1 | //===- GPUToSPIRVPass.cpp - GPU to SPIR-V Passes --------------------------===// |
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 to convert a kernel function in the GPU Dialect |
10 | // into a spirv.module operation. |
11 | // |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "mlir/Conversion/GPUToSPIRV/GPUToSPIRVPass.h" |
15 | |
16 | #include "mlir/Conversion/ArithToSPIRV/ArithToSPIRV.h" |
17 | #include "mlir/Conversion/FuncToSPIRV/FuncToSPIRV.h" |
18 | #include "mlir/Conversion/GPUToSPIRV/GPUToSPIRV.h" |
19 | #include "mlir/Conversion/MemRefToSPIRV/MemRefToSPIRV.h" |
20 | #include "mlir/Conversion/SCFToSPIRV/SCFToSPIRV.h" |
21 | #include "mlir/Conversion/VectorToSPIRV/VectorToSPIRV.h" |
22 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
23 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
24 | #include "mlir/Dialect/SPIRV/IR/SPIRVDialect.h" |
25 | #include "mlir/Dialect/SPIRV/IR/SPIRVOps.h" |
26 | #include "mlir/Dialect/SPIRV/Transforms/SPIRVConversion.h" |
27 | #include "mlir/IR/PatternMatch.h" |
28 | |
29 | namespace mlir { |
30 | #define GEN_PASS_DEF_CONVERTGPUTOSPIRV |
31 | #include "mlir/Conversion/Passes.h.inc" |
32 | } // namespace mlir |
33 | |
34 | using namespace mlir; |
35 | |
36 | namespace { |
37 | /// Pass to lower GPU Dialect to SPIR-V. The pass only converts the gpu.func ops |
38 | /// inside gpu.module ops. i.e., the function that are referenced in |
39 | /// gpu.launch_func ops. For each such function |
40 | /// |
41 | /// 1) Create a spirv::ModuleOp, and clone the function into spirv::ModuleOp |
42 | /// (the original function is still needed by the gpu::LaunchKernelOp, so cannot |
43 | /// replace it). |
44 | /// |
45 | /// 2) Lower the body of the spirv::ModuleOp. |
46 | struct GPUToSPIRVPass final : impl::ConvertGPUToSPIRVBase<GPUToSPIRVPass> { |
47 | explicit GPUToSPIRVPass(bool mapMemorySpace) |
48 | : mapMemorySpace(mapMemorySpace) {} |
49 | void runOnOperation() override; |
50 | |
51 | private: |
52 | bool mapMemorySpace; |
53 | }; |
54 | |
55 | void GPUToSPIRVPass::runOnOperation() { |
56 | MLIRContext *context = &getContext(); |
57 | ModuleOp module = getOperation(); |
58 | |
59 | SmallVector<Operation *, 1> gpuModules; |
60 | OpBuilder builder(context); |
61 | |
62 | auto targetEnvSupportsKernelCapability = [](gpu::GPUModuleOp moduleOp) { |
63 | Operation *gpuModule = moduleOp.getOperation(); |
64 | auto targetAttr = spirv::lookupTargetEnvOrDefault(op: gpuModule); |
65 | spirv::TargetEnv targetEnv(targetAttr); |
66 | return targetEnv.allows(spirv::Capability::Kernel); |
67 | }; |
68 | |
69 | module.walk([&](gpu::GPUModuleOp moduleOp) { |
70 | // Clone each GPU kernel module for conversion, given that the GPU |
71 | // launch op still needs the original GPU kernel module. |
72 | // For Vulkan Shader capabilities, we insert the newly converted SPIR-V |
73 | // module right after the original GPU module, as that's the expectation of |
74 | // the in-tree SPIR-V CPU runner (the Vulkan runner does not use this pass). |
75 | // For OpenCL Kernel capabilities, we insert the newly converted SPIR-V |
76 | // module inside the original GPU module, as that's the expectaion of the |
77 | // normal GPU compilation pipeline. |
78 | if (targetEnvSupportsKernelCapability(moduleOp)) { |
79 | builder.setInsertionPointToStart(moduleOp.getBody()); |
80 | } else { |
81 | builder.setInsertionPoint(moduleOp.getOperation()); |
82 | } |
83 | gpuModules.push_back(builder.clone(*moduleOp.getOperation())); |
84 | }); |
85 | |
86 | // Run conversion for each module independently as they can have different |
87 | // TargetEnv attributes. |
88 | for (Operation *gpuModule : gpuModules) { |
89 | spirv::TargetEnvAttr targetAttr = |
90 | spirv::lookupTargetEnvOrDefault(gpuModule); |
91 | |
92 | // Map MemRef memory space to SPIR-V storage class first if requested. |
93 | if (mapMemorySpace) { |
94 | spirv::MemorySpaceToStorageClassMap memorySpaceMap = |
95 | targetEnvSupportsKernelCapability( |
96 | dyn_cast<gpu::GPUModuleOp>(gpuModule)) |
97 | ? spirv::mapMemorySpaceToOpenCLStorageClass |
98 | : spirv::mapMemorySpaceToVulkanStorageClass; |
99 | spirv::MemorySpaceToStorageClassConverter converter(memorySpaceMap); |
100 | spirv::convertMemRefTypesAndAttrs(gpuModule, converter); |
101 | |
102 | // Check if there are any illegal ops remaining. |
103 | std::unique_ptr<ConversionTarget> target = |
104 | spirv::getMemorySpaceToStorageClassTarget(*context); |
105 | gpuModule->walk([&target, this](Operation *childOp) { |
106 | if (target->isIllegal(childOp)) { |
107 | childOp->emitOpError("failed to legalize memory space" ); |
108 | signalPassFailure(); |
109 | return WalkResult::interrupt(); |
110 | } |
111 | return WalkResult::advance(); |
112 | }); |
113 | } |
114 | |
115 | std::unique_ptr<ConversionTarget> target = |
116 | SPIRVConversionTarget::get(targetAttr); |
117 | |
118 | SPIRVConversionOptions options; |
119 | options.use64bitIndex = this->use64bitIndex; |
120 | SPIRVTypeConverter typeConverter(targetAttr, options); |
121 | populateMMAToSPIRVCoopMatrixTypeConversion(typeConverter); |
122 | |
123 | RewritePatternSet patterns(context); |
124 | populateGPUToSPIRVPatterns(typeConverter, patterns); |
125 | populateGpuWMMAToSPIRVCoopMatrixKHRConversionPatterns(typeConverter, |
126 | patterns); |
127 | |
128 | // TODO: Change SPIR-V conversion to be progressive and remove the following |
129 | // patterns. |
130 | ScfToSPIRVContext scfContext; |
131 | populateSCFToSPIRVPatterns(typeConverter, scfContext, patterns); |
132 | mlir::arith::populateArithToSPIRVPatterns(typeConverter, patterns); |
133 | populateMemRefToSPIRVPatterns(typeConverter, patterns); |
134 | populateFuncToSPIRVPatterns(typeConverter, patterns); |
135 | populateVectorToSPIRVPatterns(typeConverter, patterns); |
136 | |
137 | if (failed(applyFullConversion(gpuModule, *target, std::move(patterns)))) |
138 | return signalPassFailure(); |
139 | } |
140 | |
141 | // For OpenCL, the gpu.func op in the original gpu.module op needs to be |
142 | // replaced with an empty func.func op with the same arguments as the gpu.func |
143 | // op. The func.func op needs gpu.kernel attribute set. |
144 | module.walk([&](gpu::GPUModuleOp moduleOp) { |
145 | if (targetEnvSupportsKernelCapability(moduleOp)) { |
146 | moduleOp.walk([&](gpu::GPUFuncOp funcOp) { |
147 | builder.setInsertionPoint(funcOp); |
148 | auto newFuncOp = builder.create<func::FuncOp>( |
149 | funcOp.getLoc(), funcOp.getName(), funcOp.getFunctionType()); |
150 | auto entryBlock = newFuncOp.addEntryBlock(); |
151 | builder.setInsertionPointToEnd(entryBlock); |
152 | builder.create<func::ReturnOp>(funcOp.getLoc()); |
153 | newFuncOp->setAttr(gpu::GPUDialect::getKernelFuncAttrName(), |
154 | builder.getUnitAttr()); |
155 | funcOp.erase(); |
156 | }); |
157 | } |
158 | }); |
159 | } |
160 | |
161 | } // namespace |
162 | |
163 | std::unique_ptr<OperationPass<ModuleOp>> |
164 | mlir::createConvertGPUToSPIRVPass(bool mapMemorySpace) { |
165 | return std::make_unique<GPUToSPIRVPass>(args&: mapMemorySpace); |
166 | } |
167 | |