| 1 | //===- FoldAddIntoDest.cpp ---------------------------------------*- C++-*-===// |
| 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/Linalg/IR/Linalg.h" |
| 10 | #include "mlir/Dialect/Linalg/IR/LinalgInterfaces.h" |
| 11 | #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
| 12 | #include "mlir/IR/Dominance.h" |
| 13 | #include "mlir/Interfaces/DestinationStyleOpInterface.h" |
| 14 | |
| 15 | using namespace mlir; |
| 16 | |
| 17 | // Determine whether the value is defined to be zero. |
| 18 | static bool isDefinedAsZero(Value val) { |
| 19 | if (!val) |
| 20 | return false; |
| 21 | |
| 22 | // Check whether val is a constant scalar / vector splat / tensor splat float |
| 23 | // or integer zero. |
| 24 | if (matchPattern(value: val, pattern: m_AnyZeroFloat()) || matchPattern(value: val, pattern: m_Zero())) |
| 25 | return true; |
| 26 | |
| 27 | return TypeSwitch<Operation *, bool>(val.getDefiningOp()) |
| 28 | .Case<linalg::FillOp, linalg::CopyOp>([&](auto op) { |
| 29 | return op && op.getInputs().size() == 1 && |
| 30 | isDefinedAsZero(op.getInputs()[0]); |
| 31 | }) |
| 32 | .Default([&](auto) { return false; }); |
| 33 | } |
| 34 | |
| 35 | /// Replace a linalg.add with one operand the single user of a contraction, |
| 36 | /// which has a zero-filled, "identity-mapped" destination and is dominated by |
| 37 | /// the `other` operand, by the contraction with `other` as its dest. |
| 38 | /// |
| 39 | /// As an example, the following pseudo-code will be rewritten |
| 40 | /// %cst = arith.constant 0.000000e+00 |
| 41 | /// %empty = tensor.empty() |
| 42 | /// %zeroed = linalg.fill ins(%cst : f32) outs(%empty : !type) -> !type |
| 43 | /// %C = linalg.matmul ins(%A, %B) outs(%zeroed) |
| 44 | /// %empty2 = tensor.empty() |
| 45 | /// %zeroed2 = linalg.fill ins(%cst : f32) outs(%empty2 : !type) -> !type |
| 46 | /// %F = linalg.matmul ins(%D, %E) outs(%zeroed2) |
| 47 | /// %out = linalg.add ins(%C, %F) outs(%empty) |
| 48 | /// to: |
| 49 | /// %cst = arith.constant 0.000000e+00 |
| 50 | /// %empty = tensor.empty() |
| 51 | /// %zeroed = linalg.fill ins(%cst : f32) outs(%empty : !type) -> !type |
| 52 | /// %C = linalg.matmul ins(%A, %B) outs(%zeroed) |
| 53 | /// %out = linalg.matmul ins(%D, %E) outs(%C) |
| 54 | /// |
| 55 | struct FoldAddIntoDest final : public OpRewritePattern<linalg::AddOp> { |
| 56 | using OpRewritePattern<linalg::AddOp>::OpRewritePattern; |
| 57 | |
| 58 | LogicalResult matchAndRewrite(linalg::AddOp addOp, |
| 59 | PatternRewriter &rewriter) const override { |
| 60 | // For now, pattern only applies on tensor types (memref support is TODO). |
| 61 | if (!addOp.hasPureTensorSemantics()) |
| 62 | return failure(); |
| 63 | |
| 64 | Value dominatingOperand = nullptr; |
| 65 | linalg::LinalgOp dominatedOp = nullptr; |
| 66 | { // We will forget about which operand was left or right after this block. |
| 67 | Value lhs = addOp.getInputs()[0]; |
| 68 | Value rhs = addOp.getInputs()[1]; |
| 69 | |
| 70 | // Can only put one of addOp's operands in the dest/out arg of the other's |
| 71 | // defining op based on suitable dominance. |
| 72 | // TODO: Can be generalized to move ops around as long as that still |
| 73 | // respects use-def chains and doesn't affect side-effects. |
| 74 | if (auto rhsOp = rhs.getDefiningOp<linalg::LinalgOp>()) { |
| 75 | DominanceInfo domInfo(rhsOp); |
| 76 | if (domInfo.properlyDominates(lhs, rhsOp)) { |
| 77 | dominatingOperand = lhs; |
| 78 | dominatedOp = rhsOp; |
| 79 | } |
| 80 | } |
| 81 | if (auto lhsOp = lhs.getDefiningOp<linalg::LinalgOp>()) { |
| 82 | DominanceInfo domInfo(lhsOp); |
| 83 | if (domInfo.properlyDominates(rhs, lhsOp)) { |
| 84 | dominatingOperand = rhs; |
| 85 | dominatedOp = lhsOp; |
| 86 | } |
| 87 | } |
| 88 | if (!dominatingOperand || !dominatedOp) |
| 89 | return failure(); |
| 90 | // NB: As linalg.add's generalisation ignores the out argument in its |
| 91 | // region there is no need to perform checks on addOp's out argument. |
| 92 | } |
| 93 | |
| 94 | // When dominated op is a contraction we know it accumulates on its out arg. |
| 95 | // E.g., AddOp is not a contraction and hence ignores its out arg's value. |
| 96 | // TODO: Generalize check to also pass in case of other LinalgOps that |
| 97 | // accumulate on their out arg but are not (binary) contraction ops. |
| 98 | auto dominatedDestOp = |
| 99 | dyn_cast<DestinationStyleOpInterface>((Operation *)dominatedOp); |
| 100 | if (dominatedOp->getNumResults() != 1 || |
| 101 | !linalg::isaContractionOpInterface(linalgOp: dominatedOp) || |
| 102 | (!dominatedDestOp || dominatedDestOp.getNumDpsInits() != 1)) |
| 103 | return rewriter.notifyMatchFailure( |
| 104 | dominatedOp, "expected dominated op to be single-result " |
| 105 | "destination-passing contraction" ); |
| 106 | |
| 107 | // To change the contraction's result, `addOp` must be its only user. |
| 108 | if (!dominatedOp->getResult(0).hasOneUse()) |
| 109 | return rewriter.notifyMatchFailure( |
| 110 | dominatedOp, |
| 111 | "expected linalg.add to be single user of contraction's result" ); |
| 112 | |
| 113 | // As `dominatedOp` was already accumulating on its out argument, it is only |
| 114 | // safe to no longer use its current out arg when it is the additive ident. |
| 115 | auto *destOperand = dominatedDestOp.getDpsInitOperand(0); |
| 116 | if (!isDefinedAsZero(destOperand->get())) |
| 117 | return rewriter.notifyMatchFailure( |
| 118 | dominatedOp, "expected dominated op's dest to be additive zero" ); |
| 119 | // TODO: If the other op is a contraction and has additive ident as dest, we |
| 120 | // can swap the dests and achieve the proper sum, given suitable dominance. |
| 121 | |
| 122 | // As an operand to `addOp`, `dominatingOperand` has an identity affine_map. |
| 123 | // Hence, we can only substitute `dominatingOperand` for the dest of the |
| 124 | // contraction when dest's indexing_map corresponds to an identity map |
| 125 | // w.r.t. just the dimensions of dest, i.e. is an ordered projection. |
| 126 | SmallVector<AffineMap> indexMaps = dominatedOp.getIndexingMapsArray(); |
| 127 | int prevDimPos = -1; |
| 128 | for (auto expr : indexMaps[destOperand->getOperandNumber()].getResults()) { |
| 129 | auto dim = dyn_cast<AffineDimExpr>(expr); |
| 130 | if (!dim || prevDimPos > static_cast<int>(dim.getPosition())) |
| 131 | return rewriter.notifyMatchFailure( |
| 132 | dominatedOp, "expected index_map for contraction's dest to be an " |
| 133 | "ordered projection" ); |
| 134 | prevDimPos = dim.getPosition(); |
| 135 | } |
| 136 | |
| 137 | // Replace the additive-ident, i.e. zero, out arg of the dominated op by the |
| 138 | // dominating summand. This makes the dominated op's result the sum of both |
| 139 | // of addOp's arguments - therefore we replace addOp and it uses by it. |
| 140 | rewriter.modifyOpInPlace( |
| 141 | dominatedOp, [&]() { dominatedOp->setOperand(2, dominatingOperand); }); |
| 142 | rewriter.replaceAllOpUsesWith(addOp, dominatedOp->getResult(0)); |
| 143 | return success(); |
| 144 | } |
| 145 | }; |
| 146 | |
| 147 | void linalg::populateFoldAddIntoDestPatterns(RewritePatternSet &patterns) { |
| 148 | // Replace linalg.add when destination passing suffices for achieving the sum. |
| 149 | patterns.add<FoldAddIntoDest>(arg: patterns.getContext()); |
| 150 | } |
| 151 | |