[Mlir-commits] [mlir] [mlir][gpu]Add GPUToXeVM lowering pipeline pass. (PR #161216)

[Sang Ik Lee]

================
@@ -0,0 +1,139 @@
+//===- GPUToXeVMPipeline.cpp - Lowering pipeline to XeVM/LLVM -------------===//
+//
+// Part of the LLVM Project, 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a pass for testing the lowering to XeVM as a generally
+// usable sink pass. If XeGPU ops are used, it expects the MLIR code to have
+// XeGPU ops already embedded in gpu code.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Conversion/AffineToStandard/AffineToStandard.h"
+#include "mlir/Conversion/Passes.h"
+#include "mlir/Conversion/SCFToControlFlow/SCFToControlFlow.h"
+#include "mlir/Conversion/VectorToSCF/VectorToSCF.h"
+#include "mlir/Conversion/XeGPUToXeVM/XeGPUToXeVM.h"
+#include "mlir/Conversion/XeVMToLLVM/XeVMToLLVM.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/GPU/IR/GPUDialect.h"
+#include "mlir/Dialect/GPU/Pipelines/Passes.h"
+#include "mlir/Dialect/GPU/Transforms/Passes.h"
+#include "mlir/Dialect/LLVMIR/Transforms/RequestCWrappers.h"
+#include "mlir/Dialect/MemRef/Transforms/Passes.h"
+#include "mlir/Dialect/XeGPU/Transforms/Passes.h"
+#include "mlir/Pass/PassManager.h"
+#include "mlir/Pass/PassOptions.h"
+#include "mlir/Target/LLVM/XeVM/Target.h"
+#include "mlir/Transforms/Passes.h"
+
+using namespace mlir;
+
+namespace {
+//===----------------------------------------------------------------------===//
+// Common pipeline
+//===----------------------------------------------------------------------===//
+void buildCommonPassPipeline(
+    OpPassManager &pm, const mlir::gpu::GPUToXeVMPipelineOptions &options) {
+  // builtin.module scope passes
+  pm.addPass(createCSEPass());
+  {
+    GpuXeVMAttachTargetOptions xevmTargetOptions;
+    xevmTargetOptions.moduleMatcher = options.xevmModuleMatcher;
+    xevmTargetOptions.triple = options.zebinTriple;
+    xevmTargetOptions.chip = options.zebinChip;
+    xevmTargetOptions.optLevel = options.optLevel;
+    xevmTargetOptions.cmdOptions = options.cmdOptions;
+    pm.addPass(createGpuXeVMAttachTarget(xevmTargetOptions));
+  }
+}
+
+//===----------------------------------------------------------------------===//
+// GPUModule-specific stuff.
+//===----------------------------------------------------------------------===//
+void buildGpuPassPipeline(OpPassManager &pm,
+                          const mlir::gpu::GPUToXeVMPipelineOptions &options) {
+  if (options.xegpuOpLevel == "workgroup") {
+    pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUWgToSgDistribute());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(createLowerAffinePass());
+    pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUBlocking());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCanonicalizerPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
+  }
+  if (options.xegpuOpLevel == "subgroup" ||
+      options.xegpuOpLevel == "workgroup") {
+    pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUPropagateLayout());
+    pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUSubgroupDistribute());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCanonicalizerPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(createLoopInvariantCodeMotionPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
+    pm.addNestedPass<gpu::GPUModuleOp>(xegpu::createXeGPUVectorLinearize());
+  }
+  pm.addNestedPass<gpu::GPUModuleOp>(createConvertXeGPUToXeVMPass());
+  {
+    ConvertGpuOpsToLLVMSPVOpsOptions gpuToLLVMSPVOptions;
+    gpuToLLVMSPVOptions.use64bitIndex = options.use64bitIndex;
+    pm.addNestedPass<gpu::GPUModuleOp>(
+        createConvertGpuOpsToLLVMSPVOps(gpuToLLVMSPVOptions));
+  }
+  pm.addNestedPass<gpu::GPUModuleOp>(createConvertXeVMToLLVMPass());
+  pm.addNestedPass<gpu::GPUModuleOp>(createCSEPass());
+}
+
+//===----------------------------------------------------------------------===//
+// Host Post-GPU pipeline
+//===----------------------------------------------------------------------===//
+void buildHostPostPipeline(OpPassManager &pm,
+                           const mlir::gpu::GPUToXeVMPipelineOptions &options) {
+  pm.addNestedPass<func::FuncOp>(LLVM::createLLVMRequestCWrappersPass());
+  pm.addNestedPass<func::FuncOp>(createGpuAsyncRegionPass());
+  pm.addPass(createReconcileUnrealizedCastsPass());
+  pm.addPass(createConvertVectorToSCFPass());
+  pm.addPass(createSCFToControlFlowPass());
+  pm.addPass(memref::createExpandStridedMetadataPass());
+  pm.addPass(createFinalizeMemRefToLLVMConversionPass());
+  {
+    GpuToLLVMConversionPassOptions gpuToLLVMOptions;
+    gpuToLLVMOptions.hostBarePtrCallConv = options.hostBarePtrCallConv;
+    gpuToLLVMOptions.kernelBarePtrCallConv = options.kernelBarePtrCallConv;
+    pm.addPass(createGpuToLLVMConversionPass(gpuToLLVMOptions));
+  }
+  pm.addPass(createConvertToLLVMPass());
+  pm.addPass(createLowerAffinePass());
----------------
silee2 wrote:

Shouldn't this be called earlier? This pass will lower affine dialect into scf and standard dialects.
Calling it after createConvertToLLVMPass() does not look correct.
BTW, do we need to support `affine` in this pipeline?
Pipeline for nvvm doesn't seem to support it.

https://github.com/llvm/llvm-project/pull/161216