Merge remote-tracking branch 'origin/main' into zhangyan/fix_perf

Author: ZhangYan

cmake/imex-version.txt

@@ -1 +1 @@
-6c2e414a953b9a118bce6adac21cf9d42630e674
+20f2eef4f6c10fcbd68d358591c8b3ef4d1b97d2

include/gc/ExecutionEngine/GPURuntime/GpuOclRuntime.h

@@ -0,0 +1,29 @@
+//===-- GpuOclRuntime.h - GPU OpenCL runtime --------------------*- C++ -*-===//
+//
+// This file is licensed under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef GC_GPUOCLRUNTIME_H
+#define GC_GPUOCLRUNTIME_H
+
+namespace mlir::gc::gpu {
+constexpr char GPU_OCL_MALLOC[] = "gcGpuOclMalloc";
+constexpr char GPU_OCL_DEALLOC[] = "gcGpuOclDealloc";
+constexpr char GPU_OCL_MEMCPY[] = "gcGpuOclMemcpy";
+constexpr char GPU_OCL_KERNEL_CREATE[] = "gcGpuOclKernelCreate";
+constexpr char GPU_OCL_KERNEL_DESTROY[] = "gcGpuOclKernelDestroy";
+constexpr char GPU_OCL_KERNEL_LAUNCH[] = "gcGpuOclKernelLaunch";
+constexpr char GPU_OCL_MOD_DESTRUCTOR[] = "gcGpuOclModuleDestructor";
+} // namespace mlir::gc::gpu
+
+#ifndef GC_GPU_OCL_CONST_ONLY
+
+// TBD
+
+#else
+#undef GC_GPU_OCL_CONST_ONLY
+#endif
+#endif

include/gc/Transforms/Passes.h

@@ -20,7 +20,6 @@ namespace func {
 class FuncOp;
 } // namespace func
 
-
 namespace LLVM {
 class LLVMDialect;
 }
@@ -116,7 +115,8 @@ void populateFrontendPasses(mlir::OpPassManager &);
 void populateCPUPipeline(mlir::OpPassManager &);
 
 #ifdef GC_USE_IMEX
-void populateGPUPipeline(mlir::OpPassManager &);
+struct GPUPipelineOption;
+void populateGPUPipeline(mlir::OpPassManager &, const GPUPipelineOption &);
 #endif
 
 #define GEN_PASS_DECL

include/gc/Transforms/Passes.td

@@ -93,6 +93,20 @@ def LinalgToXeGPU : Pass<"linalg-to-xegpu", "func::FuncOp"> {
                "DPAS register block sizes MxNxK">,
   ];
 }
+
+def AddContextArg : Pass<"add-ctx-arg", "func::FuncOp"> {
+  let summary = "Add a context argument.";
+  let description = [{
+    Add a new memref argument to the function, that could be used to pass some context.
+  }];
+}
+
+def GpuToGpuOcl : Pass<"gpu-to-gpuocl", "ModuleOp"> {
+  let summary = "Convert the GPU operations to GpuOclRuntime calls.";
+  let description = [{
+    Convert the gpu alloc, dealloc, memcpy and launch operations to GpuOclRuntime calls.
+  }];
+}
 #endif // GC_USE_IMEX
 
 def IterativeTilingAndFusion : Pass<"iterative-tiling-and-fusion",

lib/gc/Transforms/GPU/AddContextArg.cpp

@@ -0,0 +1,54 @@
+//===-- AddContextArg.cpp - Add context argument ----------------*- C++ -*-===//
+//
+// This file is licensed under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+#include "mlir/Conversion/Passes.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+
+namespace mlir::gc {
+#define GEN_PASS_DECL_ADDCONTEXTARG
+#define GEN_PASS_DEF_ADDCONTEXTARG
+#include "gc/Transforms/Passes.h.inc"
+} // namespace mlir::gc
+
+using namespace mlir;
+
+namespace {
+struct AddContextArg final : gc::impl::AddContextArgBase<AddContextArg> {
+  void runOnOperation() override {
+    auto func = getOperation();
+    if (func.isExternal()) {
+      return;
+    }
+
+    auto funcType = func.getFunctionType();
+    auto argTypes = llvm::to_vector<8>(funcType.getInputs());
+    auto resultTypes = llvm::to_vector<1>(funcType.getResults());
+    auto ctx = func->getContext();
+    auto newArgType = MemRefType::get({}, IntegerType::get(ctx, 8));
+    argTypes.emplace_back(newArgType);
+    auto newFuncType = FunctionType::get(ctx, argTypes, resultTypes);
+    func.setType(newFuncType);
+    func.getBody().front().addArgument(newArgType, func.getLoc());
+
+    // Find all function calls and append the last argument of the current
+    // function to the call.
+    auto module = func->getParentOfType<ModuleOp>();
+    func.walk([&](func::CallOp call) {
+      // If the function to be called is defined in the current module, then the
+      // context arg will be added to this function signature either and, thus,
+      // wee need add the context arg to the function call.
+      if (auto callee = module.lookupSymbol<func::FuncOp>(call.getCallee());
+          !callee || callee.isExternal()) {
+        return;
+      }
+      auto args = llvm::to_vector<8>(call.getOperands());
+      args.emplace_back(func.getArgument(func.getNumArguments() - 1));
+      call->setOperands(args);
+    });
+  }
+};
+} // namespace

lib/gc/Transforms/GPU/CMakeLists.txt

@@ -1,4 +1,6 @@
 gc_add_mlir_library(GcGpuPasses
+  AddContextArg.cpp
+  GpuToGpuOcl.cpp
   LinalgToXeGPU.cpp
   Pipeline.cpp