Merge pull request #1044 from aarongreig/aaron/clCTSFixMegaBranch

[OpenCL] Combined CTS fixes

Author: kbenzie

source/adapters/opencl/common.cpp

@@ -60,6 +60,23 @@ ur_result_t mapCLErrorToUR(cl_int Result) {
     return UR_RESULT_ERROR_OUT_OF_RESOURCES;
   case CL_INVALID_MEM_OBJECT:
     return UR_RESULT_ERROR_INVALID_MEM_OBJECT;
+  case CL_INVALID_QUEUE_PROPERTIES:
+    return UR_RESULT_ERROR_INVALID_QUEUE_PROPERTIES;
+  case CL_INVALID_BUFFER_SIZE:
+    return UR_RESULT_ERROR_INVALID_BUFFER_SIZE;
+  case CL_INVALID_IMAGE_SIZE:
+    return UR_RESULT_ERROR_INVALID_IMAGE_SIZE;
+  case CL_INVALID_IMAGE_FORMAT_DESCRIPTOR:
+  case CL_INVALID_IMAGE_DESCRIPTOR:
+    return UR_RESULT_ERROR_INVALID_IMAGE_FORMAT_DESCRIPTOR;
+  case CL_IMAGE_FORMAT_NOT_SUPPORTED:
+    return UR_RESULT_ERROR_UNSUPPORTED_IMAGE_FORMAT;
+  case CL_PROFILING_INFO_NOT_AVAILABLE:
+    return UR_RESULT_ERROR_PROFILING_INFO_NOT_AVAILABLE;
+  case CL_LINK_PROGRAM_FAILURE:
+    return UR_RESULT_ERROR_PROGRAM_LINK_FAILURE;
+  case CL_INVALID_ARG_INDEX:
+    return UR_RESULT_ERROR_INVALID_KERNEL_ARGUMENT_INDEX;
   default:
     return UR_RESULT_ERROR_UNKNOWN;
   }

source/adapters/opencl/context.cpp

@@ -10,6 +10,10 @@
 
 #include "context.hpp"
 
+#include <mutex>
+#include <set>
+#include <unordered_map>
+
 ur_result_t cl_adapter::getDevicesFromContext(
     ur_context_handle_t hContext,
     std::unique_ptr<std::vector<cl_device_id>> &DevicesInCtx) {
@@ -89,10 +93,17 @@ urContextGetInfo(ur_context_handle_t hContext, ur_context_info_t propName,
   case UR_CONTEXT_INFO_NUM_DEVICES:
   case UR_CONTEXT_INFO_DEVICES:
   case UR_CONTEXT_INFO_REFERENCE_COUNT: {
-
-    CL_RETURN_ON_FAILURE(
+    size_t CheckPropSize = 0;
+    auto ClResult =
         clGetContextInfo(cl_adapter::cast<cl_context>(hContext), CLPropName,
-                         propSize, pPropValue, pPropSizeRet));
+                         propSize, pPropValue, &CheckPropSize);
+    if (pPropValue && CheckPropSize != propSize) {
+      return UR_RESULT_ERROR_INVALID_SIZE;
+    }
+    CL_RETURN_ON_FAILURE(ClResult);
+    if (pPropSizeRet) {
+      *pPropSizeRet = CheckPropSize;
+    }
     return UR_RESULT_SUCCESS;
   }
   default:
@@ -130,8 +141,53 @@ UR_APIEXPORT ur_result_t UR_APICALL urContextCreateWithNativeHandle(
 }
 
 UR_APIEXPORT ur_result_t UR_APICALL urContextSetExtendedDeleter(
-    [[maybe_unused]] ur_context_handle_t hContext,
-    [[maybe_unused]] ur_context_extended_deleter_t pfnDeleter,
-    [[maybe_unused]] void *pUserData) {
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+    ur_context_handle_t hContext, ur_context_extended_deleter_t pfnDeleter,
+    void *pUserData) {
+  static std::unordered_map<ur_context_handle_t,
+                            std::set<ur_context_extended_deleter_t>>
+      ContextCallbackMap;
+  static std::mutex ContextCallbackMutex;
+
+  {
+    std::lock_guard<std::mutex> Lock(ContextCallbackMutex);
+    // Callbacks can only be registered once and we need to avoid double
+    // allocating.
+    if (ContextCallbackMap.count(hContext) &&
+        ContextCallbackMap[hContext].count(pfnDeleter)) {
+      return UR_RESULT_SUCCESS;
+    }
+
+    ContextCallbackMap[hContext].insert(pfnDeleter);
+  }
+
+  struct ContextCallback {
+    void execute() {
+      pfnDeleter(pUserData);
+      {
+        std::lock_guard<std::mutex> Lock(*CallbackMutex);
+        (*CallbackMap)[hContext].erase(pfnDeleter);
+        if ((*CallbackMap)[hContext].empty()) {
+          CallbackMap->erase(hContext);
+        }
+      }
+      delete this;
+    }
+    ur_context_handle_t hContext;
+    ur_context_extended_deleter_t pfnDeleter;
+    void *pUserData;
+    std::unordered_map<ur_context_handle_t,
+                       std::set<ur_context_extended_deleter_t>> *CallbackMap;
+    std::mutex *CallbackMutex;
+  };
+  auto Callback =
+      new ContextCallback({hContext, pfnDeleter, pUserData, &ContextCallbackMap,
+                           &ContextCallbackMutex});
+  auto ClCallback = [](cl_context, void *pUserData) {
+    auto *C = static_cast<ContextCallback *>(pUserData);
+    C->execute();
+  };
+  CL_RETURN_ON_FAILURE(clSetContextDestructorCallback(
+      cl_adapter::cast<cl_context>(hContext), ClCallback, Callback));
+
+  return UR_RESULT_SUCCESS;
 }

source/adapters/opencl/device.cpp

@@ -345,6 +345,23 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
 
     return ReturnValue(URDeviceType);
   }
+  case UR_DEVICE_INFO_DEVICE_ID: {
+    bool Supported = false;
+    CL_RETURN_ON_FAILURE(cl_adapter::checkDeviceExtensions(
+        cl_adapter::cast<cl_device_id>(hDevice), {"cl_khr_pci_bus_info"},
+        Supported));
+
+    if (!Supported) {
+      return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
+    }
+
+    cl_device_pci_bus_info_khr PciInfo = {};
+    CL_RETURN_ON_FAILURE(clGetDeviceInfo(
+        cl_adapter::cast<cl_device_id>(hDevice), CL_DEVICE_PCI_BUS_INFO_KHR,
+        sizeof(PciInfo), &PciInfo, nullptr));
+    return ReturnValue(PciInfo.pci_device);
+  }
+
   case UR_DEVICE_INFO_BACKEND_RUNTIME_VERSION: {
     oclv::OpenCLVersion Version;
     CL_RETURN_ON_FAILURE(cl_adapter::getDeviceVersion(
@@ -760,6 +777,16 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
 
     return ReturnValue(Supported);
   }
+  case UR_DEVICE_INFO_VIRTUAL_MEMORY_SUPPORT: {
+    return ReturnValue(false);
+  }
+  case UR_DEVICE_INFO_HOST_PIPE_READ_WRITE_SUPPORTED: {
+    bool Supported = false;
+    CL_RETURN_ON_FAILURE(cl_adapter::checkDeviceExtensions(
+        cl_adapter::cast<cl_device_id>(hDevice),
+        {"cl_intel_program_scope_host_pipe"}, Supported));
+    return ReturnValue(Supported);
+  }
   case UR_DEVICE_INFO_QUEUE_PROPERTIES:
   case UR_DEVICE_INFO_QUEUE_ON_DEVICE_PROPERTIES:
   case UR_DEVICE_INFO_QUEUE_ON_HOST_PROPERTIES:
@@ -775,7 +802,7 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
     /* CL type: cl_bitfield / enum
      * UR type: ur_flags_t (uint32_t) */
 
-    cl_bitfield CLValue;
+    cl_bitfield CLValue = 0;
     CL_RETURN_ON_FAILURE(
         clGetDeviceInfo(cl_adapter::cast<cl_device_id>(hDevice), CLPropName,
                         sizeof(cl_bitfield), &CLValue, nullptr));
@@ -898,13 +925,12 @@ UR_APIEXPORT ur_result_t UR_APICALL urDeviceGetInfo(ur_device_handle_t hDevice,
    * sycl/doc/extensions/supported/sycl_ext_intel_device_info.md */
   case UR_DEVICE_INFO_UUID:
   /* This enums have no equivalent in OpenCL */
-  case UR_DEVICE_INFO_DEVICE_ID:
+  case UR_DEVICE_INFO_MAX_REGISTERS_PER_WORK_GROUP:
   case UR_DEVICE_INFO_GLOBAL_MEM_FREE:
   case UR_DEVICE_INFO_MEMORY_CLOCK_RATE:
   case UR_DEVICE_INFO_MEMORY_BUS_WIDTH:
-  case UR_DEVICE_INFO_ASYNC_BARRIER:
-  case UR_DEVICE_INFO_HOST_PIPE_READ_WRITE_SUPPORTED: {
-    return UR_RESULT_ERROR_INVALID_ENUMERATION;
+  case UR_DEVICE_INFO_ASYNC_BARRIER: {
+    return UR_RESULT_ERROR_UNSUPPORTED_ENUMERATION;
   }
   default: {
     return UR_RESULT_ERROR_INVALID_ENUMERATION;