不知道为啥在最新的master分支上删掉了一个必要的函数，先加上然后等两周后同步看看改动再说

2 files changed, 22 insertions, 0 deletions

diff --git a/mlir/include/mlir/Conversion/GPUCommon/GPUCommonPass.h b/mlir/include/mlir/Conversion/GPUCommon/GPUCommonPass.h
index 708a3fe0b23e..42d06f301ad5 100644
--- a/mlir/include/mlir/Conversion/GPUCommon/GPUCommonPass.h
+++ b/mlir/include/mlir/Conversion/GPUCommon/GPUCommonPass.h
@@ -52,6 +52,9 @@ using LoweringCallback = std::function<std::unique_ptr<llvm::Module>(
 /// typed ABI on top of GPU runtimes such as CUDA or ROCm (HIP).
 std::unique_ptr<OperationPass<ModuleOp>> createGpuToLLVMConversionPass();
 
+void populateGpuToLLVMConversionPatterns(LLVMTypeConverter &converter,
+                                         OwningRewritePatternList &patterns);
+
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_GPUCOMMON_GPUCOMMONPASS_H_
diff --git a/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp b/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp
index 81c939875953..e5adf778e098 100644
--- a/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp
+++ b/mlir/lib/Conversion/GPUCommon/GPUToLLVMConversion.cpp
@@ -798,3 +798,22 @@ std::unique_ptr<mlir::OperationPass<mlir::ModuleOp>>
 mlir::createGpuToLLVMConversionPass() {
   return std::make_unique<GpuToLLVMConversionPass>();
 }
+
+void mlir::populateGpuToLLVMConversionPatterns(
+    LLVMTypeConverter &converter, OwningRewritePatternList &patterns) {
+  auto gpuBinaryAnnotation = gpu::getDefaultGpuBinaryAnnotation();
+  converter.addConversion(
+      [context = &converter.getContext()](gpu::AsyncTokenType type) -> Type {
+        return LLVM::LLVMPointerType::get(IntegerType::get(context, 8));
+      });
+  patterns.add<ConvertAllocOpToGpuRuntimeCallPattern,
+               ConvertDeallocOpToGpuRuntimeCallPattern,
+               ConvertHostRegisterOpToGpuRuntimeCallPattern,
+               ConvertMemcpyOpToGpuRuntimeCallPattern,
+               ConvertWaitAsyncOpToGpuRuntimeCallPattern,
+               ConvertWaitOpToGpuRuntimeCallPattern,
+               ConvertAsyncYieldToGpuRuntimeCallPattern>(converter);
+  patterns.add<ConvertLaunchFuncOpToGpuRuntimeCallPattern>(converter,
+                                                           gpuBinaryAnnotation);
+  patterns.add<EraseGpuModuleOpPattern>(&converter.getContext());
+}