| 1 | //===- SparseTensorPipelines.cpp - Pipelines for sparse tensor code -------===// |
| 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 | #include "mlir/Conversion/Passes.h" |
| 10 | #include "mlir/Dialect/Arith/Transforms/Passes.h" |
| 11 | #include "mlir/Dialect/Bufferization/Transforms/Passes.h" |
| 12 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 13 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 14 | #include "mlir/Dialect/GPU/Transforms/Passes.h" |
| 15 | #include "mlir/Dialect/Linalg/Passes.h" |
| 16 | #include "mlir/Dialect/MemRef/Transforms/Passes.h" |
| 17 | #include "mlir/Dialect/SparseTensor/Pipelines/Passes.h" |
| 18 | #include "mlir/Dialect/SparseTensor/Transforms/Passes.h" |
| 19 | #include "mlir/Pass/PassManager.h" |
| 20 | #include "mlir/Transforms/Passes.h" |
| 21 | |
| 22 | //===----------------------------------------------------------------------===// |
| 23 | // Pipeline implementation. |
| 24 | //===----------------------------------------------------------------------===// |
| 25 | |
| 26 | void mlir::sparse_tensor::buildSparsifier(OpPassManager &pm, |
| 27 | const SparsifierOptions &options) { |
| 28 | // Rewrite named linalg ops into generic ops and apply fusion. |
| 29 | pm.addNestedPass<func::FuncOp>(createLinalgGeneralizeNamedOpsPass()); |
| 30 | pm.addNestedPass<func::FuncOp>(createLinalgElementwiseOpFusionPass()); |
| 31 | |
| 32 | // Sparsification and bufferization mini-pipeline. |
| 33 | pm.addPass(createSparsificationAndBufferizationPass( |
| 34 | getBufferizationOptionsForSparsification( |
| 35 | options.testBufferizationAnalysisOnly), |
| 36 | options.sparsificationOptions(), options.createSparseDeallocs, |
| 37 | options.enableRuntimeLibrary, options.enableBufferInitialization, |
| 38 | options.vectorLength, |
| 39 | /*enableVLAVectorization=*/options.armSVE, |
| 40 | /*enableSIMDIndex32=*/options.force32BitVectorIndices, |
| 41 | options.enableGPULibgen, |
| 42 | options.sparsificationOptions().sparseEmitStrategy, |
| 43 | options.sparsificationOptions().parallelizationStrategy)); |
| 44 | |
| 45 | // Bail-early for test setup. |
| 46 | if (options.testBufferizationAnalysisOnly) |
| 47 | return; |
| 48 | |
| 49 | // Storage specifier lowering and bufferization wrap-up. |
| 50 | pm.addPass(createStorageSpecifierToLLVMPass()); |
| 51 | pm.addNestedPass<func::FuncOp>(pass: createCanonicalizerPass()); |
| 52 | |
| 53 | // GPU code generation. |
| 54 | const bool gpuCodegen = options.gpuTriple.hasValue(); |
| 55 | if (gpuCodegen) { |
| 56 | pm.addPass(createSparseGPUCodegenPass()); |
| 57 | pm.addNestedPass<gpu::GPUModuleOp>(createStripDebugInfoPass()); |
| 58 | pm.addNestedPass<gpu::GPUModuleOp>(createSCFToControlFlowPass()); |
| 59 | pm.addNestedPass<gpu::GPUModuleOp>(createConvertGpuOpsToNVVMOps()); |
| 60 | } |
| 61 | |
| 62 | // Progressively lower to LLVM. Note that the convert-vector-to-llvm |
| 63 | // pass is repeated on purpose. |
| 64 | // TODO(springerm): Add sparse support to the BufferDeallocation pass and add |
| 65 | // it to this pipeline. |
| 66 | pm.addNestedPass<func::FuncOp>(createConvertLinalgToLoopsPass()); |
| 67 | pm.addNestedPass<func::FuncOp>(createConvertVectorToSCFPass()); |
| 68 | pm.addNestedPass<func::FuncOp>(memref::createExpandReallocPass()); |
| 69 | pm.addNestedPass<func::FuncOp>(createSCFToControlFlowPass()); |
| 70 | pm.addPass(memref::createExpandStridedMetadataPass()); |
| 71 | pm.addPass(pass: createLowerAffinePass()); |
| 72 | pm.addPass( |
| 73 | pass: createConvertVectorToLLVMPass(options.convertVectorToLLVMOptions())); |
| 74 | pm.addPass(pass: createFinalizeMemRefToLLVMConversionPass()); |
| 75 | pm.addNestedPass<func::FuncOp>(createConvertComplexToStandardPass()); |
| 76 | pm.addNestedPass<func::FuncOp>(arith::createArithExpandOpsPass()); |
| 77 | pm.addNestedPass<func::FuncOp>(createConvertMathToLLVMPass()); |
| 78 | pm.addPass(pass: createConvertMathToLibmPass()); |
| 79 | pm.addPass(pass: createConvertComplexToLibm()); |
| 80 | pm.addPass( |
| 81 | pass: createConvertVectorToLLVMPass(options.convertVectorToLLVMOptions())); |
| 82 | pm.addPass(pass: createConvertComplexToLLVMPass()); |
| 83 | pm.addPass( |
| 84 | pass: createConvertVectorToLLVMPass(options.convertVectorToLLVMOptions())); |
| 85 | pm.addPass(pass: createConvertFuncToLLVMPass()); |
| 86 | pm.addPass(pass: createArithToLLVMConversionPass()); |
| 87 | pm.addPass(pass: createConvertControlFlowToLLVMPass()); |
| 88 | |
| 89 | // Finalize GPU code generation. |
| 90 | if (gpuCodegen) { |
| 91 | GpuNVVMAttachTargetOptions nvvmTargetOptions; |
| 92 | nvvmTargetOptions.triple = options.gpuTriple; |
| 93 | nvvmTargetOptions.chip = options.gpuChip; |
| 94 | nvvmTargetOptions.features = options.gpuFeatures; |
| 95 | pm.addPass(pass: createGpuNVVMAttachTarget(nvvmTargetOptions)); |
| 96 | pm.addPass(pass: createGpuToLLVMConversionPass()); |
| 97 | GpuModuleToBinaryPassOptions gpuModuleToBinaryPassOptions; |
| 98 | gpuModuleToBinaryPassOptions.compilationTarget = options.gpuFormat; |
| 99 | pm.addPass(pass: createGpuModuleToBinaryPass(gpuModuleToBinaryPassOptions)); |
| 100 | } |
| 101 | |
| 102 | // Convert poison values. |
| 103 | pm.addPass(pass: createUBToLLVMConversionPass()); |
| 104 | |
| 105 | // Ensure all casts are realized. |
| 106 | pm.addPass(pass: createReconcileUnrealizedCastsPass()); |
| 107 | } |
| 108 | |
| 109 | //===----------------------------------------------------------------------===// |
| 110 | // Pipeline registration. |
| 111 | //===----------------------------------------------------------------------===// |
| 112 | |
| 113 | void mlir::sparse_tensor::registerSparseTensorPipelines() { |
| 114 | PassPipelineRegistration<SparsifierOptions>( |
| 115 | "sparsifier" , |
| 116 | "The standard pipeline for taking sparsity-agnostic IR using the" |
| 117 | " sparse-tensor type, and lowering it to LLVM IR with concrete" |
| 118 | " representations and algorithms for sparse tensors." , |
| 119 | buildSparsifier); |
| 120 | } |
| 121 | |