[UR] Add handles to opencl adapter (#17572)

Co-authored-by: omarahmed1111 <[email protected]>

Author: RossBrunton

sycl/source/detail/buffer_impl.cpp

@@ -86,17 +86,17 @@ buffer_impl::getNativeVector(backend BackendName) const {
 
     auto Adapter = Platform->getAdapter();
 
-    if (Platform->getBackend() == backend::opencl) {
-      __SYCL_OCL_CALL(clRetainMemObject, ur::cast<cl_mem>(NativeMem));
-    }
-
     ur_native_handle_t Handle = 0;
     // When doing buffer interop we don't know what device the memory should be
     // resident on, so pass nullptr for Device param. Buffer interop may not be
     // supported by all backends.
     Adapter->call<UrApiKind::urMemGetNativeHandle>(NativeMem, /*Dev*/ nullptr,
                                                    &Handle);
     Handles.push_back(Handle);
+
+    if (Platform->getBackend() == backend::opencl) {
+      __SYCL_OCL_CALL(clRetainMemObject, ur::cast<cl_mem>(Handle));
+    }
   }
 
   addInteropObject(Handles);

sycl/source/kernel.cpp

@@ -22,9 +22,10 @@ kernel::kernel(cl_kernel ClKernel, const context &SyclContext) {
   ur_kernel_handle_t hKernel = nullptr;
   ur_native_handle_t nativeHandle =
       reinterpret_cast<ur_native_handle_t>(ClKernel);
-  Adapter->call<detail::UrApiKind::urKernelCreateWithNativeHandle>(
-      nativeHandle, detail::getSyclObjImpl(SyclContext)->getHandleRef(),
-      nullptr, nullptr, &hKernel);
+  Adapter
+      ->call<errc::invalid, detail::UrApiKind::urKernelCreateWithNativeHandle>(
+          nativeHandle, detail::getSyclObjImpl(SyclContext)->getHandleRef(),
+          nullptr, nullptr, &hKernel);
   impl = std::make_shared<detail::kernel_impl>(
       hKernel, detail::getSyclObjImpl(SyclContext), nullptr, nullptr);
   // This is a special interop constructor for OpenCL, so the kernel must be

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

@@ -7,6 +7,8 @@
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
+#include "logger/ur_logger.hpp"
+#include "platform.hpp"
 
 #include "CL/cl.h"
 #include "logger/ur_logger.hpp"
@@ -18,6 +20,9 @@ struct ur_adapter_handle_t_ {
   std::mutex Mutex;
   logger::Logger &log = logger::get_logger("opencl");
 
+  std::vector<std::unique_ptr<ur_platform_handle_t_>> URPlatforms;
+  uint32_t NumPlatforms = 0;
+
   // 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.

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

@@ -10,13 +10,18 @@
 
 #include "command_buffer.hpp"
 #include "common.hpp"
+#include "context.hpp"
+#include "event.hpp"
+#include "kernel.hpp"
+#include "memory.hpp"
+#include "queue.hpp"
 
 /// The ur_exp_command_buffer_handle_t_ destructor calls CL release
 /// command-buffer to free the underlying object.
 ur_exp_command_buffer_handle_t_::~ur_exp_command_buffer_handle_t_() {
   urQueueRelease(hInternalQueue);
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hContext);
+  cl_context CLContext = hContext->CLContext;
   cl_ext::clReleaseCommandBufferKHR_fn clReleaseCommandBufferKHR = nullptr;
   cl_int Res =
       cl_ext::getExtFuncFromContext<decltype(clReleaseCommandBufferKHR)>(
@@ -43,7 +48,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferCreateExp(
   UR_RETURN_ON_FAILURE(
       urQueueCreate(hContext, hDevice, &QueueProperties, &Queue));
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hContext);
+  cl_context CLContext = hContext->CLContext;
   cl_ext::clCreateCommandBufferKHR_fn clCreateCommandBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCreateCommandBufferKHR)>(
@@ -53,7 +58,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferCreateExp(
   const bool IsUpdatable = pCommandBufferDesc->isUpdatable;
 
   ur_device_command_buffer_update_capability_flags_t UpdateCapabilities;
-  cl_device_id CLDevice = cl_adapter::cast<cl_device_id>(hDevice);
+  cl_device_id CLDevice = hDevice->CLDevice;
   CL_RETURN_ON_FAILURE(
       getDeviceCommandBufferUpdateCapabilities(CLDevice, UpdateCapabilities));
   bool DeviceSupportsUpdate = UpdateCapabilities > 0;
@@ -67,16 +72,19 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferCreateExp(
       IsUpdatable ? CL_COMMAND_BUFFER_MUTABLE_KHR : 0u, 0};
 
   cl_int Res = CL_SUCCESS;
-  auto CLCommandBuffer = clCreateCommandBufferKHR(
-      1, cl_adapter::cast<cl_command_queue *>(&Queue), Properties, &Res);
+  const cl_command_queue CLQueue = Queue->CLQueue;
+  auto CLCommandBuffer =
+      clCreateCommandBufferKHR(1, &CLQueue, Properties, &Res);
   CL_RETURN_ON_FAILURE_AND_SET_NULL(Res, phCommandBuffer);
 
   try {
     auto URCommandBuffer = std::make_unique<ur_exp_command_buffer_handle_t_>(
         Queue, hContext, hDevice, CLCommandBuffer, IsUpdatable, IsInOrder);
     *phCommandBuffer = URCommandBuffer.release();
-  } catch (...) {
+  } catch (std::bad_alloc &) {
     return UR_RESULT_ERROR_OUT_OF_RESOURCES;
+  } catch (...) {
+    return UR_RESULT_ERROR_UNKNOWN;
   }
 
   CL_RETURN_ON_FAILURE(Res);
@@ -101,7 +109,7 @@ urCommandBufferReleaseExp(ur_exp_command_buffer_handle_t hCommandBuffer) {
 UR_APIEXPORT ur_result_t UR_APICALL
 urCommandBufferFinalizeExp(ur_exp_command_buffer_handle_t hCommandBuffer) {
   UR_ASSERT(!hCommandBuffer->IsFinalized, UR_RESULT_ERROR_INVALID_OPERATION);
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clFinalizeCommandBufferKHR_fn clFinalizeCommandBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clFinalizeCommandBufferKHR)>(
@@ -133,7 +141,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendKernelLaunchExp(
   UR_ASSERT(!(phCommandHandle && !hCommandBuffer->IsUpdatable),
             UR_RESULT_ERROR_INVALID_OPERATION);
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clCommandNDRangeKernelKHR_fn clCommandNDRangeKernelKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandNDRangeKernelKHR)>(
@@ -161,10 +169,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendKernelLaunchExp(
       IsInOrder ? nullptr : pSyncPointWaitList;
   uint32_t WaitListSize = IsInOrder ? 0 : numSyncPointsInWaitList;
   CL_RETURN_ON_FAILURE(clCommandNDRangeKernelKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr, Properties,
-      cl_adapter::cast<cl_kernel>(hKernel), workDim, pGlobalWorkOffset,
-      pGlobalWorkSize, pLocalWorkSize, WaitListSize, SyncPointWaitList,
-      RetSyncPoint, OutCommandHandle));
+      hCommandBuffer->CLCommandBuffer, nullptr, Properties, hKernel->CLKernel,
+      workDim, pGlobalWorkOffset, pGlobalWorkSize, pLocalWorkSize, WaitListSize,
+      SyncPointWaitList, RetSyncPoint, OutCommandHandle));
 
   try {
     auto Handle = std::make_unique<ur_exp_command_buffer_command_handle_t_>(
@@ -224,7 +231,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyExp(
   (void)phEventWaitList;
   (void)phEvent;
   (void)phCommand;
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clCommandCopyBufferKHR_fn clCommandCopyBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandCopyBufferKHR)>(
@@ -237,10 +244,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyExp(
       IsInOrder ? nullptr : pSyncPointWaitList;
   uint32_t WaitListSize = IsInOrder ? 0 : numSyncPointsInWaitList;
   CL_RETURN_ON_FAILURE(clCommandCopyBufferKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr, nullptr,
-      cl_adapter::cast<cl_mem>(hSrcMem), cl_adapter::cast<cl_mem>(hDstMem),
-      srcOffset, dstOffset, size, WaitListSize, SyncPointWaitList, RetSyncPoint,
-      nullptr));
+      hCommandBuffer->CLCommandBuffer, nullptr, nullptr, hSrcMem->CLMemory,
+      hDstMem->CLMemory, srcOffset, dstOffset, size, WaitListSize,
+      SyncPointWaitList, RetSyncPoint, nullptr));
 
   return UR_RESULT_SUCCESS;
 }
@@ -267,7 +273,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyRectExp(
   size_t OpenCLDstRect[3]{dstOrigin.x, dstOrigin.y, dstOrigin.z};
   size_t OpenCLRegion[3]{region.width, region.height, region.depth};
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clCommandCopyBufferRectKHR_fn clCommandCopyBufferRectKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandCopyBufferRectKHR)>(
@@ -280,11 +286,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferCopyRectExp(
       IsInOrder ? nullptr : pSyncPointWaitList;
   uint32_t WaitListSize = IsInOrder ? 0 : numSyncPointsInWaitList;
   CL_RETURN_ON_FAILURE(clCommandCopyBufferRectKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr, nullptr,
-      cl_adapter::cast<cl_mem>(hSrcMem), cl_adapter::cast<cl_mem>(hDstMem),
-      OpenCLOriginRect, OpenCLDstRect, OpenCLRegion, srcRowPitch, srcSlicePitch,
-      dstRowPitch, dstSlicePitch, WaitListSize, SyncPointWaitList, RetSyncPoint,
-      nullptr));
+      hCommandBuffer->CLCommandBuffer, nullptr, nullptr, hSrcMem->CLMemory,
+      hDstMem->CLMemory, OpenCLOriginRect, OpenCLDstRect, OpenCLRegion,
+      srcRowPitch, srcSlicePitch, dstRowPitch, dstSlicePitch, WaitListSize,
+      SyncPointWaitList, RetSyncPoint, nullptr));
 
   return UR_RESULT_SUCCESS;
 }
@@ -376,7 +381,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferFillExp(
     [[maybe_unused]] ur_event_handle_t *phEvent,
     [[maybe_unused]] ur_exp_command_buffer_command_handle_t *phCommand) {
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clCommandFillBufferKHR_fn clCommandFillBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clCommandFillBufferKHR)>(
@@ -389,9 +394,9 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferAppendMemBufferFillExp(
       IsInOrder ? nullptr : pSyncPointWaitList;
   uint32_t WaitListSize = IsInOrder ? 0 : numSyncPointsInWaitList;
   CL_RETURN_ON_FAILURE(clCommandFillBufferKHR(
-      hCommandBuffer->CLCommandBuffer, nullptr, nullptr,
-      cl_adapter::cast<cl_mem>(hBuffer), pPattern, patternSize, offset, size,
-      WaitListSize, SyncPointWaitList, RetSyncPoint, nullptr));
+      hCommandBuffer->CLCommandBuffer, nullptr, nullptr, hBuffer->CLMemory,
+      pPattern, patternSize, offset, size, WaitListSize, SyncPointWaitList,
+      RetSyncPoint, nullptr));
 
   return UR_RESULT_SUCCESS;
 }
@@ -447,21 +452,25 @@ UR_APIEXPORT ur_result_t UR_APICALL urEnqueueCommandBufferExp(
     uint32_t numEventsInWaitList, const ur_event_handle_t *phEventWaitList,
     ur_event_handle_t *phEvent) {
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clEnqueueCommandBufferKHR_fn clEnqueueCommandBufferKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clEnqueueCommandBufferKHR)>(
           CLContext, cl_ext::ExtFuncPtrCache->clEnqueueCommandBufferKHRCache,
           cl_ext::EnqueueCommandBufferName, &clEnqueueCommandBufferKHR));
 
   const uint32_t NumberOfQueues = 1;
-
+  cl_event Event;
+  std::vector<cl_event> CLWaitEvents(numEventsInWaitList);
+  for (uint32_t i = 0; i < numEventsInWaitList; i++) {
+    CLWaitEvents[i] = phEventWaitList[i]->CLEvent;
+  }
+  cl_command_queue CLQueue = hQueue->CLQueue;
   CL_RETURN_ON_FAILURE(clEnqueueCommandBufferKHR(
-      NumberOfQueues, cl_adapter::cast<cl_command_queue *>(&hQueue),
-      hCommandBuffer->CLCommandBuffer, numEventsInWaitList,
-      cl_adapter::cast<const cl_event *>(phEventWaitList),
-      cl_adapter::cast<cl_event *>(phEvent)));
+      NumberOfQueues, &CLQueue, hCommandBuffer->CLCommandBuffer,
+      numEventsInWaitList, CLWaitEvents.data(), ifUrEvent(phEvent, Event)));
 
+  UR_RETURN_ON_FAILURE(createUREvent(Event, hQueue->Context, hQueue, phEvent));
   return UR_RESULT_SUCCESS;
 }
 
@@ -501,11 +510,11 @@ void updateKernelArgs(std::vector<cl_mutable_dispatch_arg_khr> &CLArgs,
   for (uint32_t i = 0; i < NumMemobjArgs; i++) {
     const ur_exp_command_buffer_update_memobj_arg_desc_t &URMemObjArg =
         ArgMemobjList[i];
+    cl_mem arg_value = URMemObjArg.hNewMemObjArg->CLMemory;
     cl_mutable_dispatch_arg_khr CLArg{
         URMemObjArg.argIndex, // arg_index
         sizeof(cl_mem),       // arg_size
-        cl_adapter::cast<const cl_mem *>(
-            &URMemObjArg.hNewMemObjArg) // arg_value
+        &arg_value            // arg_value
     };
 
     CLArgs.push_back(CLArg);
@@ -549,7 +558,7 @@ ur_result_t validateCommandDesc(
   // Verify that the device supports updating the aspects of the kernel that
   // the user is requesting.
   ur_device_handle_t URDevice = CommandBuffer->hDevice;
-  cl_device_id CLDevice = cl_adapter::cast<cl_device_id>(URDevice);
+  cl_device_id CLDevice = URDevice->CLDevice;
 
   ur_device_command_buffer_update_capability_flags_t UpdateCapabilities = 0;
   CL_RETURN_ON_FAILURE(
@@ -601,7 +610,8 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
         validateCommandDesc(hCommandBuffer, pUpdateKernelLaunch[i]));
   }
 
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
+
   cl_ext::clUpdateMutableCommandsKHR_fn clUpdateMutableCommandsKHR = nullptr;
   UR_RETURN_ON_FAILURE(
       cl_ext::getExtFuncFromContext<decltype(clUpdateMutableCommandsKHR)>(
@@ -657,8 +667,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urCommandBufferUpdateKernelLaunchExp(
       updateNDRange(CLLocalWorkSize, CommandWorkDim, LocalWorkSizePtr);
     }
 
-    cl_mutable_command_khr CLCommand =
-        cl_adapter::cast<cl_mutable_command_khr>(Command->CLMutableCommand);
+    cl_mutable_command_khr CLCommand = Command->CLMutableCommand;
     Config = cl_mutable_dispatch_config_khr{
         CLCommand,
         static_cast<cl_uint>(CLArgs.size()),    // num_args
@@ -736,7 +745,7 @@ ur_result_t UR_APICALL urCommandBufferAppendNativeCommandExp(
     uint32_t numSyncPointsInWaitList,
     const ur_exp_command_buffer_sync_point_t *pSyncPointWaitList,
     ur_exp_command_buffer_sync_point_t *pSyncPoint) {
-  cl_context CLContext = cl_adapter::cast<cl_context>(hCommandBuffer->hContext);
+  cl_context CLContext = hCommandBuffer->hContext->CLContext;
   cl_ext::clCommandBarrierWithWaitListKHR_fn clCommandBarrierWithWaitListKHR =
       nullptr;
   UR_RETURN_ON_FAILURE(

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

@@ -156,20 +156,6 @@ extern thread_local char ErrorMessage[MaxMessageSize];
 // Utility function for setting a message and warning
 [[maybe_unused]] void setErrorMessage(const char *Message,
                                       ur_result_t ErrorCode);
-
-template <class To, class From> To cast(From Value) {
-
-  if constexpr (std::is_pointer_v<From>) {
-    static_assert(std::is_pointer_v<From> == std::is_pointer_v<To>,
-                  "Cast failed pointer check");
-    return reinterpret_cast<To>(Value);
-  } else {
-    static_assert(sizeof(From) == sizeof(To), "Cast failed size check");
-    static_assert(std::is_signed_v<From> == std::is_signed_v<To>,
-                  "Cast failed sign check");
-    return static_cast<To>(Value);
-  }
-}
 } // namespace cl_adapter
 
 namespace cl_ext {