Merge pull request #2219 from kbenzie/benie/cl-adapter-opt-core-func

[CL] Support using old ICD loaders

Author: aarongreig

source/adapters/opencl/adapter.cpp

@@ -8,16 +8,49 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "adapter.hpp"
 #include "common.hpp"
-#include "logger/ur_logger.hpp"
+#include "ur/ur.hpp"
+
+#ifdef _MSC_VER
+#include <Windows.h>
+#else
+#include <dlfcn.h>
+#endif
+
+ur_adapter_handle_t_::ur_adapter_handle_t_() {
+#ifdef _MSC_VER
+  // Loading OpenCL.dll increments the libraries internal reference count.
+  auto handle = LoadLibraryA("OpenCL.dll");
+
+#define CL_CORE_FUNCTION(FUNC)                                                 \
+  FUNC = reinterpret_cast<decltype(::FUNC) *>(GetProcAddress(handle, "FUNC"));
+#include "core_functions.def"
+#undef CL_CORE_FUNCTION
+
+  // So we can safely decrement it here wihtout actually unloading OpenCL.dll.
+  FreeLibrary(handle);
+#else
+  // Loading libOpenCL.so to get the library handle but don't dlclose it as
+  // this causes a segfault when attempting to call any OpenCL entry point.
+  auto handle = dlopen("libOpenCL.so", RTLD_LOCAL);
+
+#define CL_CORE_FUNCTION(FUNC)                                                 \
+  FUNC = reinterpret_cast<decltype(::FUNC) *>(dlsym(handle, #FUNC));
+#include "core_functions.def"
+#undef CL_CORE_FUNCTION
+
+#endif
+}
 
-struct ur_adapter_handle_t_ {
-  std::atomic<uint32_t> RefCount = 0;
-  std::mutex Mutex;
-  logger::Logger &log = logger::get_logger("opencl");
-};
+static ur_adapter_handle_t adapter = nullptr;
 
-static ur_adapter_handle_t_ *adapter = nullptr;
+ur_adapter_handle_t ur::cl::getAdapter() {
+  if (!adapter) {
+    die("OpenCL adapter used before initalization or after destruction");
+  }
+  return adapter;
+}
 
 static void globalAdapterShutdown() {
   if (cl_ext::ExtFuncPtrCache) {

source/adapters/opencl/adapter.hpp

@@ -8,6 +8,26 @@
 //
 //===----------------------------------------------------------------------===//
 
-struct ur_adapter_handle_t_;
+#include "CL/cl.h"
+#include "logger/ur_logger.hpp"
 
-extern ur_adapter_handle_t_ adapter;
+struct ur_adapter_handle_t_ {
+  ur_adapter_handle_t_();
+
+  std::atomic<uint32_t> RefCount = 0;
+  std::mutex Mutex;
+  logger::Logger &log = logger::get_logger("opencl");
+
+  // Function pointers to core OpenCL entry points which may not exist in older
+  // versions of the OpenCL-ICD-Loader are tracked here and initialized by
+  // dynamically loading the symbol by name.
+#define CL_CORE_FUNCTION(FUNC) decltype(::FUNC) *FUNC = nullptr;
+#include "core_functions.def"
+#undef CL_CORE_FUNCTION
+};
+
+namespace ur {
+namespace cl {
+ur_adapter_handle_t getAdapter();
+} // namespace cl
+} // namespace ur

source/adapters/opencl/common.hpp

@@ -239,10 +239,6 @@ cl_int(CL_API_CALL *)(cl_command_queue, cl_program, const char *, cl_bool,
                       size_t, size_t, void *, cl_uint, const cl_event *,
                       cl_event *);
 
-using clSetProgramSpecializationConstant_fn = CL_API_ENTRY
-cl_int(CL_API_CALL *)(cl_program program, cl_uint spec_id, size_t spec_size,
-                      const void *spec_value);
-
 using clEnqueueReadHostPipeINTEL_fn = CL_API_ENTRY
 cl_int(CL_API_CALL *)(cl_command_queue queue, cl_program program,
                       const char *pipe_symbol, cl_bool blocking, void *ptr,

source/adapters/opencl/context.cpp

@@ -9,6 +9,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "context.hpp"
+#include "adapter.hpp"
 
 #include <mutex>
 #include <set>
@@ -169,6 +170,13 @@ UR_APIEXPORT ur_result_t UR_APICALL urContextCreateWithNativeHandle(
 UR_APIEXPORT ur_result_t UR_APICALL urContextSetExtendedDeleter(
     ur_context_handle_t hContext, ur_context_extended_deleter_t pfnDeleter,
     void *pUserData) {
+  if (!ur::cl::getAdapter()->clSetContextDestructorCallback) {
+    ur::cl::getAdapter()->log.warning(
+        "clSetContextDestructorCallback not found, consider upgrading the "
+        "OpenCL-ICD-Loader to the latest version.");
+    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  }
+
   static std::unordered_map<ur_context_handle_t,
                             std::set<ur_context_extended_deleter_t>>
       ContextCallbackMap;
@@ -212,7 +220,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urContextSetExtendedDeleter(
     auto *C = static_cast<ContextCallback *>(pUserData);
     C->execute();
   };
-  CL_RETURN_ON_FAILURE(clSetContextDestructorCallback(
+  CL_RETURN_ON_FAILURE(ur::cl::getAdapter()->clSetContextDestructorCallback(
       cl_adapter::cast<cl_context>(hContext), ClCallback, Callback));
 
   return UR_RESULT_SUCCESS;

source/adapters/opencl/core_functions.def

@@ -0,0 +1,14 @@
+// Introduced in OpenCL 2.0
+// TODO: clCreateCommandQueueWithProperties
+
+// Introduced in OpenCL 2.1
+// TODO: clGetKernelSubGroupInfo
+// TODO: clCreateProgramWithIL
+// TODO: clGetHostTimer
+// TODO: clGetDeviceAndHostTimer
+
+// Introduced in OpenCL 2.2
+CL_CORE_FUNCTION(clSetProgramSpecializationConstant)
+
+// Introduced in OpenCL 3.0
+CL_CORE_FUNCTION(clSetContextDestructorCallback)

source/adapters/opencl/extension_functions.def

@@ -13,7 +13,6 @@ CL_EXTENSION_FUNC(clEnqueueWriteGlobalVariable)
 CL_EXTENSION_FUNC(clEnqueueReadGlobalVariable)
 CL_EXTENSION_FUNC(clEnqueueReadHostPipeINTEL)
 CL_EXTENSION_FUNC(clEnqueueWriteHostPipeINTEL)
-CL_EXTENSION_FUNC(clSetProgramSpecializationConstant)
 CL_EXTENSION_FUNC(clCreateCommandBufferKHR)
 CL_EXTENSION_FUNC(clRetainCommandBufferKHR)
 CL_EXTENSION_FUNC(clReleaseCommandBufferKHR)

source/adapters/opencl/program.cpp

@@ -8,6 +8,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "adapter.hpp"
 #include "common.hpp"
 #include "context.hpp"
 #include "device.hpp"
@@ -392,50 +393,17 @@ UR_APIEXPORT ur_result_t UR_APICALL urProgramSetSpecializationConstants(
                                        sizeof(cl_platform_id), &CurPlatform,
                                        nullptr));
 
-  oclv::OpenCLVersion PlatVer;
-  cl_adapter::getPlatformVersion(CurPlatform, PlatVer);
-
-  bool UseExtensionLookup = false;
-  if (PlatVer < oclv::V2_2) {
-    UseExtensionLookup = true;
-  } else {
-    for (cl_device_id Dev : *DevicesInCtx) {
-      oclv::OpenCLVersion DevVer;
-
-      UR_RETURN_ON_FAILURE(cl_adapter::getDeviceVersion(Dev, DevVer));
-
-      if (DevVer < oclv::V2_2) {
-        UseExtensionLookup = true;
-        break;
-      }
-    }
-  }
-
-  if (UseExtensionLookup == false) {
+  if (ur::cl::getAdapter()->clSetProgramSpecializationConstant) {
     for (uint32_t i = 0; i < count; ++i) {
-      CL_RETURN_ON_FAILURE(clSetProgramSpecializationConstant(
-          CLProg, pSpecConstants[i].id, pSpecConstants[i].size,
-          pSpecConstants[i].pValue));
+      CL_RETURN_ON_FAILURE(
+          ur::cl::getAdapter()->clSetProgramSpecializationConstant(
+              CLProg, pSpecConstants[i].id, pSpecConstants[i].size,
+              pSpecConstants[i].pValue));
     }
   } else {
-    cl_ext::clSetProgramSpecializationConstant_fn
-        SetProgramSpecializationConstant = nullptr;
-    const ur_result_t URResult = cl_ext::getExtFuncFromContext<
-        decltype(SetProgramSpecializationConstant)>(
-        Ctx, cl_ext::ExtFuncPtrCache->clSetProgramSpecializationConstantCache,
-        cl_ext::SetProgramSpecializationConstantName,
-        &SetProgramSpecializationConstant);
-
-    if (URResult != UR_RESULT_SUCCESS) {
-      return URResult;
-    }
-
-    for (uint32_t i = 0; i < count; ++i) {
-      CL_RETURN_ON_FAILURE(SetProgramSpecializationConstant(
-          CLProg, pSpecConstants[i].id, pSpecConstants[i].size,
-          pSpecConstants[i].pValue));
-    }
+    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
   }
+
   return UR_RESULT_SUCCESS;
 }