| 1 | //===- ExtractSliceFromReshapeUtils.cpp - Slice reshape rewrites ----------===// |
| 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 rewrites that replace slices of reshape results with |
| 10 | // aggregated slices of the reshape source. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | #include "mlir/Dialect/Affine/IR/AffineOps.h" |
| 14 | #include "mlir/Dialect/Arith/Utils/Utils.h" |
| 15 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 16 | #include "mlir/Dialect/Tensor/Transforms/TransformUtils.h" |
| 17 | #include "mlir/Dialect/Utils/ReshapeOpsUtils.h" |
| 18 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 19 | #include "mlir/IR/BuiltinTypes.h" |
| 20 | #include "mlir/IR/OpDefinition.h" |
| 21 | #include "llvm/ADT/STLExtras.h" |
| 22 | |
| 23 | using namespace mlir; |
| 24 | using namespace mlir::affine; |
| 25 | using namespace mlir::tensor; |
| 26 | |
| 27 | /// A tuple that represents (dimension number, dimension value). |
| 28 | using DimAndIndex = std::tuple<unsigned, Value>; |
| 29 | |
| 30 | /// Transform `dimAndIndex` from the output index space of a (non-rank-reducing) |
| 31 | /// slice described by `sliceParams` into the input index space. |
| 32 | static DimAndIndex invertSliceIndexing(OpBuilder &b, Location loc, |
| 33 | ArrayRef<Range> sliceParams, |
| 34 | const DimAndIndex &dimAndIndex) { |
| 35 | AffineExpr d0, s0, s1; |
| 36 | bindDims(ctx: b.getContext(), exprs&: d0); |
| 37 | bindSymbols(ctx: b.getContext(), exprs&: s0, exprs&: s1); |
| 38 | auto [dim, indexValue] = dimAndIndex; |
| 39 | assert(dim < sliceParams.size() && "slice should be non rank-reducing" ); |
| 40 | return std::make_pair( |
| 41 | x&: dim, y: affine::makeComposedAffineApply( |
| 42 | b, loc, e: s0 + d0 * s1, |
| 43 | operands: {indexValue, sliceParams[dim].offset, sliceParams[dim].stride})); |
| 44 | } |
| 45 | |
| 46 | /// Transform `dimAndIndex` from the result tensor index space of a |
| 47 | /// CollapseShapeOp to the source tensor index space. |
| 48 | static ValueRange invertCollapseShapeIndexing( |
| 49 | OpBuilder &b, Location loc, ArrayRef<ReassociationIndices> reassociation, |
| 50 | ArrayRef<OpFoldResult> reshapeSourceShape, const DimAndIndex &dimAndIndex) { |
| 51 | const auto &[dim, indexValue] = dimAndIndex; |
| 52 | SmallVector<OpFoldResult> basis; |
| 53 | for (int64_t i : reassociation[dim]) |
| 54 | basis.push_back(Elt: reshapeSourceShape[i]); |
| 55 | auto delinearized = |
| 56 | b.create<AffineDelinearizeIndexOp>(location: loc, args: indexValue, args&: basis); |
| 57 | return delinearized->getResults(); |
| 58 | } |
| 59 | |
| 60 | FailureOr<ExtractSliceFromCollapseHelper> |
| 61 | tensor::ExtractSliceFromCollapseHelper::( |
| 62 | OpBuilder &b, tensor::CollapseShapeOp collapseOp, |
| 63 | tensor::ExtractSliceOp ) { |
| 64 | if (extractOp.getSource().getDefiningOp<tensor::CollapseShapeOp>() != |
| 65 | collapseOp) |
| 66 | return failure(); |
| 67 | SmallVector<Range> ranges; |
| 68 | ranges.reserve(N: extractOp.getSourceType().getRank()); |
| 69 | for (const auto &[o, s, st] : |
| 70 | llvm::zip(t: extractOp.getMixedOffsets(), u: extractOp.getMixedSizes(), |
| 71 | args: extractOp.getMixedStrides())) { |
| 72 | ranges.push_back(Elt: {.offset: o, .size: s, .stride: st}); |
| 73 | } |
| 74 | return ExtractSliceFromCollapseHelper::create(b, op: collapseOp, sliceParams: ranges); |
| 75 | } |
| 76 | |
| 77 | FailureOr<ExtractSliceFromCollapseHelper> |
| 78 | tensor::ExtractSliceFromCollapseHelper::(OpBuilder &b, |
| 79 | tensor::CollapseShapeOp op, |
| 80 | ArrayRef<Range> sliceParams) { |
| 81 | // Don't perform this pattern if the collapse op can be simplified by |
| 82 | // a rank-reducing extract slice. |
| 83 | if (succeeded(Result: mlir::getSimplifyCollapseShapeWithRankReducingSliceInfo( |
| 84 | sourceType: op.getSrcType(), reassociationIndices: op.getReassociationIndices()))) |
| 85 | return failure(); |
| 86 | |
| 87 | // Materialize the output shape of the collapse_shape operation. This will |
| 88 | // create IR describing the output shape in terms of the input shape. |
| 89 | ReifiedRankedShapedTypeDims reifiedShapes; |
| 90 | if (failed(Result: reifyResultShapes(b, op, reifiedReturnShapes&: reifiedShapes))) |
| 91 | return failure(); |
| 92 | SmallVector<OpFoldResult> &collapseShapeOutputShape = reifiedShapes[0]; |
| 93 | SmallVector<ReassociationIndices> reassociationIndices = |
| 94 | op.getReassociationIndices(); |
| 95 | |
| 96 | // Determine which of the CollapseShapeOp's result dimensions are sliced |
| 97 | // and/or linearized. |
| 98 | llvm::SmallBitVector linearizedDimensions = |
| 99 | getLinearizedDimensions(reassociationIndices); |
| 100 | llvm::SmallBitVector slicedDimensions = |
| 101 | getSlicedDimensions(sliceInputShape: collapseShapeOutputShape, sliceParams); |
| 102 | |
| 103 | auto collapseShapeInputShape = |
| 104 | tensor::getMixedSizes(builder&: b, loc: op.getLoc(), value: op.getSrc()); |
| 105 | |
| 106 | SmallVector<Value> tileSizes; |
| 107 | for (unsigned i = 0; i < sliceParams.size(); i++) { |
| 108 | if (slicedDimensions[i] && linearizedDimensions[i]) |
| 109 | tileSizes.push_back( |
| 110 | Elt: getValueOrCreateConstantIndexOp(b, loc: op.getLoc(), ofr: sliceParams[i].size)); |
| 111 | } |
| 112 | |
| 113 | return ExtractSliceFromCollapseHelper( |
| 114 | op, collapseShapeInputShape, collapseShapeOutputShape, sliceParams, |
| 115 | linearizedDimensions, slicedDimensions, tileSizes); |
| 116 | } |
| 117 | |
| 118 | std::pair<Value, SmallVector<Range>> |
| 119 | tensor::ExtractSliceFromCollapseHelper::emitLoopNestBody( |
| 120 | OpBuilder &builder, Location loc, ValueRange tileInductionVars) { |
| 121 | // Create the helper class for forming the slice parameters. |
| 122 | const SmallVector<ReassociationIndices> reassociationIndices = |
| 123 | collapseShapeOp.getReassociationIndices(); |
| 124 | SliceFromCollapseHelper helper(reassociationIndices, collapseShapeInputShape, |
| 125 | collapseShapeOutputShape, sliceParams); |
| 126 | |
| 127 | // Get the indices of the tiled dims (linearized by the collapse_shape |
| 128 | // and sliced by the extract_slice) invert the index spaces |
| 129 | // transformations. |
| 130 | SmallVector<ValueRange> multiIndices; |
| 131 | unsigned loopIdx = 0; |
| 132 | for (unsigned i = 0, e = linearizedDimensions.size(); i < e; i++) { |
| 133 | if (linearizedDimensions[i] && slicedDimensions[i]) { |
| 134 | DimAndIndex tb = |
| 135 | invertSliceIndexing(b&: builder, loc, sliceParams, |
| 136 | dimAndIndex: std::make_tuple(args&: i, args: tileInductionVars[loopIdx++])); |
| 137 | multiIndices.push_back(Elt: invertCollapseShapeIndexing( |
| 138 | b&: builder, loc, reassociation: reassociationIndices, reshapeSourceShape: collapseShapeInputShape, dimAndIndex: tb)); |
| 139 | } |
| 140 | } |
| 141 | |
| 142 | SmallVector<Range> = |
| 143 | helper.getExtractSliceParams(ctx: builder.getContext(), multiIndices); |
| 144 | |
| 145 | Value subTileResult = builder.create<tensor::ExtractSliceOp>( |
| 146 | location: loc, args: collapseShapeOp.getSrc(), args&: extractParams); |
| 147 | |
| 148 | SmallVector<Range> insertParams = |
| 149 | helper.getInsertSliceParams(ctx: builder.getContext(), tileIndices: tileInductionVars); |
| 150 | |
| 151 | // Collapse the dimensions of the source slice back down. |
| 152 | Value collapsedResult = builder.create<tensor::CollapseShapeOp>( |
| 153 | location: loc, args&: subTileResult, args: reassociationIndices); |
| 154 | return std::make_pair(x&: collapsedResult, y&: insertParams); |
| 155 | } |
| 156 | |
| 157 | FailureOr<Operation *> |
| 158 | tensor::( |
| 159 | tensor::CollapseShapeOp op, RewriterBase &rewriter) { |
| 160 | SmallVector<ReassociationIndices> reassociationIndices = |
| 161 | op.getReassociationIndices(); |
| 162 | RankedTensorType sourceType = op.getSrcType(); |
| 163 | FailureOr<CollapseShapeRankReducingSliceSimplificationInfo> info = |
| 164 | getSimplifyCollapseShapeWithRankReducingSliceInfo(sourceType, |
| 165 | reassociationIndices); |
| 166 | if (failed(Result: info)) |
| 167 | return failure(); |
| 168 | |
| 169 | // Create the rank-reducing extract slice op. |
| 170 | auto zero = rewriter.getIndexAttr(value: 0); |
| 171 | auto one = rewriter.getIndexAttr(value: 1); |
| 172 | SmallVector<OpFoldResult> offsets(sourceType.getRank(), zero); |
| 173 | SmallVector<OpFoldResult> sizes = |
| 174 | tensor::getMixedSizes(builder&: rewriter, loc: op.getLoc(), value: op.getSrc()); |
| 175 | SmallVector<OpFoldResult> strides(sourceType.getRank(), one); |
| 176 | auto sliceOp = rewriter.create<tensor::ExtractSliceOp>( |
| 177 | location: op.getLoc(), args&: info->sliceResultType, args: op.getSrc(), args&: offsets, args&: sizes, args&: strides); |
| 178 | |
| 179 | if (!info->newReassociationIndices.has_value()) { |
| 180 | rewriter.replaceOp(op, newValues: sliceOp.getResult()); |
| 181 | return sliceOp.getOperation(); |
| 182 | } |
| 183 | |
| 184 | return rewriter |
| 185 | .replaceOpWithNewOp<tensor::CollapseShapeOp>( |
| 186 | op, args: sliceOp.getResult(), args&: *info->newReassociationIndices) |
| 187 | .getOperation(); |
| 188 | } |
| 189 | |