[UR] Use reference counting on OpenCL adapters

Author: RossBrunton

unified-runtime/source/adapters/opencl/adapter.cpp

@@ -18,6 +18,11 @@
 #include <dlfcn.h>
 #endif
 
+// There can only be one OpenCL adapter alive at a time.
+// If it is alive (more get/retains than releases called), this is a pointer to
+// it.
+static ur_adapter_handle_t liveAdapter = nullptr;
+
 ur_adapter_handle_t_::ur_adapter_handle_t_() {
 #ifdef _MSC_VER
 
@@ -42,45 +47,38 @@ ur_adapter_handle_t_::ur_adapter_handle_t_() {
 #undef CL_CORE_FUNCTION
 
 #endif // _MSC_VER
+  assert(!liveAdapter);
+  liveAdapter = this;
 }
 
-static ur_adapter_handle_t adapter = nullptr;
+ur_adapter_handle_t_::~ur_adapter_handle_t_() {
+  assert(liveAdapter == this);
+  liveAdapter = nullptr;
+}
 
 ur_adapter_handle_t ur::cl::getAdapter() {
-  if (!adapter) {
+  if (!liveAdapter) {
     die("OpenCL adapter used before initalization or after destruction");
   }
-  return adapter;
-}
-
-static void globalAdapterShutdown() {
-  if (cl_ext::ExtFuncPtrCache) {
-    delete cl_ext::ExtFuncPtrCache;
-    cl_ext::ExtFuncPtrCache = nullptr;
-  }
-  if (adapter) {
-    delete adapter;
-    adapter = nullptr;
-  }
+  return liveAdapter;
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL
 urAdapterGet(uint32_t NumEntries, ur_adapter_handle_t *phAdapters,
              uint32_t *pNumAdapters) {
+  static std::mutex AdapterConstructionMutex{};
+
   if (NumEntries > 0 && phAdapters) {
-    // Sometimes urAdaterGet may be called after the library already been torn
-    // down, we also need to create a temporary handle for it.
-    if (!adapter) {
-      adapter = new ur_adapter_handle_t_();
-      atexit(globalAdapterShutdown);
-    }
+    std::lock_guard<std::mutex> Lock{AdapterConstructionMutex};
 
-    std::lock_guard<std::mutex> Lock{adapter->Mutex};
-    if (adapter->RefCount++ == 0) {
-      cl_ext::ExtFuncPtrCache = new cl_ext::ExtFuncPtrCacheT();
+    if (!liveAdapter) {
+      *phAdapters = new ur_adapter_handle_t_();
+    } else {
+      *phAdapters = liveAdapter;
     }
 
-    *phAdapters = adapter;
+    auto &adapter = *phAdapters;
+    adapter->RefCount++;
   }
 
   if (pNumAdapters) {
@@ -90,21 +88,16 @@ urAdapterGet(uint32_t NumEntries, ur_adapter_handle_t *phAdapters,
   return UR_RESULT_SUCCESS;
 }
 
-UR_APIEXPORT ur_result_t UR_APICALL urAdapterRetain(ur_adapter_handle_t) {
-  ++adapter->RefCount;
+UR_APIEXPORT ur_result_t UR_APICALL
+urAdapterRetain(ur_adapter_handle_t hAdapter) {
+  ++hAdapter->RefCount;
   return UR_RESULT_SUCCESS;
 }
 
-UR_APIEXPORT ur_result_t UR_APICALL urAdapterRelease(ur_adapter_handle_t) {
-  // Check first if the adapter is valid pointer
-  if (adapter) {
-    std::lock_guard<std::mutex> Lock{adapter->Mutex};
-    if (--adapter->RefCount == 0) {
-      if (cl_ext::ExtFuncPtrCache) {
-        delete cl_ext::ExtFuncPtrCache;
-        cl_ext::ExtFuncPtrCache = nullptr;
-      }
-    }
+UR_APIEXPORT ur_result_t UR_APICALL
+urAdapterRelease(ur_adapter_handle_t hAdapter) {
+  if (--hAdapter->RefCount == 0) {
+    delete hAdapter;
   }
   return UR_RESULT_SUCCESS;
 }
@@ -117,18 +110,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urAdapterGetLastError(
   return UR_RESULT_SUCCESS;
 }
 
-UR_APIEXPORT ur_result_t UR_APICALL urAdapterGetInfo(ur_adapter_handle_t,
-                                                     ur_adapter_info_t propName,
-                                                     size_t propSize,
-                                                     void *pPropValue,
-                                                     size_t *pPropSizeRet) {
+UR_APIEXPORT ur_result_t UR_APICALL
+urAdapterGetInfo(ur_adapter_handle_t hAdapter, ur_adapter_info_t propName,
+                 size_t propSize, void *pPropValue, size_t *pPropSizeRet) {
   UrReturnHelper ReturnValue(propSize, pPropValue, pPropSizeRet);
 
   switch (propName) {
   case UR_ADAPTER_INFO_BACKEND:
     return ReturnValue(UR_ADAPTER_BACKEND_OPENCL);
   case UR_ADAPTER_INFO_REFERENCE_COUNT:
-    return ReturnValue(adapter->RefCount.load());
+    return ReturnValue(hAdapter->RefCount.load());
   case UR_ADAPTER_INFO_VERSION:
     return ReturnValue(uint32_t{1});
   default:

unified-runtime/source/adapters/opencl/adapter.hpp

@@ -7,18 +7,22 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+#pragma once
+
 #include "logger/ur_logger.hpp"
 #include "platform.hpp"
 
 #include "CL/cl.h"
+#include "common.hpp"
 #include "logger/ur_logger.hpp"
 
 struct ur_adapter_handle_t_ {
   ur_adapter_handle_t_();
+  ~ur_adapter_handle_t_();
 
   std::atomic<uint32_t> RefCount = 0;
-  std::mutex Mutex;
   logger::Logger &log = logger::get_logger("opencl");
+  cl_ext::ExtFuncPtrCacheT fnCache{};
 
   std::vector<std::unique_ptr<ur_platform_handle_t_>> URPlatforms;
   uint32_t NumPlatforms = 0;
@@ -34,5 +38,8 @@ struct ur_adapter_handle_t_ {
 namespace ur {
 namespace cl {
 ur_adapter_handle_t getAdapter();
+inline cl_ext::ExtFuncPtrCacheT &getExtFnPtrCache() {
+  return getAdapter()->fnCache;
+}
 } // namespace cl
 } // namespace ur

unified-runtime/source/adapters/opencl/command_buffer.cpp

@@ -9,6 +9,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "command_buffer.hpp"
+#include "adapter.hpp"
 #include "common.hpp"
 #include "context.hpp"
 #include "event.hpp"
@@ -25,7 +26,7 @@ ur_exp_command_buffer_handle_t_::~ur_exp_command_buffer_handle_t_() {
   cl_ext::clReleaseCommandBufferKHR_fn clReleaseCommandBufferKHR = nullptr;
   cl_int Res =
       cl_ext::getExtFuncFromContext<decltype(clReleaseCommandBufferKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clReleaseCommandBufferKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clReleaseCommandBufferKHRCache,
           cl_ext::ReleaseCommandBufferName, &clReleaseCommandBufferKHR);
   assert(Res == CL_SUCCESS);
   (void)Res;
@@ -53,7 +54,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferCreateExp(
   cl_ext::clCreateCommandBufferKHR_fn clCreateCommandBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCreateCommandBufferKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clCreateCommandBufferKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clCreateCommandBufferKHRCache,
           cl_ext::CreateCommandBufferName, &clCreateCommandBufferKHR));
 
   const bool IsUpdatable = pCommandBufferDesc->isUpdatable;
@@ -114,7 +115,7 @@ urCommandBufferFinalizeExp(ur_exp_command_buffer_handle_t hCommandBuffer) {
   cl_ext::clFinalizeCommandBufferKHR_fn clFinalizeCommandBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clFinalizeCommandBufferKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clFinalizeCommandBufferKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clFinalizeCommandBufferKHRCache,
           cl_ext::FinalizeCommandBufferName, &clFinalizeCommandBufferKHR));
 
   CL_RETURN_ON_FAILURE(
@@ -146,7 +147,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendKernelLaunchExp(
   cl_ext::clCommandNDRangeKernelKHR_fn clCommandNDRangeKernelKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandNDRangeKernelKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clCommandNDRangeKernelKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clCommandNDRangeKernelKHRCache,
           cl_ext::CommandNRRangeKernelName, &clCommandNDRangeKernelKHR));
 
   cl_mutable_command_khr CommandHandle = nullptr;
@@ -236,7 +237,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyExp(
   cl_ext::clCommandCopyBufferKHR_fn clCommandCopyBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandCopyBufferKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clCommandCopyBufferKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clCommandCopyBufferKHRCache,
           cl_ext::CommandCopyBufferName, &clCommandCopyBufferKHR));
 
   const bool IsInOrder = hCommandBuffer->IsInOrder;
@@ -278,7 +279,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyRectExp(
   cl_ext::clCommandCopyBufferRectKHR_fn clCommandCopyBufferRectKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandCopyBufferRectKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clCommandCopyBufferRectKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clCommandCopyBufferRectKHRCache,
           cl_ext::CommandCopyBufferRectName, &clCommandCopyBufferRectKHR));
 
   const bool IsInOrder = hCommandBuffer->IsInOrder;
@@ -386,7 +387,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferFillExp(
   cl_ext::clCommandFillBufferKHR_fn clCommandFillBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandFillBufferKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clCommandFillBufferKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clCommandFillBufferKHRCache,
           cl_ext::CommandFillBufferName, &clCommandFillBufferKHR));
 
   const bool IsInOrder = hCommandBuffer->IsInOrder;
@@ -457,7 +458,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferEnqueueExp(
   cl_ext::clEnqueueCommandBufferKHR_fn clEnqueueCommandBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clEnqueueCommandBufferKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clEnqueueCommandBufferKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clEnqueueCommandBufferKHRCache,
           cl_ext::EnqueueCommandBufferName, &clEnqueueCommandBufferKHR));
 
   const uint32_t NumberOfQueues = 1;
@@ -615,7 +616,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
   cl_ext::clUpdateMutableCommandsKHR_fn clUpdateMutableCommandsKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clUpdateMutableCommandsKHR)>(
-          CLContext, cl_ext::ExtFuncPtrCache->clUpdateMutableCommandsKHRCache,
+          CLContext, ur::cl::getExtFnPtrCache().clUpdateMutableCommandsKHRCache,
           cl_ext::UpdateMutableCommandsName, &clUpdateMutableCommandsKHR));
 
   if (!hCommandBuffer->IsFinalized || !hCommandBuffer->IsUpdatable)

unified-runtime/source/adapters/opencl/common.hpp

@@ -342,11 +342,6 @@ struct ExtFuncPtrCacheT {
 #undef CL_EXTENSION_FUNC
   }
 };
-// A raw pointer is used here since the lifetime of this map has to be tied to
-// piTeardown to avoid issues with static destruction order (a user application
-// might have static objects that indirectly access this cache in their
-// destructor).
-inline ExtFuncPtrCacheT *ExtFuncPtrCache;
 
 // USM helper function to get an extension function pointer
 template <typename T>

unified-runtime/source/adapters/opencl/context.cpp

@@ -125,20 +125,9 @@ urContextGetInfo(ur_context_handle_t hContext, ur_context_info_t propName,
 
 UR_APIEXPORT ur_result_t UR_APICALL
 urContextRelease(ur_context_handle_t hContext) {
-  // If we're reasonably sure this context is about to be detroyed we should
-  // clear the ext function pointer cache. This isn't foolproof sadly but it
-  // should drastically reduce the chances of the pathological case described
-  // in the comments in common.hpp.
   static std::mutex contextReleaseMutex;
-  auto clContext = hContext->CLContext;
 
   std::lock_guard<std::mutex> lock(contextReleaseMutex);
-  size_t refCount = hContext->getReferenceCount();
-  // ExtFuncPtrCache is destroyed in an atexit() callback, so it doesn't
-  // necessarily outlive the adapter (or all the contexts).
-  if (refCount == 1 && cl_ext::ExtFuncPtrCache) {
-    cl_ext::ExtFuncPtrCache->clearCache(clContext);
-  }
 
   if (hContext->decrementReferenceCount() == 0) {
     delete hContext;

unified-runtime/source/adapters/opencl/context.hpp

@@ -9,6 +9,7 @@
 //===----------------------------------------------------------------------===//
 #pragma once
 
+#include "adapter.hpp"
 #include "common.hpp"
 #include "device.hpp"
 
@@ -29,6 +30,9 @@ struct ur_context_handle_t_ {
       Devices.emplace_back(phDevices[i]);
       urDeviceRetain(phDevices[i]);
     }
+    // The context retains a reference to the adapter so it can clear the
+    // function ptr cache on distruction
+    urAdapterRetain(ur::cl::getAdapter());
     RefCount = 1;
   }
 
@@ -42,6 +46,13 @@ struct ur_context_handle_t_ {
                                     const ur_device_handle_t *phDevices,
                                     ur_context_handle_t &Context);
   ~ur_context_handle_t_() {
+    // If we're reasonably sure this context is about to be destroyed we should
+    // clear the ext function pointer cache. This isn't foolproof sadly but it
+    // should drastically reduce the chances of the pathological case described
+    // in the comments in common.hpp.
+    ur::cl::getExtFnPtrCache().clearCache(CLContext);
+    urAdapterRelease(ur::cl::getAdapter());
+
     for (uint32_t i = 0; i < DeviceCount; i++) {
       urDeviceRelease(Devices[i]);
     }

unified-runtime/source/adapters/opencl/enqueue.cpp

@@ -8,6 +8,7 @@
 //
 //===----------------------------------------------------------------------===//
 
+#include "adapter.hpp"
 #include "common.hpp"
 #include "context.hpp"
 #include "event.hpp"
@@ -410,7 +411,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueDeviceGlobalVariableWrite(
   MapUREventsToCL(numEventsInWaitList, phEventWaitList, CLWaitEvents);
   cl_ext::clEnqueueWriteGlobalVariable_fn F = nullptr;
   UR_RETURN_ON_FAILURE(cl_ext::getExtFuncFromContext<decltype(F)>(
-      Ctx, cl_ext::ExtFuncPtrCache->clEnqueueWriteGlobalVariableCache,
+      Ctx, ur::cl::getExtFnPtrCache().clEnqueueWriteGlobalVariableCache,
       cl_ext::EnqueueWriteGlobalVariableName, &F));
 
   cl_int Res =
@@ -432,7 +433,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueDeviceGlobalVariableRead(
   MapUREventsToCL(numEventsInWaitList, phEventWaitList, CLWaitEvents);
   cl_ext::clEnqueueReadGlobalVariable_fn F = nullptr;
   UR_RETURN_ON_FAILURE(cl_ext::getExtFuncFromContext<decltype(F)>(
-      Ctx, cl_ext::ExtFuncPtrCache->clEnqueueReadGlobalVariableCache,
+      Ctx, ur::cl::getExtFnPtrCache().clEnqueueReadGlobalVariableCache,
       cl_ext::EnqueueReadGlobalVariableName, &F));
 
   cl_int Res =
@@ -456,7 +457,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueReadHostPipe(
   cl_ext::clEnqueueReadHostPipeINTEL_fn FuncPtr = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<cl_ext::clEnqueueReadHostPipeINTEL_fn>(
-          CLContext, cl_ext::ExtFuncPtrCache->clEnqueueReadHostPipeINTELCache,
+          CLContext, ur::cl::getExtFnPtrCache().clEnqueueReadHostPipeINTELCache,
           cl_ext::EnqueueReadHostPipeName, &FuncPtr));
 
   if (FuncPtr) {
@@ -484,7 +485,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueWriteHostPipe(
   cl_ext::clEnqueueWriteHostPipeINTEL_fn FuncPtr = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<cl_ext::clEnqueueWriteHostPipeINTEL_fn>(
-          CLContext, cl_ext::ExtFuncPtrCache->clEnqueueWriteHostPipeINTELCache,
+          CLContext,
+          ur::cl::getExtFnPtrCache().clEnqueueWriteHostPipeINTELCache,
           cl_ext::EnqueueWriteHostPipeName, &FuncPtr));
 
   if (FuncPtr) {