[L0] Add Support for External Semaphores

- Added support for using the Intel L0 driver experimental extension for
  external sempahores.
- This implementation enables support for external semaphores that will
  exist in a future L0 Intel GPU Driver implementation.
- The functionality outlined in this commit mirrors the planned
  functionality that will be published in the next offical L0 spec release 1.12.
  Once the L0 1.12 spec is released, then the usage of the header definitions
  of this functionality will be updated to match the official spec while
  supporting the previous driver implementation.

Signed-off-by: Neil R. Spruit <[email protected]>

Author: nrspruit

source/adapters/level_zero/common.hpp

@@ -531,3 +531,76 @@ extern thread_local int32_t ErrorAdapterNativeCode;
                                       int32_t AdapterErrorCode);
 
 #define L0_DRIVER_INORDER_MIN_VERSION 29534
+
+// Definitions for the External Semaphore Extension
+
+#ifndef ZE_INTEL_EXTERNAL_SEMAPHORE_EXP_NAME
+/// @brief Event sync mode extension name
+#define ZE_INTEL_EXTERNAL_SEMAPHORE_EXP_NAME                                   \
+  "ZE_intel_experimental_external_semaphore"
+#endif // ZE_INTEL_EXTERNAL_SEMAPHORE_EXP_NAME
+
+typedef enum _ze_intel_external_semaphore_exp_version_t {
+  ZE_EXTERNAL_SEMAPHORE_EXP_VERSION_1_0 =
+      ZE_MAKE_VERSION(1, 0), ///< version 1.0
+  ZE_EXTERNAL_SEMAPHORE_EXP_VERSION_CURRENT =
+      ZE_MAKE_VERSION(1, 0), ///< latest known version
+  ZE_EXTERNAL_SEMAPHORE_EXP_VERSION_FORCE_UINT32 = 0x7fffffff
+} ze_intel_external_semaphore_exp_version_t;
+typedef enum _ze_intel_external_semaphore_exp_flags_t {
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_FD,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_WIN32,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_WIN32_KMT,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_D3D12_FENCE,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_D3D11_FENCE,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_KEYED_MUTEX,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_KEYED_MUTEX_KMT,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_TIMELINE_SEMAPHORE_FD,
+  ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_TIMELINE_SEMAPHORE_WIN32
+} ze_intel_external_semaphore_exp_flags_t;
+
+typedef struct _ze_intel_external_semaphore_exp_desc_t {
+  ze_structure_type_t stype;
+  const void *pNext;
+  ze_intel_external_semaphore_exp_flags_t flags;
+} ze_intel_external_semaphore_exp_desc_t;
+
+typedef struct _ze_intel_external_semaphore_win32_exp_desc_t {
+  ze_structure_type_t stype;
+  const void *pNext;
+  void *handle;
+  const char *name;
+} ze_intel_external_semaphore_win32_exp_desc_t;
+
+typedef struct _ze_intel_external_semaphore_fd_exp_desc_t {
+  ze_structure_type_t stype;
+  const void *pNext;
+  int fd;
+} ze_intel_external_semaphore_desc_fd_exp_desc_t;
+
+typedef struct _ze_intel_external_semaphore_signal_exp_params_t {
+  ze_structure_type_t stype;
+  const void *pNext;
+  uint64_t value;
+} ze_intel_external_semaphore_signal_exp_params_t;
+
+typedef struct _ze_intel_external_semaphore_wait_exp_params_t {
+  ze_structure_type_t stype;
+  const void *pNext;
+
+  uint64_t value;
+} ze_intel_external_semaphore_wait_exp_params_t;
+
+typedef struct _ze_intel_external_semaphore_exp_handle_t
+    *ze_intel_external_semaphore_exp_handle_t;
+
+#define ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_EXP_DESC                    \
+  (ze_structure_type_t)0x0003001E
+#define ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_WIN32_EXP_DESC              \
+  (ze_structure_type_t)0x0003001F
+#define ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_FD_EXP_DESC                 \
+  (ze_structure_type_t)0x00030023
+#define ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_EXP           \
+  (ze_structure_type_t)0x00030024
+#define ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_WAIT_PARAMS_EXP             \
+  (ze_structure_type_t)0x00030025

source/adapters/level_zero/image.cpp

@@ -1190,41 +1190,130 @@ ur_result_t urBindlessImagesImportExternalSemaphoreExp(
     ur_exp_external_semaphore_type_t semHandleType,
     ur_exp_external_semaphore_desc_t *pExternalSemaphoreDesc,
     ur_exp_external_semaphore_handle_t *phExternalSemaphoreHandle) {
-  std::ignore = hContext;
-  std::ignore = hDevice;
-  std::ignore = semHandleType;
-  std::ignore = pExternalSemaphoreDesc;
-  std::ignore = phExternalSemaphoreHandle;
-  logger::error(logger::LegacyMessage("[UR][L0] {} function not implemented!"),
-                "{} function not implemented!", __FUNCTION__);
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+
+  auto UrPlatform = hContext->getPlatform();
+  if (UrPlatform->ZeExternalSemaphoreExt.Supported == false) {
+    logger::error(logger::LegacyMessage("[UR][L0] "),
+                  " {} function not supported!", __FUNCTION__);
+    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  }
+  ze_intel_external_semaphore_exp_desc_t SemDesc = {
+      ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_EXP_DESC, nullptr,
+      ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_FD};
+  ze_intel_external_semaphore_exp_handle_t ExtSemaphoreHandle;
+  ze_intel_external_semaphore_desc_fd_exp_desc_t FDExpDesc = {
+      ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_FD_EXP_DESC, nullptr, 0};
+  _ze_intel_external_semaphore_win32_exp_desc_t Win32ExpDesc = {
+      ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_WIN32_EXP_DESC, nullptr,
+      nullptr, nullptr};
+  void *pNext = const_cast<void *>(pExternalSemaphoreDesc->pNext);
+  while (pNext != nullptr) {
+    const ur_base_desc_t *BaseDesc = static_cast<const ur_base_desc_t *>(pNext);
+    if (BaseDesc->stype == UR_STRUCTURE_TYPE_EXP_FILE_DESCRIPTOR) {
+      auto FileDescriptor =
+          static_cast<const ur_exp_file_descriptor_t *>(pNext);
+      FDExpDesc.fd = FileDescriptor->fd;
+      SemDesc.pNext = &FDExpDesc;
+      SemDesc.flags = ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_FD;
+    } else if (BaseDesc->stype == UR_STRUCTURE_TYPE_EXP_WIN32_HANDLE) {
+      SemDesc.pNext = &Win32ExpDesc;
+      auto Win32Handle = static_cast<const ur_exp_win32_handle_t *>(pNext);
+      switch (semHandleType) {
+      case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_WIN32_NT:
+        SemDesc.flags = ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_WIN32;
+        break;
+      case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_WIN32_NT_DX12_FENCE:
+        SemDesc.flags = ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_D3D12_FENCE;
+        break;
+      case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_OPAQUE_FD:
+        SemDesc.flags = ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_FD;
+        break;
+      default:
+        return UR_RESULT_ERROR_INVALID_VALUE;
+      }
+      Win32ExpDesc.handle = Win32Handle->handle;
+    }
+    pNext = const_cast<void *>(BaseDesc->pNext);
+  }
+
+  ZE2UR_CALL(UrPlatform->ZeExternalSemaphoreExt.zexImportExternalSemaphoreExp,
+             (hDevice->ZeDevice, &ExtSemaphoreHandle, &SemDesc));
+  *phExternalSemaphoreHandle =
+      (ur_exp_external_semaphore_handle_t)ExtSemaphoreHandle;
+
+  return UR_RESULT_SUCCESS;
 }
 
 ur_result_t urBindlessImagesReleaseExternalSemaphoreExp(
     ur_context_handle_t hContext, ur_device_handle_t hDevice,
     ur_exp_external_semaphore_handle_t hExternalSemaphore) {
-  std::ignore = hContext;
   std::ignore = hDevice;
-  std::ignore = hExternalSemaphore;
-  logger::error(logger::LegacyMessage("[UR][L0] {} function not implemented!"),
-                "{} function not implemented!", __FUNCTION__);
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  auto UrPlatform = hContext->getPlatform();
+  if (UrPlatform->ZeExternalSemaphoreExt.Supported == false) {
+    logger::error(logger::LegacyMessage("[UR][L0] "),
+                  " {} function not supported!", __FUNCTION__);
+    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  }
+  ZE2UR_CALL(
+      UrPlatform->ZeExternalSemaphoreExt.zexDeviceReleaseExternalSemaphoreExp,
+      ((ze_intel_external_semaphore_exp_handle_t)hExternalSemaphore));
+
+  return UR_RESULT_SUCCESS;
 }
 
 ur_result_t urBindlessImagesWaitExternalSemaphoreExp(
     ur_queue_handle_t hQueue, ur_exp_external_semaphore_handle_t hSemaphore,
     bool hasValue, uint64_t waitValue, uint32_t numEventsInWaitList,
     const ur_event_handle_t *phEventWaitList, ur_event_handle_t *phEvent) {
-  std::ignore = hQueue;
-  std::ignore = hSemaphore;
-  std::ignore = hasValue;
-  std::ignore = waitValue;
-  std::ignore = numEventsInWaitList;
-  std::ignore = phEventWaitList;
-  std::ignore = phEvent;
-  logger::error(logger::LegacyMessage("[UR][L0] "),
-                " {} function not implemented!", __FUNCTION__);
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  auto UrPlatform = hQueue->Context->getPlatform();
+  if (UrPlatform->ZeExternalSemaphoreExt.Supported == false) {
+    logger::error(logger::LegacyMessage("[UR][L0] "),
+                  " {} function not supported!", __FUNCTION__);
+    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  }
+
+  bool UseCopyEngine = false;
+
+  // We want to batch these commands to avoid extra submissions (costly)
+  bool OkToBatch = true;
+
+  _ur_ze_event_list_t TmpWaitList;
+  UR_CALL(TmpWaitList.createAndRetainUrZeEventList(
+      numEventsInWaitList, phEventWaitList, hQueue, UseCopyEngine));
+
+  // Get a new command list to be used on this call
+  ur_command_list_ptr_t CommandList{};
+  UR_CALL(hQueue->Context->getAvailableCommandList(
+      hQueue, CommandList, UseCopyEngine, numEventsInWaitList, phEventWaitList,
+      OkToBatch, nullptr /*ForcedCmdQueue*/));
+
+  ze_event_handle_t ZeEvent = nullptr;
+  ur_event_handle_t InternalEvent;
+  bool IsInternal = phEvent == nullptr;
+  ur_event_handle_t *Event = phEvent ? phEvent : &InternalEvent;
+  UR_CALL(createEventAndAssociateQueue(hQueue, Event,
+                                       UR_COMMAND_EXTERNAL_SEMAPHORE_WAIT_EXP,
+                                       CommandList, IsInternal,
+                                       /*IsMultiDevice*/ false));
+  UR_CALL(setSignalEvent(hQueue, UseCopyEngine, &ZeEvent, Event,
+                         numEventsInWaitList, phEventWaitList,
+                         CommandList->second.ZeQueue));
+  (*Event)->WaitList = TmpWaitList;
+
+  const auto &ZeCommandList = CommandList->first;
+  const auto &WaitList = (*Event)->WaitList;
+
+  ze_intel_external_semaphore_wait_exp_params_t WaitParams = {
+      ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_WAIT_PARAMS_EXP, nullptr, 0};
+  WaitParams.value = hasValue ? waitValue : 0;
+  const ze_intel_external_semaphore_exp_handle_t hExtSemaphore =
+      reinterpret_cast<ze_intel_external_semaphore_exp_handle_t>(hSemaphore);
+  ZE2UR_CALL(UrPlatform->ZeExternalSemaphoreExt
+                 .zexCommandListAppendWaitExternalSemaphoresExp,
+             (ZeCommandList, &hExtSemaphore, &WaitParams, 1, ZeEvent,
+              WaitList.Length, WaitList.ZeEventList));
+
+  return UR_RESULT_SUCCESS;
 }
 
 ur_result_t urBindlessImagesSignalExternalSemaphoreExp(
@@ -1238,9 +1327,56 @@ ur_result_t urBindlessImagesSignalExternalSemaphoreExp(
   std::ignore = numEventsInWaitList;
   std::ignore = phEventWaitList;
   std::ignore = phEvent;
-  logger::error(logger::LegacyMessage("[UR][L0] {} function not implemented!"),
-                "{} function not implemented!", __FUNCTION__);
-  return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  auto UrPlatform = hQueue->Context->getPlatform();
+  if (UrPlatform->ZeExternalSemaphoreExt.Supported == false) {
+    logger::error(logger::LegacyMessage("[UR][L0] "),
+                  " {} function not supported!", __FUNCTION__);
+    return UR_RESULT_ERROR_UNSUPPORTED_FEATURE;
+  }
+
+  bool UseCopyEngine = false;
+
+  // We want to batch these commands to avoid extra submissions (costly)
+  bool OkToBatch = true;
+
+  _ur_ze_event_list_t TmpWaitList;
+  UR_CALL(TmpWaitList.createAndRetainUrZeEventList(
+      numEventsInWaitList, phEventWaitList, hQueue, UseCopyEngine));
+
+  // Get a new command list to be used on this call
+  ur_command_list_ptr_t CommandList{};
+  UR_CALL(hQueue->Context->getAvailableCommandList(
+      hQueue, CommandList, UseCopyEngine, numEventsInWaitList, phEventWaitList,
+      OkToBatch, nullptr /*ForcedCmdQueue*/));
+
+  ze_event_handle_t ZeEvent = nullptr;
+  ur_event_handle_t InternalEvent;
+  bool IsInternal = phEvent == nullptr;
+  ur_event_handle_t *Event = phEvent ? phEvent : &InternalEvent;
+  UR_CALL(createEventAndAssociateQueue(hQueue, Event,
+                                       UR_COMMAND_EXTERNAL_SEMAPHORE_SIGNAL_EXP,
+                                       CommandList, IsInternal,
+                                       /*IsMultiDevice*/ false));
+  UR_CALL(setSignalEvent(hQueue, UseCopyEngine, &ZeEvent, Event,
+                         numEventsInWaitList, phEventWaitList,
+                         CommandList->second.ZeQueue));
+  (*Event)->WaitList = TmpWaitList;
+
+  const auto &ZeCommandList = CommandList->first;
+  const auto &WaitList = (*Event)->WaitList;
+
+  ze_intel_external_semaphore_signal_exp_params_t SignalParams = {
+      ZE_INTEL_STRUCTURE_TYPE_EXTERNAL_SEMAPHORE_SIGNAL_PARAMS_EXP, nullptr, 0};
+  SignalParams.value = hasValue ? signalValue : 0;
+  const ze_intel_external_semaphore_exp_handle_t hExtSemaphore =
+      reinterpret_cast<ze_intel_external_semaphore_exp_handle_t>(hSemaphore);
+
+  ZE2UR_CALL(UrPlatform->ZeExternalSemaphoreExt
+                 .zexCommandListAppendSignalExternalSemaphoresExp,
+             (ZeCommandList, &hExtSemaphore, &SignalParams, 1, ZeEvent,
+              WaitList.Length, WaitList.ZeEventList));
+
+  return UR_RESULT_SUCCESS;
 }
 
 } // namespace ur::level_zero

source/adapters/level_zero/platform.cpp

@@ -221,6 +221,7 @@ ur_result_t ur_platform_handle_t_::initialize() {
              (ZeDriver, &Count, ZeExtensions.data()));
 
   bool MutableCommandListSpecExtensionSupported = false;
+  bool ZeIntelExternalSemaphoreExtensionSupported = false;
   for (auto &extension : ZeExtensions) {
     // Check if global offset extension is available
     if (strncmp(extension.name, ZE_GLOBAL_OFFSET_EXP_NAME,
@@ -252,6 +253,13 @@ ur_result_t ur_platform_handle_t_::initialize() {
         MutableCommandListSpecExtensionSupported = true;
       }
     }
+    // Check if extension is available for External Sempahores
+    if (strncmp(extension.name, ZE_INTEL_EXTERNAL_SEMAPHORE_EXP_NAME,
+                strlen(ZE_INTEL_EXTERNAL_SEMAPHORE_EXP_NAME) + 1) == 0) {
+      if (extension.version == ZE_EXTERNAL_SEMAPHORE_EXP_VERSION_1_0) {
+        ZeIntelExternalSemaphoreExtensionSupported = true;
+      }
+    }
     zeDriverExtensionMap[extension.name] = extension.version;
   }
 
@@ -286,6 +294,38 @@ ur_result_t ur_platform_handle_t_::initialize() {
   // If yes, then set up L0 API pointers if the platform supports it.
   ZeUSMImport.setZeUSMImport(this);
 
+  if (ZeIntelExternalSemaphoreExtensionSupported) {
+    ZeExternalSemaphoreExt.Supported |=
+        (ZE_CALL_NOCHECK(
+             zeDriverGetExtensionFunctionAddress,
+             (ZeDriver, "zeIntelDeviceImportExternalSemaphoreExp",
+              reinterpret_cast<void **>(
+                  &ZeExternalSemaphoreExt.zexImportExternalSemaphoreExp))) ==
+         0);
+    ZeExternalSemaphoreExt.Supported |=
+        (ZE_CALL_NOCHECK(
+             zeDriverGetExtensionFunctionAddress,
+             (ZeDriver, "zeIntelCommandListAppendWaitExternalSemaphoresExp",
+              reinterpret_cast<void **>(
+                  &ZeExternalSemaphoreExt
+                       .zexCommandListAppendWaitExternalSemaphoresExp))) == 0);
+    ZeExternalSemaphoreExt.Supported |=
+        (ZE_CALL_NOCHECK(
+             zeDriverGetExtensionFunctionAddress,
+             (ZeDriver, "zeIntelCommandListAppendSignalExternalSemaphoresExp",
+              reinterpret_cast<void **>(
+                  &ZeExternalSemaphoreExt
+                       .zexCommandListAppendSignalExternalSemaphoresExp))) ==
+         0);
+    ZeExternalSemaphoreExt.Supported |=
+        (ZE_CALL_NOCHECK(zeDriverGetExtensionFunctionAddress,
+                         (ZeDriver, "zeIntelDeviceReleaseExternalSemaphoreExp",
+                          reinterpret_cast<void **>(
+                              &ZeExternalSemaphoreExt
+                                   .zexDeviceReleaseExternalSemaphoreExp))) ==
+         0);
+  }
+
   // Check if mutable command list extension is supported and initialize
   // function pointers.
   if (MutableCommandListSpecExtensionSupported) {

source/adapters/level_zero/platform.hpp

@@ -114,4 +114,24 @@ struct ur_platform_handle_t_ : public _ur_platform {
         ze_command_list_handle_t, const ze_mutable_command_id_exp_desc_t *,
         uint32_t, ze_kernel_handle_t *, uint64_t *) = nullptr;
   } ZeMutableCmdListExt;
-};
+
+  // Structure with function pointers for External Semaphore Extension.
+  struct ZeExternalSemaphoreExtension {
+    bool Supported = false;
+    ze_result_t (*zexImportExternalSemaphoreExp)(
+        ze_device_handle_t, ze_intel_external_semaphore_exp_handle_t *,
+        const ze_intel_external_semaphore_exp_desc_t *);
+    ze_result_t (*zexCommandListAppendWaitExternalSemaphoresExp)(
+        ze_command_list_handle_t,
+        const ze_intel_external_semaphore_exp_handle_t *,
+        const ze_intel_external_semaphore_wait_exp_params_t *, unsigned int,
+        ze_event_handle_t, uint32_t, ze_event_handle_t *);
+    ze_result_t (*zexCommandListAppendSignalExternalSemaphoresExp)(
+        ze_command_list_handle_t,
+        const ze_intel_external_semaphore_exp_handle_t *,
+        const ze_intel_external_semaphore_signal_exp_params_t *, size_t,
+        ze_event_handle_t, uint32_t, ze_event_handle_t *);
+    ze_result_t (*zexDeviceReleaseExternalSemaphoreExp)(
+        ze_intel_external_semaphore_exp_handle_t);
+  } ZeExternalSemaphoreExt;
+};