1//===- GPUToNVVMPipeline.cpp - Test lowering to NVVM as a sink pass -------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements a pass for testing the lowering to NVVM as a generally
10// usable sink pass.
11//
12//===----------------------------------------------------------------------===//
13
14#include "mlir/Conversion/AffineToStandard/AffineToStandard.h"
15#include "mlir/Conversion/ArithToLLVM/ArithToLLVM.h"
16#include "mlir/Conversion/FuncToLLVM/ConvertFuncToLLVMPass.h"
17#include "mlir/Conversion/GPUCommon/GPUCommonPass.h"
18#include "mlir/Conversion/GPUToNVVM/GPUToNVVMPass.h"
19#include "mlir/Conversion/IndexToLLVM/IndexToLLVM.h"
20#include "mlir/Conversion/MathToLLVM/MathToLLVM.h"
21#include "mlir/Conversion/NVGPUToNVVM/NVGPUToNVVM.h"
22#include "mlir/Conversion/NVVMToLLVM/NVVMToLLVM.h"
23#include "mlir/Conversion/ReconcileUnrealizedCasts/ReconcileUnrealizedCasts.h"
24#include "mlir/Conversion/SCFToControlFlow/SCFToControlFlow.h"
25#include "mlir/Conversion/VectorToSCF/VectorToSCF.h"
26#include "mlir/Dialect/GPU/IR/GPUDialect.h"
27#include "mlir/Dialect/GPU/Pipelines/Passes.h"
28#include "mlir/Dialect/GPU/Transforms/NVVM/Passes.h"
29#include "mlir/Dialect/GPU/Transforms/Passes.h"
30#include "mlir/Dialect/MemRef/Transforms/Passes.h"
31#include "mlir/Pass/PassManager.h"
32#include "mlir/Pass/PassOptions.h"
33#include "mlir/Transforms/Passes.h"
34
35using namespace mlir;
36
37namespace {
38
39//===----------------------------------------------------------------------===//
40// Common pipeline
41//===----------------------------------------------------------------------===//
42void buildCommonPassPipeline(
43 OpPassManager &pm, const mlir::gpu::GPUToNVVMPipelineOptions &options) {
44 pm.addPass(pass: createConvertNVGPUToNVVMPass());
45 pm.addPass(pass: createGpuKernelOutliningPass());
46 pm.addPass(pass: createConvertVectorToSCFPass());
47 pm.addPass(pass: createSCFToControlFlowPass());
48 pm.addPass(pass: createConvertNVVMToLLVMPass());
49 pm.addPass(pass: createConvertFuncToLLVMPass());
50 pm.addPass(pass: memref::createExpandStridedMetadataPass());
51
52 GpuNVVMAttachTargetOptions nvvmTargetOptions;
53 nvvmTargetOptions.triple = options.cubinTriple;
54 nvvmTargetOptions.chip = options.cubinChip;
55 nvvmTargetOptions.features = options.cubinFeatures;
56 nvvmTargetOptions.optLevel = options.optLevel;
57 nvvmTargetOptions.cmdOptions = options.cmdOptions;
58 pm.addPass(pass: createGpuNVVMAttachTarget(options: nvvmTargetOptions));
59 pm.addPass(pass: createLowerAffinePass());
60 pm.addPass(pass: createArithToLLVMConversionPass());
61 ConvertIndexToLLVMPassOptions convertIndexToLLVMPassOpt;
62 convertIndexToLLVMPassOpt.indexBitwidth = options.indexBitWidth;
63 pm.addPass(pass: createConvertIndexToLLVMPass(options: convertIndexToLLVMPassOpt));
64 pm.addPass(pass: createCanonicalizerPass());
65 pm.addPass(pass: createCSEPass());
66}
67
68//===----------------------------------------------------------------------===//
69// GPUModule-specific stuff.
70//===----------------------------------------------------------------------===//
71void buildGpuPassPipeline(OpPassManager &pm,
72 const mlir::gpu::GPUToNVVMPipelineOptions &options) {
73 ConvertGpuOpsToNVVMOpsOptions opt;
74 opt.useBarePtrCallConv = options.kernelUseBarePtrCallConv;
75 opt.indexBitwidth = options.indexBitWidth;
76 opt.allowPatternRollback = options.allowPatternRollback;
77 pm.addNestedPass<gpu::GPUModuleOp>(pass: createConvertGpuOpsToNVVMOps(options: opt));
78 pm.addNestedPass<gpu::GPUModuleOp>(pass: createCanonicalizerPass());
79 pm.addNestedPass<gpu::GPUModuleOp>(pass: createCSEPass());
80 pm.addNestedPass<gpu::GPUModuleOp>(pass: createReconcileUnrealizedCastsPass());
81}
82
83//===----------------------------------------------------------------------===//
84// Host Post-GPU pipeline
85//===----------------------------------------------------------------------===//
86void buildHostPostPipeline(OpPassManager &pm,
87 const mlir::gpu::GPUToNVVMPipelineOptions &options) {
88 GpuToLLVMConversionPassOptions opt;
89 opt.hostBarePtrCallConv = options.hostUseBarePtrCallConv;
90 opt.kernelBarePtrCallConv = options.kernelUseBarePtrCallConv;
91 pm.addPass(pass: createGpuToLLVMConversionPass(options: opt));
92
93 GpuModuleToBinaryPassOptions gpuModuleToBinaryPassOptions;
94 gpuModuleToBinaryPassOptions.compilationTarget = options.cubinFormat;
95 pm.addPass(pass: createGpuModuleToBinaryPass(options: gpuModuleToBinaryPassOptions));
96 pm.addPass(pass: createConvertMathToLLVMPass());
97 pm.addPass(pass: createCanonicalizerPass());
98 pm.addPass(pass: createCSEPass());
99 pm.addPass(pass: createReconcileUnrealizedCastsPass());
100}
101
102} // namespace
103
104void mlir::gpu::buildLowerToNVVMPassPipeline(
105 OpPassManager &pm, const GPUToNVVMPipelineOptions &options) {
106 // Common pipelines
107 buildCommonPassPipeline(pm, options);
108
109 // GPUModule-specific stuff
110 buildGpuPassPipeline(pm, options);
111
112 // Host post-GPUModule-specific stuff
113 buildHostPostPipeline(pm, options);
114}
115
116void mlir::gpu::registerGPUToNVVMPipeline() {
117 PassPipelineRegistration<GPUToNVVMPipelineOptions>(
118 "gpu-lower-to-nvvm-pipeline",
119 "The default pipeline lowers main dialects (arith, memref, scf, "
120 "vector, gpu, and nvgpu) to NVVM. It starts by lowering GPU code to the "
121 "specified compilation target (default is fatbin) then lowers the host "
122 "code.",
123 buildLowerToNVVMPassPipeline);
124}
125