| 1 | //===- AffineExpandIndexOpsAsAffine.cpp - Expand index ops to apply 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 to expand affine index ops into one or more more |
| 10 | // fundamental operations. |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "mlir/Dialect/Affine/Transforms/Passes.h" |
| 14 | |
| 15 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 16 | #include "mlir/Dialect/Affine/Transforms/Transforms.h" |
| 17 | #include "mlir/Dialect/Affine/Utils.h" |
| 18 | #include "mlir/Dialect/Arith/Utils/Utils.h" |
| 19 | #include "mlir/Dialect/UB/IR/UBOps.h" |
| 20 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 21 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 22 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| 23 | |
| 24 | namespace mlir { |
| 25 | namespace affine { |
| 26 | #define GEN_PASS_DEF_AFFINEEXPANDINDEXOPSASAFFINE |
| 27 | #include "mlir/Dialect/Affine/Transforms/Passes.h.inc" |
| 28 | } // namespace affine |
| 29 | } // namespace mlir |
| 30 | |
| 31 | using namespace mlir; |
| 32 | using namespace mlir::affine; |
| 33 | |
| 34 | namespace { |
| 35 | /// Lowers `affine.delinearize_index` into a sequence of division and remainder |
| 36 | /// operations via affine.apply. For vector types, unrolls to per-element |
| 37 | /// scalar affine.apply operations. |
| 38 | struct LowerDelinearizeIndexOps |
| 39 | : public OpRewritePattern<AffineDelinearizeIndexOp> { |
| 40 | using OpRewritePattern<AffineDelinearizeIndexOp>::OpRewritePattern; |
| 41 | LogicalResult matchAndRewrite(AffineDelinearizeIndexOp op, |
| 42 | PatternRewriter &rewriter) const override { |
| 43 | Location loc = op.getLoc(); |
| 44 | Value linearIndex = op.getLinearIndex(); |
| 45 | auto vecTy = dyn_cast<VectorType>(Val: linearIndex.getType()); |
| 46 | |
| 47 | // Scalar case: use the existing affine lowering path. |
| 48 | if (!vecTy) { |
| 49 | FailureOr<SmallVector<Value>> multiIndex = |
| 50 | delinearizeIndex(b&: rewriter, loc, linearIndex, basis: op.getEffectiveBasis(), |
| 51 | /*hasOuterBound=*/false); |
| 52 | if (failed(Result: multiIndex)) |
| 53 | return failure(); |
| 54 | rewriter.replaceOp(op, newValues: *multiIndex); |
| 55 | return success(); |
| 56 | } |
| 57 | |
| 58 | // Vector case: unroll to per-element scalar affine.apply operations |
| 59 | // using StaticTileOffsetRange for multi-dimensional vector support. |
| 60 | if (vecTy.isScalable()) |
| 61 | return rewriter.notifyMatchFailure(arg&: op, msg: "scalable vectors not supported" ); |
| 62 | |
| 63 | unsigned numResults = op.getNumResults(); |
| 64 | ArrayRef<int64_t> shape = vecTy.getShape(); |
| 65 | SmallVector<int64_t> tileShape(shape.size(), 1); |
| 66 | |
| 67 | SmallVector<Value> resultVecs(numResults); |
| 68 | Value poison = ub::PoisonOp::create(builder&: rewriter, location: loc, result: vecTy); |
| 69 | for (unsigned r = 0; r < numResults; ++r) |
| 70 | resultVecs[r] = poison; |
| 71 | |
| 72 | for (SmallVector<int64_t> pos : StaticTileOffsetRange(shape, tileShape)) { |
| 73 | Value scalar = vector::ExtractOp::create(builder&: rewriter, location: loc, source: linearIndex, position: pos); |
| 74 | |
| 75 | FailureOr<SmallVector<Value>> scalarResults = |
| 76 | delinearizeIndex(b&: rewriter, loc, linearIndex: scalar, basis: op.getEffectiveBasis(), |
| 77 | /*hasOuterBound=*/false); |
| 78 | if (failed(Result: scalarResults)) |
| 79 | return failure(); |
| 80 | |
| 81 | for (unsigned r = 0; r < numResults; ++r) |
| 82 | resultVecs[r] = vector::InsertOp::create( |
| 83 | builder&: rewriter, location: loc, valueToStore: (*scalarResults)[r], dest: resultVecs[r], position: pos); |
| 84 | } |
| 85 | |
| 86 | rewriter.replaceOp(op, newValues: resultVecs); |
| 87 | return success(); |
| 88 | } |
| 89 | }; |
| 90 | |
| 91 | /// Lowers `affine.linearize_index` into a sequence of multiplications and |
| 92 | /// additions via affine.apply. For vector types, unrolls to per-element |
| 93 | /// scalar affine.apply operations. |
| 94 | struct LowerLinearizeIndexOps final : OpRewritePattern<AffineLinearizeIndexOp> { |
| 95 | using Base::Base; |
| 96 | LogicalResult matchAndRewrite(AffineLinearizeIndexOp op, |
| 97 | PatternRewriter &rewriter) const override { |
| 98 | Location loc = op.getLoc(); |
| 99 | auto vecTy = dyn_cast<VectorType>(Val: op.getLinearIndex().getType()); |
| 100 | |
| 101 | // Scalar case: use the existing affine lowering path. |
| 102 | if (!vecTy) { |
| 103 | // Should be folded away, included here for safety. |
| 104 | if (op.getMultiIndex().empty()) { |
| 105 | rewriter.replaceOpWithNewOp<arith::ConstantIndexOp>(op, args: 0); |
| 106 | return success(); |
| 107 | } |
| 108 | |
| 109 | SmallVector<OpFoldResult> multiIndex = |
| 110 | getAsOpFoldResult(values: op.getMultiIndex()); |
| 111 | OpFoldResult linearIndex = |
| 112 | linearizeIndex(builder&: rewriter, loc, multiIndex, basis: op.getMixedBasis()); |
| 113 | Value linearIndexValue = |
| 114 | getValueOrCreateConstantIntOp(b&: rewriter, loc, ofr: linearIndex); |
| 115 | rewriter.replaceOp(op, newValues: linearIndexValue); |
| 116 | return success(); |
| 117 | } |
| 118 | |
| 119 | // Vector case: unroll to per-element scalar affine.apply operations |
| 120 | // using StaticTileOffsetRange for multi-dimensional vector support. |
| 121 | if (vecTy.isScalable()) |
| 122 | return rewriter.notifyMatchFailure(arg&: op, msg: "scalable vectors not supported" ); |
| 123 | |
| 124 | ArrayRef<int64_t> shape = vecTy.getShape(); |
| 125 | SmallVector<int64_t> tileShape(shape.size(), 1); |
| 126 | ValueRange multiIndex = op.getMultiIndex(); |
| 127 | |
| 128 | Value result = ub::PoisonOp::create(builder&: rewriter, location: loc, result: vecTy); |
| 129 | |
| 130 | for (SmallVector<int64_t> pos : StaticTileOffsetRange(shape, tileShape)) { |
| 131 | SmallVector<OpFoldResult> scalarIndices; |
| 132 | for (Value vec : multiIndex) |
| 133 | scalarIndices.push_back( |
| 134 | Elt: vector::ExtractOp::create(builder&: rewriter, location: loc, source: vec, position: pos).getResult()); |
| 135 | |
| 136 | OpFoldResult linearIndex = |
| 137 | linearizeIndex(builder&: rewriter, loc, multiIndex: scalarIndices, basis: op.getMixedBasis()); |
| 138 | Value scalarResult = |
| 139 | getValueOrCreateConstantIntOp(b&: rewriter, loc, ofr: linearIndex); |
| 140 | |
| 141 | result = |
| 142 | vector::InsertOp::create(builder&: rewriter, location: loc, valueToStore: scalarResult, dest: result, position: pos); |
| 143 | } |
| 144 | |
| 145 | rewriter.replaceOp(op, newValues: result); |
| 146 | return success(); |
| 147 | } |
| 148 | }; |
| 149 | |
| 150 | class ExpandAffineIndexOpsAsAffinePass |
| 151 | : public affine::impl::AffineExpandIndexOpsAsAffineBase< |
| 152 | ExpandAffineIndexOpsAsAffinePass> { |
| 153 | public: |
| 154 | ExpandAffineIndexOpsAsAffinePass() = default; |
| 155 | |
| 156 | void runOnOperation() override { |
| 157 | MLIRContext *context = &getContext(); |
| 158 | RewritePatternSet patterns(context); |
| 159 | populateAffineExpandIndexOpsAsAffinePatterns(patterns); |
| 160 | if (failed(Result: applyPatternsGreedily(op: getOperation(), patterns: std::move(patterns)))) |
| 161 | return signalPassFailure(); |
| 162 | } |
| 163 | }; |
| 164 | |
| 165 | } // namespace |
| 166 | |
| 167 | void mlir::affine::populateAffineExpandIndexOpsAsAffinePatterns( |
| 168 | RewritePatternSet &patterns) { |
| 169 | patterns.insert<LowerDelinearizeIndexOps, LowerLinearizeIndexOps>( |
| 170 | arg: patterns.getContext()); |
| 171 | } |
| 172 | |
| 173 | std::unique_ptr<Pass> mlir::affine::createAffineExpandIndexOpsAsAffinePass() { |
| 174 | return std::make_unique<ExpandAffineIndexOpsAsAffinePass>(); |
| 175 | } |
| 176 | |