Merge pull request #2139 from nrspruit/zeHandle_copy_dependencies

[L0] Pass and track event dependencies required before executing Memory Copy buffer inits
oneapi-src · Oct 2, 2024 · 00f958f · 00f958f
2 parents cc2d590 + 98a67a2
commit 00f958f
Show file tree

Hide file tree

Showing 6 changed files with 146 additions and 59 deletions.
diff --git a/.github/workflows/multi_device.yml b/.github/workflows/multi_device.yml
@@ -30,12 +30,11 @@ jobs:
     - name: Install pip packages
       run: pip install -r third_party/requirements.txt
 
-    # TODO: enable once test failure are fixed/ignored
-    # - name: Download DPC++
-    #   run: |
-    #     wget -O ${{github.workspace}}/dpcpp_compiler.tar.gz https://github.com/intel/llvm/releases/download/nightly-2024-09-27/sycl_linux.tar.gz
-    #     mkdir dpcpp_compiler
-    #     tar -xvf ${{github.workspace}}/dpcpp_compiler.tar.gz -C dpcpp_compiler
+    - name: Download DPC++
+      run: |
+        wget -O ${{github.workspace}}/dpcpp_compiler.tar.gz https://github.com/intel/llvm/releases/download/nightly-2024-01-29/sycl_linux.tar.gz
+        mkdir dpcpp_compiler
+        tar -xvf ${{github.workspace}}/dpcpp_compiler.tar.gz -C dpcpp_compiler
 
     - name: Configure CMake
       shell: bash -el {0}
@@ -49,6 +48,8 @@ jobs:
         -DUR_BUILD_TESTS=ON
         -DUR_BUILD_ADAPTER_${{matrix.adapter.name}}=ON
         -DUR_TEST_DEVICES_COUNT=2
+        -DUR_DPCXX=${{github.workspace}}/dpcpp_compiler/bin/clang++
+        -DUR_SYCL_LIBRARY_DIR=${{github.workspace}}/dpcpp_compiler/lib
 
     - name: Build
       run: cmake --build ${{github.workspace}}/build -j $(nproc)
@@ -60,4 +61,4 @@ jobs:
 
     - name: Test adapters
       working-directory: ${{github.workspace}}/build
-      run: env UR_CTS_ADAPTER_PLATFORM="${{matrix.adapter.platform}}" ctest -C ${{matrix.build_type}} --output-on-failure -L "conformance" --timeout 180
+      run: env UR_CTS_ADAPTER_PLATFORM="${{matrix.adapter.platform}}" ctest -C ${{matrix.build_type}} --output-on-failure -L "conformance" -E "enqueue|kernel|program|integration|exp_command_buffer|exp_enqueue_native|exp_launch_properties|exp_usm_p2p" --timeout 180
diff --git a/source/adapters/level_zero/command_buffer.cpp b/source/adapters/level_zero/command_buffer.cpp
@@ -672,7 +672,7 @@ setKernelPendingArguments(ur_exp_command_buffer_handle_t CommandBuffer,
     char **ZeHandlePtr = nullptr;
     if (Arg.Value) {
       UR_CALL(Arg.Value->getZeHandlePtr(ZeHandlePtr, Arg.AccessMode,
-                                        CommandBuffer->Device));
+                                        CommandBuffer->Device, nullptr, 0u));
     }
     ZE2UR_CALL(zeKernelSetArgumentValue,
                (Kernel->ZeKernel, Arg.Index, Arg.Size, ZeHandlePtr));
@@ -826,10 +826,10 @@ ur_result_t urCommandBufferAppendMemBufferCopyExp(
 
   char *ZeHandleSrc;
   UR_CALL(SrcBuffer->getZeHandle(ZeHandleSrc, ur_mem_handle_t_::read_only,
-                                 CommandBuffer->Device));
+                                 CommandBuffer->Device, nullptr, 0u));
   char *ZeHandleDst;
   UR_CALL(DstBuffer->getZeHandle(ZeHandleDst, ur_mem_handle_t_::write_only,
-                                 CommandBuffer->Device));
+                                 CommandBuffer->Device, nullptr, 0u));
 
   bool PreferCopyEngine = (SrcBuffer->OnHost || DstBuffer->OnHost);
 
@@ -858,10 +858,10 @@ ur_result_t urCommandBufferAppendMemBufferCopyRectExp(
 
   char *ZeHandleSrc;
   UR_CALL(SrcBuffer->getZeHandle(ZeHandleSrc, ur_mem_handle_t_::read_only,
-                                 CommandBuffer->Device));
+                                 CommandBuffer->Device, nullptr, 0u));
   char *ZeHandleDst;
   UR_CALL(DstBuffer->getZeHandle(ZeHandleDst, ur_mem_handle_t_::write_only,
-                                 CommandBuffer->Device));
+                                 CommandBuffer->Device, nullptr, 0u));
 
   bool PreferCopyEngine = (SrcBuffer->OnHost || DstBuffer->OnHost);
 
@@ -884,7 +884,7 @@ ur_result_t urCommandBufferAppendMemBufferWriteExp(
 
   char *ZeHandleDst = nullptr;
   UR_CALL(Buffer->getZeHandle(ZeHandleDst, ur_mem_handle_t_::write_only,
-                              CommandBuffer->Device));
+                              CommandBuffer->Device, nullptr, 0u));
   // Always prefer copy engine for writes
   bool PreferCopyEngine = true;
 
@@ -908,7 +908,7 @@ ur_result_t urCommandBufferAppendMemBufferWriteRectExp(
 
   char *ZeHandleDst = nullptr;
   UR_CALL(Buffer->getZeHandle(ZeHandleDst, ur_mem_handle_t_::write_only,
-                              CommandBuffer->Device));
+                              CommandBuffer->Device, nullptr, 0u));
 
   // Always prefer copy engine for writes
   bool PreferCopyEngine = true;
@@ -930,7 +930,7 @@ ur_result_t urCommandBufferAppendMemBufferReadExp(
 
   char *ZeHandleSrc = nullptr;
   UR_CALL(Buffer->getZeHandle(ZeHandleSrc, ur_mem_handle_t_::read_only,
-                              CommandBuffer->Device));
+                              CommandBuffer->Device, nullptr, 0u));
 
   // Always prefer copy engine for reads
   bool PreferCopyEngine = true;
@@ -953,7 +953,7 @@ ur_result_t urCommandBufferAppendMemBufferReadRectExp(
 
   char *ZeHandleSrc;
   UR_CALL(Buffer->getZeHandle(ZeHandleSrc, ur_mem_handle_t_::read_only,
-                              CommandBuffer->Device));
+                              CommandBuffer->Device, nullptr, 0u));
 
   // Always prefer copy engine for reads
   bool PreferCopyEngine = true;
@@ -1078,7 +1078,7 @@ ur_result_t urCommandBufferAppendMemBufferFillExp(
   char *ZeHandleDst = nullptr;
   _ur_buffer *UrBuffer = reinterpret_cast<_ur_buffer *>(Buffer);
   UR_CALL(UrBuffer->getZeHandle(ZeHandleDst, ur_mem_handle_t_::write_only,
-                                CommandBuffer->Device));
+                                CommandBuffer->Device, nullptr, 0u));
 
   return enqueueCommandBufferFillHelper(
       UR_COMMAND_MEM_BUFFER_FILL, CommandBuffer, ZeHandleDst + Offset,
@@ -1512,7 +1512,7 @@ ur_result_t updateKernelCommand(
     char **ZeHandlePtr = nullptr;
     if (NewMemObjArg) {
       UR_CALL(NewMemObjArg->getZeHandlePtr(ZeHandlePtr, UrAccessMode,
-                                           CommandBuffer->Device));
+                                           CommandBuffer->Device, nullptr, 0u));
     }
 
     auto ZeMutableArgDesc =

diff --git a/source/adapters/level_zero/kernel.cpp b/source/adapters/level_zero/kernel.cpp
@@ -113,7 +113,8 @@ ur_result_t urEnqueueKernelLaunch(
     char **ZeHandlePtr = nullptr;
     if (Arg.Value) {
       UR_CALL(Arg.Value->getZeHandlePtr(ZeHandlePtr, Arg.AccessMode,
-                                        Queue->Device));
+                                        Queue->Device, EventWaitList,
+                                        NumEventsInWaitList));
     }
     ZE2UR_CALL(zeKernelSetArgumentValue,
                (ZeKernel, Arg.Index, Arg.Size, ZeHandlePtr));
@@ -273,7 +274,8 @@ ur_result_t urEnqueueCooperativeKernelLaunchExp(
     char **ZeHandlePtr = nullptr;
     if (Arg.Value) {
       UR_CALL(Arg.Value->getZeHandlePtr(ZeHandlePtr, Arg.AccessMode,
-                                        Queue->Device));
+                                        Queue->Device, EventWaitList,
+                                        NumEventsInWaitList));
     }
     ZE2UR_CALL(zeKernelSetArgumentValue,
                (ZeKernel, Arg.Index, Arg.Size, ZeHandlePtr));