Merge pull request #2258 from aarongreig/aaron/tryUseExtensionSubgroupInfo

Use extension version of clGetKernelSubGroupInfo when necessary.

Author: callumfare

source/adapters/opencl/common.hpp

@@ -218,6 +218,8 @@ CONSTFIX char CommandFillBufferName[] = "clCommandFillBufferKHR";
 CONSTFIX char EnqueueCommandBufferName[] = "clEnqueueCommandBufferKHR";
 CONSTFIX char GetCommandBufferInfoName[] = "clGetCommandBufferInfoKHR";
 CONSTFIX char UpdateMutableCommandsName[] = "clUpdateMutableCommandsKHR";
+CONSTFIX char CreateProgramWithILName[] = "clCreateProgramWithILKHR";
+CONSTFIX char GetKernelSubGroupInfoName[] = "clGetKernelSubGroupInfoKHR";
 
 #undef CONSTFIX
 
@@ -316,6 +318,13 @@ 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);
 
+using clCreateProgramWithILKHR_fn = CL_API_ENTRY
+cl_program(CL_API_CALL *)(cl_context, const void *, size_t, cl_int *);
+
+using clGetKernelSubGroupInfoKHR_fn = CL_API_ENTRY
+cl_int(CL_API_CALL *)(cl_kernel, cl_device_id, cl_kernel_sub_group_info, size_t,
+                      const void *, size_t, void *, size_t *);
+
 template <typename T> struct FuncPtrCache {
   std::map<cl_context, T> Map;
   std::mutex Mutex;

source/adapters/opencl/device.cpp

@@ -61,7 +61,8 @@ ur_result_t cl_adapter::checkDeviceExtensions(
       // doesn't report them.
       if (isIntelFPGAEmuDevice(Dev) &&
           (Ext == "cl_intel_device_attribute_query" ||
-           Ext == "cl_intel_required_subgroup_size")) {
+           Ext == "cl_intel_required_subgroup_size" ||
+           Ext == "cl_khr_subgroups")) {
         Supported = true;
         continue;
       }

source/adapters/opencl/extension_functions.def

@@ -24,3 +24,5 @@ CL_EXTENSION_FUNC(clCommandFillBufferKHR)
 CL_EXTENSION_FUNC(clEnqueueCommandBufferKHR)
 CL_EXTENSION_FUNC(clGetCommandBufferInfoKHR)
 CL_EXTENSION_FUNC(clUpdateMutableCommandsKHR)
+CL_EXTENSION_FUNC(clCreateProgramWithILKHR)
+CL_EXTENSION_FUNC(clGetKernelSubGroupInfoKHR)

source/adapters/opencl/kernel.cpp

@@ -8,6 +8,7 @@
 //
 //===----------------------------------------------------------------------===//
 #include "common.hpp"
+#include "device.hpp"
 
 #include <algorithm>
 #include <cstddef>
@@ -189,11 +190,39 @@ urKernelGetSubGroupInfo(ur_kernel_handle_t hKernel, ur_device_handle_t hDevice,
     InputValueSize = MaxDims * sizeof(size_t);
   }
 
-  cl_int Ret = clGetKernelSubGroupInfo(cl_adapter::cast<cl_kernel>(hKernel),
-                                       cl_adapter::cast<cl_device_id>(hDevice),
-                                       mapURKernelSubGroupInfoToCL(propName),
-                                       InputValueSize, InputValue.get(),
-                                       sizeof(size_t), &RetVal, pPropSizeRet);
+  // We need to allow for the possibility that this device runs an older CL and
+  // supports the original khr subgroup extension.
+  cl_ext::clGetKernelSubGroupInfoKHR_fn GetKernelSubGroupInfo = nullptr;
+
+  oclv::OpenCLVersion DevVer;
+  CL_RETURN_ON_FAILURE(cl_adapter::getDeviceVersion(
+      cl_adapter::cast<cl_device_id>(hDevice), DevVer));
+
+  if (DevVer < oclv::V2_1) {
+    bool SubgroupExtSupported = false;
+
+    UR_RETURN_ON_FAILURE(cl_adapter::checkDeviceExtensions(
+        cl_adapter::cast<cl_device_id>(hDevice), {"cl_khr_subgroups"},
+        SubgroupExtSupported));
+    if (!SubgroupExtSupported) {
+      return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+    }
+    cl_context Context = nullptr;
+    CL_RETURN_ON_FAILURE(clGetKernelInfo(cl_adapter::cast<cl_kernel>(hKernel),
+                                         CL_KERNEL_CONTEXT, sizeof(Context),
+                                         &Context, nullptr));
+    UR_RETURN_ON_FAILURE(cl_ext::getExtFuncFromContext(
+        Context, cl_ext::ExtFuncPtrCache->clGetKernelSubGroupInfoKHRCache,
+        cl_ext::GetKernelSubGroupInfoName, &GetKernelSubGroupInfo));
+  } else {
+    GetKernelSubGroupInfo = clGetKernelSubGroupInfo;
+  }
+
+  cl_int Ret = GetKernelSubGroupInfo(cl_adapter::cast<cl_kernel>(hKernel),
+                                     cl_adapter::cast<cl_device_id>(hDevice),
+                                     mapURKernelSubGroupInfoToCL(propName),
+                                     InputValueSize, InputValue.get(),
+                                     sizeof(size_t), &RetVal, pPropSizeRet);
 
   if (Ret == CL_INVALID_OPERATION) {
     // clGetKernelSubGroupInfo returns CL_INVALID_OPERATION if the device does

source/adapters/opencl/program.cpp

@@ -99,16 +99,15 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramCreateWithIL(
       }
     }
 
-    using ApiFuncT =
-        cl_program(CL_API_CALL *)(cl_context, const void *, size_t, cl_int *);
-    ApiFuncT FuncPtr =
-        reinterpret_cast<ApiFuncT>(clGetExtensionFunctionAddressForPlatform(
-            CurPlatform, "clCreateProgramWithILKHR"));
+    cl_ext::clCreateProgramWithILKHR_fn CreateProgramWithIL = nullptr;
 
-    assert(FuncPtr != nullptr);
+    UR_RETURN_ON_FAILURE(cl_ext::getExtFuncFromContext(
+        cl_adapter::cast<cl_context>(hContext),
+        cl_ext::ExtFuncPtrCache->clCreateProgramWithILKHRCache,
+        cl_ext::CreateProgramWithILName, &CreateProgramWithIL));
 
-    *phProgram = cl_adapter::cast<ur_program_handle_t>(
-        FuncPtr(cl_adapter::cast<cl_context>(hContext), pIL, length, &Err));
+    *phProgram = cl_adapter::cast<ur_program_handle_t>(CreateProgramWithIL(
+        cl_adapter::cast<cl_context>(hContext), pIL, length, &Err));
   }
 
   // INVALID_VALUE is only returned in three circumstances according to the cl