| 1 | //===- NamedOpConversions.cpp - Implements conversions between named ops --===// |
| 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 conversions between named ops that can be seens as |
| 10 | // canonicalizations of named ops. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "mlir/Dialect/Linalg/Passes.h" |
| 15 | |
| 16 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
| 17 | #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
| 18 | #include "mlir/IR/PatternMatch.h" |
| 19 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| 20 | #include "llvm/ADT/SmallVector.h" |
| 21 | #include "llvm/ADT/TypeSwitch.h" |
| 22 | |
| 23 | namespace mlir { |
| 24 | #define GEN_PASS_DEF_LINALGNAMEDOPCONVERSIONPASS |
| 25 | #include "mlir/Dialect/Linalg/Passes.h.inc" |
| 26 | } // namespace mlir |
| 27 | |
| 28 | using namespace mlir; |
| 29 | using namespace mlir::linalg; |
| 30 | |
| 31 | static llvm::SmallVector<int64_t> getIndicesVector(int start, int end) { |
| 32 | return llvm::to_vector<2>(llvm::seq<int64_t>(start, end)); |
| 33 | } |
| 34 | |
| 35 | static LogicalResult |
| 36 | matchAndReplaceDepthwiseConv(Operation *operation, Value input, Value kernel, |
| 37 | Value iZp, Value kZp, Value init, Attribute stride, |
| 38 | Attribute dilation, PatternRewriter &rewriter) { |
| 39 | Location loc = operation->getLoc(); |
| 40 | auto linalgOp = dyn_cast<LinalgOp>(operation); |
| 41 | // Exit out on the memref version of this operation. |
| 42 | if (!linalgOp || !linalgOp.hasPureTensorSemantics()) |
| 43 | return failure(); |
| 44 | |
| 45 | auto result = operation->getResult(idx: 0); |
| 46 | |
| 47 | auto kernelTy = dyn_cast<RankedTensorType>(kernel.getType()); |
| 48 | auto initTy = dyn_cast<RankedTensorType>(init.getType()); |
| 49 | auto resultTy = dyn_cast<RankedTensorType>(result.getType()); |
| 50 | if (!kernelTy || !initTy || !resultTy) |
| 51 | return failure(); |
| 52 | |
| 53 | if (kernelTy.getDimSize(3) != 1) |
| 54 | return failure(); |
| 55 | |
| 56 | // Collapse kernel dims. |
| 57 | SmallVector<ReassociationIndices, 4> collapsedKernelDims = { |
| 58 | getIndicesVector(start: 0, end: 1), getIndicesVector(start: 1, end: 2), getIndicesVector(start: 2, end: 4)}; |
| 59 | auto newKernelTy = RankedTensorType::get( |
| 60 | {kernelTy.getDimSize(0), kernelTy.getDimSize(1), kernelTy.getDimSize(2)}, |
| 61 | kernelTy.getElementType()); |
| 62 | auto collapsedKernel = rewriter.create<tensor::CollapseShapeOp>( |
| 63 | loc, newKernelTy, kernel, collapsedKernelDims); |
| 64 | |
| 65 | // Collapse init dims. |
| 66 | SmallVector<ReassociationIndices, 4> collapsedInitDims = { |
| 67 | getIndicesVector(start: 0, end: 1), getIndicesVector(start: 1, end: 2), getIndicesVector(start: 2, end: 3), |
| 68 | getIndicesVector(start: 3, end: 5)}; |
| 69 | auto newInitTy = |
| 70 | RankedTensorType::get({initTy.getDimSize(0), initTy.getDimSize(1), |
| 71 | initTy.getDimSize(2), initTy.getDimSize(3)}, |
| 72 | initTy.getElementType()); |
| 73 | auto collapsedInit = rewriter.create<tensor::CollapseShapeOp>( |
| 74 | loc, newInitTy, init, collapsedInitDims); |
| 75 | |
| 76 | SmallVector<NamedAttribute> preservedAttrs; |
| 77 | Operation *newConv = |
| 78 | TypeSwitch<Operation *, Operation *>(operation) |
| 79 | .Case<DepthwiseConv2DNhwcHwcmOp>([&](auto op) { |
| 80 | preservedAttrs = getPrunedAttributeList(op); |
| 81 | return rewriter.create<DepthwiseConv2DNhwcHwcOp>( |
| 82 | loc, newInitTy, ValueRange{input, collapsedKernel}, |
| 83 | ValueRange{collapsedInit}, stride, dilation); |
| 84 | }) |
| 85 | .Case<DepthwiseConv2DNhwcHwcmQOp>([&](auto op) { |
| 86 | preservedAttrs = getPrunedAttributeList(op); |
| 87 | return rewriter.create<DepthwiseConv2DNhwcHwcQOp>( |
| 88 | loc, newInitTy, ValueRange{input, collapsedKernel, iZp, kZp}, |
| 89 | ValueRange{collapsedInit}, stride, dilation); |
| 90 | }) |
| 91 | .Default([](Operation *op) { return nullptr; }); |
| 92 | if (!newConv) |
| 93 | return failure(); |
| 94 | for (auto attr : preservedAttrs) |
| 95 | newConv->setAttr(attr.getName(), attr.getValue()); |
| 96 | |
| 97 | // Expand dimensions back out to |
| 98 | rewriter.replaceOpWithNewOp<tensor::ExpandShapeOp>( |
| 99 | operation, resultTy, newConv->getResult(0), collapsedInitDims); |
| 100 | return success(); |
| 101 | } |
| 102 | |
| 103 | namespace { |
| 104 | struct SimplifyDepthwiseConvOp |
| 105 | : public OpRewritePattern<DepthwiseConv2DNhwcHwcmOp> { |
| 106 | using OpRewritePattern<DepthwiseConv2DNhwcHwcmOp>::OpRewritePattern; |
| 107 | |
| 108 | LogicalResult matchAndRewrite(DepthwiseConv2DNhwcHwcmOp op, |
| 109 | PatternRewriter &rewriter) const override { |
| 110 | Operation *operation = op.getOperation(); |
| 111 | Value input = op.getDpsInputOperand(0)->get(); |
| 112 | Value kernel = op.getDpsInputOperand(1)->get(); |
| 113 | Value init = op.getDpsInitOperand(0)->get(); |
| 114 | |
| 115 | auto stride = op.getStrides(); |
| 116 | auto dilation = op.getDilations(); |
| 117 | |
| 118 | return matchAndReplaceDepthwiseConv(operation, input, kernel, nullptr, |
| 119 | nullptr, init, stride, dilation, |
| 120 | rewriter); |
| 121 | } |
| 122 | }; |
| 123 | |
| 124 | struct SimplifyDepthwiseConvQOp |
| 125 | : public OpRewritePattern<DepthwiseConv2DNhwcHwcmQOp> { |
| 126 | using OpRewritePattern<DepthwiseConv2DNhwcHwcmQOp>::OpRewritePattern; |
| 127 | |
| 128 | LogicalResult matchAndRewrite(DepthwiseConv2DNhwcHwcmQOp op, |
| 129 | PatternRewriter &rewriter) const override { |
| 130 | Operation *operation = op.getOperation(); |
| 131 | Value input = op.getDpsInputOperand(0)->get(); |
| 132 | Value kernel = op.getDpsInputOperand(1)->get(); |
| 133 | Value iZp = op.getDpsInputOperand(2)->get(); |
| 134 | Value kZp = op.getDpsInputOperand(3)->get(); |
| 135 | Value init = op.getDpsInitOperand(0)->get(); |
| 136 | |
| 137 | auto stride = op.getStrides(); |
| 138 | auto dilation = op.getDilations(); |
| 139 | |
| 140 | return matchAndReplaceDepthwiseConv(operation, input, kernel, iZp, kZp, |
| 141 | init, stride, dilation, rewriter); |
| 142 | } |
| 143 | }; |
| 144 | |
| 145 | struct LinalgNamedOpConversionPass |
| 146 | : public impl::LinalgNamedOpConversionPassBase< |
| 147 | LinalgNamedOpConversionPass> { |
| 148 | using impl::LinalgNamedOpConversionPassBase< |
| 149 | LinalgNamedOpConversionPass>::LinalgNamedOpConversionPassBase; |
| 150 | |
| 151 | void runOnOperation() override { |
| 152 | Operation *op = getOperation(); |
| 153 | RewritePatternSet patterns(op->getContext()); |
| 154 | populateLinalgNamedOpConversionPatterns(patterns); |
| 155 | if (failed(applyPatternsGreedily(op, std::move(patterns)))) |
| 156 | return signalPassFailure(); |
| 157 | } |
| 158 | }; |
| 159 | } // namespace |
| 160 | |
| 161 | void mlir::linalg::populateLinalgNamedOpConversionPatterns( |
| 162 | RewritePatternSet &patterns) { |
| 163 | patterns.add<SimplifyDepthwiseConvOp, SimplifyDepthwiseConvQOp>( |
| 164 | arg: patterns.getContext()); |
| 165 | } |
| 166 | |