| 1 | //===- ShuffleVectorFMAOps.cpp --------------------------------------------===// |
| 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/Dialect/Vector/IR/VectorOps.h" |
| 10 | #include "mlir/Dialect/X86/Transforms.h" |
| 11 | #include "mlir/Dialect/X86/X86Dialect.h" |
| 12 | |
| 13 | #include "mlir/IR/PatternMatch.h" |
| 14 | |
| 15 | #include "mlir/Pass/Pass.h" |
| 16 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| 17 | |
| 18 | using namespace mlir; |
| 19 | using namespace mlir::vector; |
| 20 | using namespace mlir::x86; |
| 21 | |
| 22 | namespace { |
| 23 | |
| 24 | // Validates whether the given operation is an x86 operation and has only |
| 25 | // one consumer. |
| 26 | static bool validateFMAOperands(Value op) { |
| 27 | if (auto cvt = op.getDefiningOp<x86::avx::CvtPackedEvenIndexedToF32Op>()) |
| 28 | return cvt.getResult().hasOneUse(); |
| 29 | |
| 30 | if (auto bcst = op.getDefiningOp<x86::avx::BcstToPackedF32Op>()) |
| 31 | return bcst.getResult().hasOneUse(); |
| 32 | |
| 33 | return false; |
| 34 | } |
| 35 | |
| 36 | // Validates the vector.fma operation on the following conditions: |
| 37 | // (i) one of the lhs or rhs defining operation should be |
| 38 | // CvtPackedEvenIndexedToF32Op, (ii) the lhs or rhs defining operation should be |
| 39 | // an x86 operation and has only one consumer, (iii) all operations |
| 40 | // are in the same block, and (iv) ths FMA has only one user. |
| 41 | static bool validateVectorFMAOp(vector::FMAOp fmaOp) { |
| 42 | Value lhs = fmaOp.getLhs(); |
| 43 | Value rhs = fmaOp.getRhs(); |
| 44 | |
| 45 | if (!isa<x86::avx::CvtPackedEvenIndexedToF32Op>(Val: lhs.getDefiningOp()) && |
| 46 | !isa<x86::avx::CvtPackedEvenIndexedToF32Op>(Val: rhs.getDefiningOp())) |
| 47 | return false; |
| 48 | |
| 49 | if (!validateFMAOperands(op: lhs) || !validateFMAOperands(op: rhs)) |
| 50 | return false; |
| 51 | |
| 52 | if (lhs.getDefiningOp()->getBlock() != rhs.getDefiningOp()->getBlock()) |
| 53 | return false; |
| 54 | |
| 55 | if (lhs.getDefiningOp()->getBlock() != fmaOp->getBlock()) |
| 56 | return false; |
| 57 | |
| 58 | if (!fmaOp.getResult().hasOneUse()) |
| 59 | return false; |
| 60 | |
| 61 | Operation *consumer = *fmaOp.getResult().getUsers().begin(); |
| 62 | if (consumer->getBlock() != fmaOp->getBlock()) |
| 63 | return false; |
| 64 | |
| 65 | return true; |
| 66 | } |
| 67 | |
| 68 | // Moves vector.fma along with the lhs and rhs defining operation before its |
| 69 | // consumer. If the consumer is vector.ShapeCastOp and has only one user then |
| 70 | // move before the consumer of vector.ShapeCastOp. |
| 71 | // TODO: Move before first consumer, if there are multiple. |
| 72 | static void moveFMA(PatternRewriter &rewriter, vector::FMAOp fmaOp) { |
| 73 | Operation *consumer = *fmaOp.getResult().getUsers().begin(); |
| 74 | |
| 75 | if (auto shapeCastOp = dyn_cast<vector::ShapeCastOp>(Val: consumer)) { |
| 76 | if (shapeCastOp.getResult().hasOneUse()) { |
| 77 | Operation *nxtConsumer = *shapeCastOp.getResult().getUsers().begin(); |
| 78 | if (nxtConsumer->getBlock() == fmaOp->getBlock()) { |
| 79 | consumer = *shapeCastOp.getResult().getUsers().begin(); |
| 80 | rewriter.moveOpBefore(op: fmaOp.getLhs().getDefiningOp(), existingOp: consumer); |
| 81 | rewriter.moveOpBefore(op: fmaOp.getRhs().getDefiningOp(), existingOp: consumer); |
| 82 | rewriter.moveOpBefore(op: fmaOp.getOperation(), existingOp: consumer); |
| 83 | rewriter.moveOpBefore(op: shapeCastOp.getOperation(), existingOp: consumer); |
| 84 | return; |
| 85 | } |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | rewriter.moveOpBefore(op: fmaOp.getLhs().getDefiningOp(), existingOp: consumer); |
| 90 | rewriter.moveOpBefore(op: fmaOp.getRhs().getDefiningOp(), existingOp: consumer); |
| 91 | rewriter.moveOpBefore(op: fmaOp.getOperation(), existingOp: consumer); |
| 92 | |
| 93 | return; |
| 94 | } |
| 95 | |
| 96 | // Shuffle FMAs with x86 operations as operands such that |
| 97 | // FMAs are grouped with respect to odd/even packed index. |
| 98 | // |
| 99 | // For example: |
| 100 | // ``` |
| 101 | // %1 = x86.avx.bcst_to_f32.packed |
| 102 | // %2 = x86.avx.cvt.packed.odd.indexed_to_f32 |
| 103 | // %3 = vector.fma %1, %2, %arg1 |
| 104 | // %4 = x86.avx.bcst_to_f32.packed |
| 105 | // %5 = x86.avx.cvt.packed.even.indexed_to_f32 |
| 106 | // %6 = vector.fma %4, %5, %3 |
| 107 | // %7 = x86.avx.bcst_to_f32.packed |
| 108 | // %8 = x86.avx.cvt.packed.odd.indexed_to_f32 |
| 109 | // %9 = vector.fma %7, %8, %arg2 |
| 110 | // %10 = x86.avx.bcst_to_f32.packed |
| 111 | // %11 = x86.avx.cvt.packed.even.indexed_to_f32 |
| 112 | // %12 = vector.fma %10, %11, %9 |
| 113 | // yield %6, %12 |
| 114 | // ``` |
| 115 | // to |
| 116 | // ``` |
| 117 | // %1 = x86.avx.bcst_to_f32.packed |
| 118 | // %2 = x86.avx.cvt.packed.odd.indexed_to_f32 |
| 119 | // %3 = vector.fma %1, %2, %arg1 |
| 120 | // %7 = x86.avx.bcst_to_f32.packed |
| 121 | // %8 = x86.avx.cvt.packed.odd.indexed_to_f32 |
| 122 | // %9 = vector.fma %7, %8, %arg2 |
| 123 | // %4 = x86.avx.bcst_to_f32.packed |
| 124 | // %5 = x86.avx.cvt.packed.even.indexed_to_f32 |
| 125 | // %6 = vector.fma %4, %5, %3 |
| 126 | // %10 = x86.avx.bcst_to_f32.packed |
| 127 | // %11 = x86.avx.cvt.packed.even.indexed_to_f32 |
| 128 | // %12 = vector.fma %10, %11, %9 |
| 129 | // yield %9, %12 |
| 130 | // ``` |
| 131 | // TODO: Shuffling supported only if the FMA, lhs/rhs defining operations |
| 132 | // have only one consumer. Have to extend this pass for multiple consumers. |
| 133 | struct ShuffleVectorFMAOps : public OpRewritePattern<vector::FMAOp> { |
| 134 | using OpRewritePattern<vector::FMAOp>::OpRewritePattern; |
| 135 | |
| 136 | LogicalResult matchAndRewrite(vector::FMAOp fmaOp, |
| 137 | PatternRewriter &rewriter) const override { |
| 138 | |
| 139 | if (!validateVectorFMAOp(fmaOp)) |
| 140 | return failure(); |
| 141 | |
| 142 | llvm::SmallVector<vector::FMAOp> fmaOps; |
| 143 | Operation *nextOp = fmaOp; |
| 144 | bool stopAtNextDependentFMA = true; |
| 145 | |
| 146 | // Break the loop and return failure if the immediate next FMA op |
| 147 | // have CvtPackedEvenIndexedToF32Op in it's lhs/rhs defining ops. |
| 148 | while ((nextOp = nextOp->getNextNode())) { |
| 149 | auto fma = dyn_cast<vector::FMAOp>(Val: nextOp); |
| 150 | if (!fma) |
| 151 | continue; |
| 152 | |
| 153 | bool hasX86CvtOperand = isa<x86::avx::CvtPackedEvenIndexedToF32Op>( |
| 154 | Val: fma.getLhs().getDefiningOp()) || |
| 155 | isa<x86::avx::CvtPackedEvenIndexedToF32Op>( |
| 156 | Val: fma.getRhs().getDefiningOp()); |
| 157 | |
| 158 | if (hasX86CvtOperand && stopAtNextDependentFMA) |
| 159 | break; |
| 160 | |
| 161 | if (validateVectorFMAOp(fmaOp: fma)) |
| 162 | fmaOps.push_back(Elt: fma); |
| 163 | |
| 164 | stopAtNextDependentFMA = false; |
| 165 | } |
| 166 | |
| 167 | if (fmaOps.empty()) |
| 168 | return rewriter.notifyMatchFailure( |
| 169 | arg&: fmaOp, msg: "No eligible FMA operations were found: the operation may " |
| 170 | "already be shuffled, there may be no following FMAs, or the " |
| 171 | "following FMAs do not satisfy the shuffle conditions." ); |
| 172 | |
| 173 | fmaOps.push_back(Elt: fmaOp); |
| 174 | for (auto fmaOp : fmaOps) |
| 175 | moveFMA(rewriter, fmaOp); |
| 176 | |
| 177 | return success(); |
| 178 | } |
| 179 | }; |
| 180 | |
| 181 | } // namespace |
| 182 | |
| 183 | void x86::populateShuffleVectorFMAOpsPatterns(RewritePatternSet &patterns) { |
| 184 | patterns.add<ShuffleVectorFMAOps>(arg: patterns.getContext()); |
| 185 | } |
| 186 | |