISPATCH_CONFIG_KHR};
+  const void *configs[1] = {&dispatch_config};
+  CL_RETURN_ON_FAILURE(clUpdateMutableCommandsKHR(
+      hCommandBuffer->CLCommandBuffer, num_configs, config_types, configs));
 
   return UR_RESULT_SUCCESS;
 }

source/adapters/opencl/common.hpp

@@ -272,33 +272,36 @@ cl_int(CL_API_CALL *)(cl_command_buffer_khr command_buffer);
 
 using clCommandNDRangeKernelKHR_fn = CL_API_ENTRY cl_int(CL_API_CALL *)(
     cl_command_buffer_khr command_buffer, cl_command_queue command_queue,
-    const cl_ndrange_kernel_command_properties_khr *properties,
-    cl_kernel kernel, cl_uint work_dim, const size_t *global_work_offset,
+    const cl_command_properties_khr *properties, cl_kernel kernel,
+    cl_uint work_dim, const size_t *global_work_offset,
     const size_t *global_work_size, const size_t *local_work_size,
     cl_uint num_sync_points_in_wait_list,
     const cl_sync_point_khr *sync_point_wait_list,
     cl_sync_point_khr *sync_point, cl_mutable_command_khr *mutable_handle);
 
 using clCommandCopyBufferKHR_fn = CL_API_ENTRY cl_int(CL_API_CALL *)(
     cl_command_buffer_khr command_buffer, cl_command_queue command_queue,
-    cl_mem src_buffer, cl_mem dst_buffer, size_t src_offset, size_t dst_offset,
-    size_t size, cl_uint num_sync_points_in_wait_list,
+    const cl_command_properties_khr *properties, cl_mem src_buffer,
+    cl_mem dst_buffer, size_t src_offset, size_t dst_offset, size_t size,
+    cl_uint num_sync_points_in_wait_list,
     const cl_sync_point_khr *sync_point_wait_list,
     cl_sync_point_khr *sync_point, cl_mutable_command_khr *mutable_handle);
 
 using clCommandCopyBufferRectKHR_fn = CL_API_ENTRY cl_int(CL_API_CALL *)(
     cl_command_buffer_khr command_buffer, cl_command_queue command_queue,
-    cl_mem src_buffer, cl_mem dst_buffer, const size_t *src_origin,
-    const size_t *dst_origin, const size_t *region, size_t src_row_pitch,
-    size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch,
+    const cl_command_properties_khr *properties, cl_mem src_buffer,
+    cl_mem dst_buffer, const size_t *src_origin, const size_t *dst_origin,
+    const size_t *region, size_t src_row_pitch, size_t src_slice_pitch,
+    size_t dst_row_pitch, size_t dst_slice_pitch,
     cl_uint num_sync_points_in_wait_list,
     const cl_sync_point_khr *sync_point_wait_list,
     cl_sync_point_khr *sync_point, cl_mutable_command_khr *mutable_handle);
 
 using clCommandFillBufferKHR_fn = CL_API_ENTRY cl_int(CL_API_CALL *)(
     cl_command_buffer_khr command_buffer, cl_command_queue command_queue,
-    cl_mem buffer, const void *pattern, size_t pattern_size, size_t offset,
-    size_t size, cl_uint num_sync_points_in_wait_list,
+    const cl_command_properties_khr *properties, cl_mem buffer,
+    const void *pattern, size_t pattern_size, size_t offset, size_t size,
+    cl_uint num_sync_points_in_wait_list,
     const cl_sync_point_khr *sync_point_wait_list,
     cl_sync_point_khr *sync_point, cl_mutable_command_khr *mutable_handle);
 
@@ -313,8 +316,9 @@ using clGetCommandBufferInfoKHR_fn = CL_API_ENTRY cl_int(CL_API_CALL *)(
     size_t param_value_size, void *param_value, size_t *param_value_size_ret);
 
 using clUpdateMutableCommandsKHR_fn = CL_API_ENTRY
-cl_int(CL_API_CALL *)(cl_command_buffer_khr command_buffer,
-                      const cl_mutable_base_config_khr *mutable_config);
+cl_int(CL_API_CALL *)(cl_command_buffer_khr command_buffer, cl_uint num_configs,
+                      const cl_command_buffer_update_type_khr *config_types,
+                      const void **configs);
 
 template <typename T> struct FuncPtrCache {
   std::map<cl_context, T> Map;