Add support for timeline semaphores (#17395)

Add support for timeline semaphores on linux and windows for CUDA
backend and preliminary support in L0 backend. Update
vulkan_interop/sampled_images.cpp to remove redundant wait when using
semaphores.

Author: DBDuncan

include/ur_api.h

@@ -123,6 +123,8 @@ Enums
     * ${X}_EXP_EXTERNAL_SEMAPHORE_TYPE_OPAQUE_FD
     * ${X}_EXP_EXTERNAL_SEMAPHORE_TYPE_WIN32_NT
     * ${X}_EXP_EXTERNAL_SEMAPHORE_TYPE_WIN32_NT_DX12_FENCE
+    * ${X}_EXP_EXTERNAL_SEMAPHORE_TYPE_TIMELINE_FD
+    * ${X}_EXP_EXTERNAL_SEMAPHORE_TYPE_TIMELINE_WIN32_NT
 
 * ${x}_function_t
     * ${X}_FUNCTION_USM_PITCHED_ALLOC_EXP
@@ -258,6 +260,8 @@ Changelog
 |          || Added support for DtoD usm pitch copies                    |
 |          || Added support for HtoH copies                              |
 +----------+-------------------------------------------------------------+
+| 20.0     | Added timeline semaphores for CUDA and L0 backends          |
++----------+-------------------------------------------------------------+
 
 Contributors
 --------------------------------------------------------------------------------

include/ur_print.hpp

@@ -207,11 +207,15 @@ class: $xBindlessImages
 name: $x_exp_external_semaphore_type_t
 etors:
   - name: OPAQUE_FD
-    desc: "Opaque file descriptor"
+    desc: "Binary semaphore opaque file descriptor"
   - name: WIN32_NT
-    desc: "Win32 NT handle"
+    desc: "Binary semaphore Win32 NT handle"
   - name: WIN32_NT_DX12_FENCE
-    desc: "Win32 NT DirectX 12 fence handle"
+    desc: "Fence semaphore Win32 NT DirectX 12 handle"
+  - name: TIMELINE_FD
+    desc: "Timeline semaphore opaque file descriptor"
+  - name: TIMELINE_WIN32_NT
+    desc: "Timeline semaphore Win32 NT handle"
 --- #--------------------------------------------------------------------------
 type: struct
 desc: "File descriptor"

scripts/core/EXP-BINDLESS-IMAGES.rst

@@ -1234,7 +1234,17 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImportExternalSemaphoreExp(
             static_cast<const ur_exp_file_descriptor_t *>(pNext);
 
         extSemDesc.handle.fd = FileDescriptor->fd;
-        extSemDesc.type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD;
+        switch (semHandleType) {
+        case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_OPAQUE_FD:
+          extSemDesc.type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_OPAQUE_FD;
+          break;
+        case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_TIMELINE_FD:
+          extSemDesc.type =
+              CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_FD;
+          break;
+        default:
+          return UR_RESULT_ERROR_INVALID_VALUE;
+        }
       } else if (BaseDesc->stype == UR_STRUCTURE_TYPE_EXP_WIN32_HANDLE) {
         auto Win32Handle = static_cast<const ur_exp_win32_handle_t *>(pNext);
         switch (semHandleType) {
@@ -1244,7 +1254,10 @@ UR_APIEXPORT ur_result_t UR_APICALL urBindlessImagesImportExternalSemaphoreExp(
         case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_WIN32_NT_DX12_FENCE:
           extSemDesc.type = CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_D3D12_FENCE;
           break;
-        case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_OPAQUE_FD:
+        case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_TIMELINE_WIN32_NT:
+          extSemDesc.type =
+              CU_EXTERNAL_SEMAPHORE_HANDLE_TYPE_TIMELINE_SEMAPHORE_WIN32;
+          break;
         default:
           return UR_RESULT_ERROR_INVALID_VALUE;
         }

scripts/core/exp-bindless-images.yml

@@ -799,7 +799,16 @@ ur_result_t urBindlessImagesImportExternalSemaphoreExp(
           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;
+      switch (semHandleType) {
+      case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_OPAQUE_FD:
+        SemDesc.flags = ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_OPAQUE_FD;
+        break;
+      case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_TIMELINE_FD:
+        SemDesc.flags = ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_TIMELINE_SEMAPHORE_FD;
+        break;
+      default:
+        return UR_RESULT_ERROR_INVALID_VALUE;
+      }
     } else if (BaseDesc->stype == UR_STRUCTURE_TYPE_EXP_WIN32_HANDLE) {
       SemDesc.pNext = &Win32ExpDesc;
       auto Win32Handle = static_cast<const ur_exp_win32_handle_t *>(pNext);
@@ -810,8 +819,9 @@ ur_result_t urBindlessImagesImportExternalSemaphoreExp(
       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;
+      case UR_EXP_EXTERNAL_SEMAPHORE_TYPE_TIMELINE_WIN32_NT:
+        SemDesc.flags =
+            ZE_EXTERNAL_SEMAPHORE_EXP_FLAGS_TIMELINE_SEMAPHORE_WIN32;
         break;
       default:
         return UR_RESULT_ERROR_INVALID_VALUE;