| 1 | //===- LowerVectorInterleave.cpp - Lower 'vector.interleave' operation ----===// |
| 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 target-independent rewrites and utilities to lower the |
| 10 | // 'vector.interleave' operation. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 15 | #include "mlir/Dialect/Vector/Transforms/LoweringPatterns.h" |
| 16 | #include "mlir/Dialect/Vector/Utils/VectorUtils.h" |
| 17 | #include "mlir/IR/BuiltinTypes.h" |
| 18 | #include "mlir/IR/PatternMatch.h" |
| 19 | #include "llvm/ADT/SmallVectorExtras.h" |
| 20 | |
| 21 | #define DEBUG_TYPE "vector-interleave-lowering" |
| 22 | |
| 23 | using namespace mlir; |
| 24 | using namespace mlir::vector; |
| 25 | |
| 26 | namespace { |
| 27 | |
| 28 | /// A one-shot unrolling of vector.interleave to the `targetRank`. |
| 29 | /// |
| 30 | /// Example: |
| 31 | /// |
| 32 | /// ```mlir |
| 33 | /// vector.interleave %a, %b : vector<1x2x3x4xi64> -> vector<1x2x3x8xi64> |
| 34 | /// ``` |
| 35 | /// Would be unrolled to: |
| 36 | /// ```mlir |
| 37 | /// %result = arith.constant dense<0> : vector<1x2x3x8xi64> |
| 38 | /// %0 = vector.extract %a[0, 0, 0] ─┐ |
| 39 | /// : vector<4xi64> from vector<1x2x3x4xi64> | |
| 40 | /// %1 = vector.extract %b[0, 0, 0] | |
| 41 | /// : vector<4xi64> from vector<1x2x3x4xi64> | - Repeated 6x for |
| 42 | /// %2 = vector.interleave %0, %1 : | all leading positions |
| 43 | /// : vector<4xi64> -> vector<8xi64> | |
| 44 | /// %3 = vector.insert %2, %result [0, 0, 0] | |
| 45 | /// : vector<8xi64> into vector<1x2x3x8xi64> ┘ |
| 46 | /// ``` |
| 47 | /// |
| 48 | /// Note: If any leading dimension before the `targetRank` is scalable the |
| 49 | /// unrolling will stop before the scalable dimension. |
| 50 | class UnrollInterleaveOp final : public OpRewritePattern<vector::InterleaveOp> { |
| 51 | public: |
| 52 | UnrollInterleaveOp(int64_t targetRank, MLIRContext *context, |
| 53 | PatternBenefit benefit = 1) |
| 54 | : OpRewritePattern(context, benefit), targetRank(targetRank) {}; |
| 55 | |
| 56 | LogicalResult matchAndRewrite(vector::InterleaveOp op, |
| 57 | PatternRewriter &rewriter) const override { |
| 58 | VectorType resultType = op.getResultVectorType(); |
| 59 | auto unrollIterator = vector::createUnrollIterator(vType: resultType, targetRank); |
| 60 | if (!unrollIterator) |
| 61 | return failure(); |
| 62 | |
| 63 | auto loc = op.getLoc(); |
| 64 | Value result = arith::ConstantOp::create(builder&: rewriter, location: loc, result: resultType, |
| 65 | value: rewriter.getZeroAttr(type: resultType)); |
| 66 | for (auto position : *unrollIterator) { |
| 67 | Value = |
| 68 | ExtractOp::create(builder&: rewriter, location: loc, source: op.getLhs(), position); |
| 69 | Value = |
| 70 | ExtractOp::create(builder&: rewriter, location: loc, source: op.getRhs(), position); |
| 71 | Value interleave = |
| 72 | InterleaveOp::create(builder&: rewriter, location: loc, lhs: extractLhs, rhs: extractRhs); |
| 73 | result = InsertOp::create(builder&: rewriter, location: loc, valueToStore: interleave, dest: result, position); |
| 74 | } |
| 75 | |
| 76 | rewriter.replaceOp(op, newValues: result); |
| 77 | return success(); |
| 78 | } |
| 79 | |
| 80 | private: |
| 81 | int64_t targetRank = 1; |
| 82 | }; |
| 83 | |
| 84 | /// A one-shot unrolling of vector.deinterleave to the `targetRank`. |
| 85 | /// |
| 86 | /// Example: |
| 87 | /// |
| 88 | /// ```mlir |
| 89 | /// %0, %1 = vector.deinterleave %a : vector<1x2x3x8xi64> -> vector<1x2x3x4xi64> |
| 90 | /// ``` |
| 91 | /// Would be unrolled to: |
| 92 | /// ```mlir |
| 93 | /// %result = arith.constant dense<0> : vector<1x2x3x4xi64> |
| 94 | /// %0 = vector.extract %a[0, 0, 0] ─┐ |
| 95 | /// : vector<8xi64> from vector<1x2x3x8xi64> | |
| 96 | /// %1, %2 = vector.deinterleave %0 | |
| 97 | /// : vector<8xi64> -> vector<4xi64> | -- Initial deinterleave |
| 98 | /// %3 = vector.insert %1, %result [0, 0, 0] | operation unrolled. |
| 99 | /// : vector<4xi64> into vector<1x2x3x4xi64> | |
| 100 | /// %4 = vector.insert %2, %result [0, 0, 0] | |
| 101 | /// : vector<4xi64> into vector<1x2x3x4xi64> ┘ |
| 102 | /// %5 = vector.extract %a[0, 0, 1] ─┐ |
| 103 | /// : vector<8xi64> from vector<1x2x3x8xi64> | |
| 104 | /// %6, %7 = vector.deinterleave %5 | |
| 105 | /// : vector<8xi64> -> vector<4xi64> | -- Recursive pattern for |
| 106 | /// %8 = vector.insert %6, %3 [0, 0, 1] | subsequent unrolled |
| 107 | /// : vector<4xi64> into vector<1x2x3x4xi64> | deinterleave |
| 108 | /// %9 = vector.insert %7, %4 [0, 0, 1] | operations. Repeated |
| 109 | /// : vector<4xi64> into vector<1x2x3x4xi64> ┘ 5x in this case. |
| 110 | /// ``` |
| 111 | /// |
| 112 | /// Note: If any leading dimension before the `targetRank` is scalable the |
| 113 | /// unrolling will stop before the scalable dimension. |
| 114 | class UnrollDeinterleaveOp final |
| 115 | : public OpRewritePattern<vector::DeinterleaveOp> { |
| 116 | public: |
| 117 | UnrollDeinterleaveOp(int64_t targetRank, MLIRContext *context, |
| 118 | PatternBenefit benefit = 1) |
| 119 | : OpRewritePattern(context, benefit), targetRank(targetRank) {}; |
| 120 | |
| 121 | LogicalResult matchAndRewrite(vector::DeinterleaveOp op, |
| 122 | PatternRewriter &rewriter) const override { |
| 123 | VectorType resultType = op.getResultVectorType(); |
| 124 | auto unrollIterator = vector::createUnrollIterator(vType: resultType, targetRank); |
| 125 | if (!unrollIterator) |
| 126 | return failure(); |
| 127 | |
| 128 | auto loc = op.getLoc(); |
| 129 | Value emptyResult = arith::ConstantOp::create( |
| 130 | builder&: rewriter, location: loc, result: resultType, value: rewriter.getZeroAttr(type: resultType)); |
| 131 | Value evenResult = emptyResult; |
| 132 | Value oddResult = emptyResult; |
| 133 | |
| 134 | for (auto position : *unrollIterator) { |
| 135 | auto = |
| 136 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: op.getSource(), position); |
| 137 | auto deinterleave = |
| 138 | vector::DeinterleaveOp::create(builder&: rewriter, location: loc, source: extractSrc); |
| 139 | evenResult = vector::InsertOp::create( |
| 140 | builder&: rewriter, location: loc, valueToStore: deinterleave.getResultEven(), dest: evenResult, position); |
| 141 | oddResult = vector::InsertOp::create( |
| 142 | builder&: rewriter, location: loc, valueToStore: deinterleave.getResultOdd(), dest: oddResult, position); |
| 143 | } |
| 144 | rewriter.replaceOp(op, newValues: ValueRange{evenResult, oddResult}); |
| 145 | return success(); |
| 146 | } |
| 147 | |
| 148 | private: |
| 149 | int64_t targetRank = 1; |
| 150 | }; |
| 151 | |
| 152 | /// Rewrite vector.interleave op into an equivalent vector.shuffle op, when |
| 153 | /// applicable: `sourceType` must be 0D or 1D, and non-scalable. |
| 154 | /// |
| 155 | /// Example: |
| 156 | /// |
| 157 | /// ```mlir |
| 158 | /// vector.interleave %a, %b : vector<7xi16> -> vector<14xi16> |
| 159 | /// ``` |
| 160 | /// |
| 161 | /// Is rewritten into: |
| 162 | /// |
| 163 | /// ```mlir |
| 164 | /// vector.shuffle %arg0, %arg1 [0, 7, 1, 8, 2, 9, 3, 10, 4, 11, 5, 12, 6, 13] |
| 165 | /// : vector<7xi16>, vector<7xi16> |
| 166 | /// ``` |
| 167 | struct InterleaveToShuffle final : OpRewritePattern<vector::InterleaveOp> { |
| 168 | using Base::Base; |
| 169 | |
| 170 | LogicalResult matchAndRewrite(vector::InterleaveOp op, |
| 171 | PatternRewriter &rewriter) const override { |
| 172 | VectorType sourceType = op.getSourceVectorType(); |
| 173 | if (sourceType.getRank() > 1 || sourceType.isScalable()) { |
| 174 | return failure(); |
| 175 | } |
| 176 | int64_t n = sourceType.getNumElements(); |
| 177 | auto seq = llvm::seq<int64_t>(Size: 2 * n); |
| 178 | auto zip = llvm::map_to_vector( |
| 179 | C&: seq, F: [n](int64_t i) { return (i % 2 ? n : 0) + i / 2; }); |
| 180 | rewriter.replaceOpWithNewOp<ShuffleOp>(op, args: op.getLhs(), args: op.getRhs(), args&: zip); |
| 181 | return success(); |
| 182 | } |
| 183 | }; |
| 184 | |
| 185 | /// Rewrite vector.deinterleave op into two equivalent vector.shuffle ops, when |
| 186 | /// applicable: `sourceType` must be 1D and non-scalable. |
| 187 | /// |
| 188 | /// Example: |
| 189 | /// |
| 190 | /// ```mlir |
| 191 | /// %evens, %odds = vector.deinterleave %arg0 : vector<4xi32> -> vector<2xi32> |
| 192 | /// ``` |
| 193 | /// |
| 194 | /// Is rewritten into: |
| 195 | /// |
| 196 | /// ```mlir |
| 197 | /// %evens = vector.shuffle %arg0, %arg0 [0, 2] : vector<4xi32>, vector<4xi32> |
| 198 | /// %odds = vector.shuffle %arg0, %arg0 [1, 3] : vector<4xi32>, vector<4xi32> |
| 199 | /// ``` |
| 200 | struct DeinterleaveToShuffle final : OpRewritePattern<vector::DeinterleaveOp> { |
| 201 | using Base::Base; |
| 202 | |
| 203 | LogicalResult matchAndRewrite(vector::DeinterleaveOp op, |
| 204 | PatternRewriter &rewriter) const override { |
| 205 | VectorType sourceType = op.getSourceVectorType(); |
| 206 | if (sourceType.getRank() != 1 || sourceType.isScalable()) { |
| 207 | return failure(); |
| 208 | } |
| 209 | |
| 210 | auto seq = llvm::seq<int64_t>(Size: sourceType.getNumElements() / 2); |
| 211 | auto evenZip = llvm::map_to_vector(C&: seq, F: [](int64_t i) { return i * 2; }); |
| 212 | auto oddZip = llvm::map_to_vector(C&: evenZip, F: [](int64_t i) { return i + 1; }); |
| 213 | |
| 214 | Value evenResult = vector::ShuffleOp::create( |
| 215 | builder&: rewriter, location: op.getLoc(), v1: op.getOperand(), v2: op.getOperand(), mask: evenZip); |
| 216 | Value oddResult = vector::ShuffleOp::create( |
| 217 | builder&: rewriter, location: op.getLoc(), v1: op.getOperand(), v2: op.getOperand(), mask: oddZip); |
| 218 | |
| 219 | rewriter.replaceOp(op, newValues: ValueRange{evenResult, oddResult}); |
| 220 | return success(); |
| 221 | } |
| 222 | }; |
| 223 | |
| 224 | } // namespace |
| 225 | |
| 226 | void mlir::vector::populateVectorInterleaveLoweringPatterns( |
| 227 | RewritePatternSet &patterns, int64_t targetRank, PatternBenefit benefit) { |
| 228 | patterns.add<UnrollInterleaveOp, UnrollDeinterleaveOp>( |
| 229 | arg&: targetRank, args: patterns.getContext(), args&: benefit); |
| 230 | } |
| 231 | |
| 232 | void mlir::vector::populateVectorInterleaveToShufflePatterns( |
| 233 | RewritePatternSet &patterns, PatternBenefit benefit) { |
| 234 | patterns.add<InterleaveToShuffle>(arg: patterns.getContext(), args&: benefit); |
| 235 | } |
| 236 | |
| 237 | void mlir::vector::populateVectorDeinterleaveToShufflePatterns( |
| 238 | RewritePatternSet &patterns, PatternBenefit benefit) { |
| 239 | patterns.add<DeinterleaveToShuffle>(arg: patterns.getContext(), args&: benefit); |
| 240 | } |
| 241 | |