| 1 | //===- FoldIntoPackAndUnpackPatterns.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/Linalg/IR/Linalg.h" |
| 10 | #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
| 11 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 12 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 13 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 14 | #include "mlir/IR/BuiltinTypeInterfaces.h" |
| 15 | #include "mlir/IR/PatternMatch.h" |
| 16 | |
| 17 | namespace mlir { |
| 18 | namespace linalg { |
| 19 | namespace { |
| 20 | |
| 21 | /// Returns the number of shape sizes that is either dynamic or greater than 1. |
| 22 | static int64_t getNumGtOneDims(ArrayRef<int64_t> shape) { |
| 23 | return llvm::count_if( |
| 24 | Range&: shape, P: [](int64_t v) { return ShapedType::isDynamic(dValue: v) || v > 1; }); |
| 25 | } |
| 26 | |
| 27 | /// Returns the index of the first non-unit size in `sizes`. Returns -1 if |
| 28 | /// there are no non-unit sizes. |
| 29 | static int64_t getFirstNonUnitSizeIdx(ArrayRef<int64_t> sizes) { |
| 30 | const auto *it = llvm::find_if(Range&: sizes, P: [](int64_t dim) { return dim != 1; }); |
| 31 | return (it != sizes.end()) ? std::distance(first: sizes.begin(), last: it) : -1; |
| 32 | } |
| 33 | |
| 34 | /// Check whether `op` is effectively a 1D pack/unpack. Example: |
| 35 | /// |
| 36 | /// %pack = linalg.pack %src |
| 37 | /// inner_dims_pos = [0, 1] |
| 38 | /// inner_tiles = [1, 2] into %dest |
| 39 | /// : tensor<1x32xf32> -> tensor<1x16x1x2xf32> |
| 40 | /// |
| 41 | /// Returns success() if there is: |
| 42 | /// * only 1 non-unit dim in the un-packed domain, |
| 43 | /// * only 1 non-unit inner tile size, and |
| 44 | /// * the unique non-unit tile size is applied to the unique non-unit |
| 45 | /// un-packed dim. |
| 46 | template <typename PackOrUnpackOp> |
| 47 | static LogicalResult isPackOnEffectively1D(RewriterBase &rewriter, |
| 48 | PackOrUnpackOp *op) { |
| 49 | // Obtain the unpacked shape. |
| 50 | auto pack = dyn_cast<linalg::PackOp>(op); |
| 51 | auto unpack = dyn_cast<linalg::UnPackOp>(op); |
| 52 | |
| 53 | ArrayRef<int64_t> unpackedShape = pack ? pack->getSourceType().getShape() |
| 54 | : unpack->getDestType().getShape(); |
| 55 | |
| 56 | // Obtain the inner tile sizes. |
| 57 | ArrayRef<int64_t> innerTileSizes = op->getStaticInnerTiles(); |
| 58 | |
| 59 | // Make sure that there is exactly single non-unit unpacked dim. |
| 60 | if (getNumGtOneDims(shape: unpackedShape) != 1) { |
| 61 | return rewriter.notifyMatchFailure( |
| 62 | *op, "expects non-packed domain to have at most one non-unit dims" ); |
| 63 | } |
| 64 | |
| 65 | // Make sure that there is at most one non-unit inner tile size. |
| 66 | auto numNonUnitInnerTiles = getNumGtOneDims(shape: innerTileSizes); |
| 67 | if (numNonUnitInnerTiles > 1) { |
| 68 | return rewriter.notifyMatchFailure( |
| 69 | *op, "expects at most one non-unit inner tiles" ); |
| 70 | } |
| 71 | |
| 72 | // If there are no non-unit tiles, there is nothing else to check. |
| 73 | if (numNonUnitInnerTiles == 0) |
| 74 | return success(); |
| 75 | |
| 76 | // Get the index of the unique non-unit unpacked dim. |
| 77 | int64_t nonUnitDimIdx = getFirstNonUnitSizeIdx(sizes: unpackedShape); |
| 78 | |
| 79 | // Get the index of the dim that the unique non-unit tile is applied to. |
| 80 | int64_t nonUnitTileDestDimIdx = getFirstNonUnitSizeIdx(sizes: innerTileSizes); |
| 81 | |
| 82 | // Make sure that the unique non-unit tile is applied to the unique unit dim. |
| 83 | if (nonUnitTileDestDimIdx != nonUnitDimIdx) { |
| 84 | return rewriter.notifyMatchFailure( |
| 85 | *op, "expects at most one non-unit inner tiles" ); |
| 86 | } |
| 87 | |
| 88 | return success(); |
| 89 | } |
| 90 | |
| 91 | // If the `linalgOp` represents a transpose, return the permutation vector for |
| 92 | // the transpose. Otherwise, return failure. |
| 93 | static FailureOr<SmallVector<int64_t>> |
| 94 | getTransposeOpPermutation(linalg::LinalgOp linalgOp) { |
| 95 | if (auto transposeOp = dyn_cast<linalg::TransposeOp>(Val: linalgOp.getOperation())) |
| 96 | return SmallVector<int64_t>(transposeOp.getPermutation()); |
| 97 | if (linalgOp.getNumParallelLoops() != linalgOp.getNumLoops()) |
| 98 | return failure(); |
| 99 | |
| 100 | if (linalgOp.getNumDpsInputs() != 1 || linalgOp.getNumDpsInits() != 1) |
| 101 | return failure(); |
| 102 | auto mapRange = linalgOp.getIndexingMapsArray(); |
| 103 | if (!mapRange.front().isPermutation() || !mapRange.back().isPermutation() || |
| 104 | mapRange.front() == mapRange.back()) { |
| 105 | return failure(); |
| 106 | } |
| 107 | if (!llvm::hasSingleElement(C&: linalgOp.getBlock()->getOperations())) |
| 108 | return failure(); |
| 109 | AffineMap outMap = mapRange.back(); |
| 110 | AffineMap inMap = mapRange.front(); |
| 111 | // To get the permutation, look at each output index and find which |
| 112 | // dimension in the input we're reading from for that index. |
| 113 | return llvm::map_to_vector(C: outMap.getResults(), |
| 114 | F: [&](AffineExpr expr) -> int64_t { |
| 115 | return *inMap.getResultPosition(input: expr); |
| 116 | }); |
| 117 | } |
| 118 | |
| 119 | /// Packing one-dimensional tensor can be expressed as an expand shape op. |
| 120 | struct SimplifyPackToExpandShape : public OpRewritePattern<PackOp> { |
| 121 | using OpRewritePattern<PackOp>::OpRewritePattern; |
| 122 | |
| 123 | FailureOr<Value> |
| 124 | insertExpand(RewriterBase &rewriter, Location loc, Value operand, |
| 125 | Type newOperandType, |
| 126 | ArrayRef<ReassociationIndices> reassociation) const { |
| 127 | if (operand.getType() == newOperandType) |
| 128 | return operand; |
| 129 | return tensor::ExpandShapeOp::create(builder&: rewriter, location: loc, resultType: newOperandType, src: operand, |
| 130 | reassociation) |
| 131 | .getResult(); |
| 132 | } |
| 133 | |
| 134 | /// Returns success() if it is only packing on the innermost dimension. |
| 135 | LogicalResult isPackOnInnerMostDim(RewriterBase &rewriter, |
| 136 | PackOp packOp) const { |
| 137 | auto outerDimsPerm = packOp.getOuterDimsPerm(); |
| 138 | if (!outerDimsPerm.empty() && !isIdentityPermutation(permutation: outerDimsPerm)) { |
| 139 | return rewriter.notifyMatchFailure( |
| 140 | arg&: packOp, |
| 141 | msg: "expects outer_dims_perm is empty or an identity permutation" ); |
| 142 | } |
| 143 | |
| 144 | int64_t srcRank = packOp.getSourceRank(); |
| 145 | ArrayRef<int64_t> dimsPos = packOp.getInnerDimsPos(); |
| 146 | if (dimsPos.size() != 1 || (dimsPos[0] + 1 != srcRank)) { |
| 147 | return rewriter.notifyMatchFailure( |
| 148 | arg&: packOp, msg: "expects packing at the innermost dimension" ); |
| 149 | } |
| 150 | return success(); |
| 151 | } |
| 152 | |
| 153 | LogicalResult matchAndRewrite(PackOp packOp, |
| 154 | PatternRewriter &rewriter) const override { |
| 155 | if (packOp.getPaddingValue()) |
| 156 | return rewriter.notifyMatchFailure(arg&: packOp, msg: "expects no padding value" ); |
| 157 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 158 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 159 | // recommended, see #225650 for details. |
| 160 | if (!packOp.hasPureTensorSemantics()) |
| 161 | return failure(); |
| 162 | |
| 163 | ShapedType sourceType = packOp.getSourceType(); |
| 164 | if (failed(Result: isPackOnInnerMostDim(rewriter, packOp)) && |
| 165 | failed(Result: isPackOnEffectively1D(rewriter, op: &packOp)) && |
| 166 | !packOp.isLikePad()) { |
| 167 | return failure(); |
| 168 | } |
| 169 | |
| 170 | ShapedType destType = packOp.getDestType(); |
| 171 | auto reassociation = |
| 172 | getReassociationIndicesForReshape(sourceType, targetType: destType); |
| 173 | if (!reassociation) |
| 174 | return failure(); |
| 175 | FailureOr<Value> expanded = |
| 176 | insertExpand(rewriter, loc: packOp.getLoc(), operand: packOp.getSource(), newOperandType: destType, |
| 177 | reassociation: *reassociation); |
| 178 | if (failed(Result: expanded)) { |
| 179 | return rewriter.notifyMatchFailure( |
| 180 | arg&: packOp, msg: "unable to expand source of tensor.pack" ); |
| 181 | } |
| 182 | rewriter.replaceOp(op: packOp, newValues: *expanded); |
| 183 | return success(); |
| 184 | } |
| 185 | }; |
| 186 | |
| 187 | struct SimplifyUnPackToCollapseShape : public OpRewritePattern<UnPackOp> { |
| 188 | using OpRewritePattern<UnPackOp>::OpRewritePattern; |
| 189 | |
| 190 | Value insertCollapse(RewriterBase &rewriter, Location loc, Value operand, |
| 191 | Type newOperandType, ArrayAttr reassociation) const { |
| 192 | if (operand.getType() == newOperandType) |
| 193 | return operand; |
| 194 | return tensor::CollapseShapeOp::create(builder&: rewriter, location: loc, result: newOperandType, |
| 195 | src: operand, reassociation); |
| 196 | } |
| 197 | |
| 198 | /// Returns success() if it is unpacking on the innermost dimension. |
| 199 | LogicalResult isUnpackOnInnerMostDim(RewriterBase &rewriter, |
| 200 | UnPackOp unpackOp) const { |
| 201 | auto outerDimsPerm = unpackOp.getOuterDimsPerm(); |
| 202 | if (!outerDimsPerm.empty() && !isIdentityPermutation(permutation: outerDimsPerm)) { |
| 203 | return rewriter.notifyMatchFailure( |
| 204 | arg&: unpackOp, |
| 205 | msg: "expects outer_dims_perm is empty or an identity permutation" ); |
| 206 | } |
| 207 | |
| 208 | ShapedType sourceType = unpackOp.getSourceType(); |
| 209 | ShapedType destType = unpackOp.getDestType(); |
| 210 | if (!sourceType.hasStaticShape() || !destType.hasStaticShape()) |
| 211 | return rewriter.notifyMatchFailure(arg&: unpackOp, msg: "expects static shapes" ); |
| 212 | |
| 213 | ArrayRef<int64_t> dimsPos = unpackOp.getInnerDimsPos(); |
| 214 | if (dimsPos.size() != 1 || (dimsPos[0] + 1 != destType.getRank())) { |
| 215 | return rewriter.notifyMatchFailure( |
| 216 | arg&: unpackOp, msg: "expects unpacking on the innermost dimension" ); |
| 217 | } |
| 218 | |
| 219 | return success(); |
| 220 | } |
| 221 | |
| 222 | LogicalResult matchAndRewrite(UnPackOp unpackOp, |
| 223 | PatternRewriter &rewriter) const override { |
| 224 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 225 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 226 | // recommended, see #225650 for details. |
| 227 | if (!unpackOp.hasPureTensorSemantics()) |
| 228 | return failure(); |
| 229 | |
| 230 | ShapedType destType = unpackOp.getDestType(); |
| 231 | if (failed(Result: isUnpackOnInnerMostDim(rewriter, unpackOp)) && |
| 232 | failed(Result: isPackOnEffectively1D(rewriter, op: &unpackOp)) && |
| 233 | !unpackOp.isLikeUnPad()) { |
| 234 | return failure(); |
| 235 | } |
| 236 | |
| 237 | ShapedType sourceType = unpackOp.getSourceType(); |
| 238 | auto reassociation = |
| 239 | getReassociationIndicesForReshape(sourceType, targetType: destType); |
| 240 | if (!reassociation) |
| 241 | return failure(); |
| 242 | Value collapsed = insertCollapse( |
| 243 | rewriter, loc: unpackOp.getLoc(), operand: unpackOp.getSource(), newOperandType: destType, |
| 244 | reassociation: getReassociationIndicesAttribute(b&: rewriter, reassociation: *reassociation)); |
| 245 | rewriter.replaceOp(op: unpackOp, newValues: collapsed); |
| 246 | return success(); |
| 247 | } |
| 248 | }; |
| 249 | |
| 250 | /// Fold a `pad` -> `pack` into `pack` if they have the same padding values and |
| 251 | /// the pad op has zero low paddings, or if `pack` has no padding values. |
| 252 | struct FoldPadWithPackOp : public OpRewritePattern<PackOp> { |
| 253 | public: |
| 254 | FoldPadWithPackOp(MLIRContext *context, ControlFoldIntoPackUnpackFn controlFn) |
| 255 | : OpRewritePattern<PackOp>(context), controlFn(std::move(controlFn)) {} |
| 256 | |
| 257 | LogicalResult matchAndRewrite(PackOp packOp, |
| 258 | PatternRewriter &rewriter) const override { |
| 259 | auto padOp = packOp.getSource().getDefiningOp<tensor::PadOp>(); |
| 260 | |
| 261 | if (!padOp || padOp.getNofold() || !padOp.hasZeroLowPad()) |
| 262 | return failure(); |
| 263 | |
| 264 | // User controlled folding function. |
| 265 | if (controlFn && !controlFn(&packOp.getSourceMutable())) |
| 266 | return failure(); |
| 267 | |
| 268 | Value constantPaddingValue = padOp.getConstantPaddingValue(); |
| 269 | if (!constantPaddingValue) |
| 270 | return failure(); |
| 271 | |
| 272 | if (auto paddingValue = packOp.getPaddingValue()) |
| 273 | if (!isEqualConstantIntOrValue(ofr1: paddingValue, ofr2: constantPaddingValue)) |
| 274 | return failure(); |
| 275 | |
| 276 | // Folding is not allowed if it were to introduce artificial padding. |
| 277 | // Folding is also disabled in the case of dynamic dimensions and/or tile |
| 278 | // sizes - that is because it would be impossible to compute the padding |
| 279 | // size and hence to establish whether "artificial" padding would be |
| 280 | // created. |
| 281 | ShapedType unpackedType = packOp.getSourceType(); |
| 282 | SmallVector<int64_t> outerShapeWithoutTranspose = |
| 283 | getPackedOuterShapeWithoutTransposition(packOrUnPack: packOp); |
| 284 | for (auto [pos, tileSize, high] : |
| 285 | llvm::zip_equal(t: packOp.getInnerDimsPos(), u: packOp.getStaticInnerTiles(), |
| 286 | args: padOp.getMixedHighPad())) { |
| 287 | if (unpackedType.isDynamicDim(idx: pos)) |
| 288 | return failure(); |
| 289 | if (ShapedType::isDynamic(dValue: outerShapeWithoutTranspose[pos])) |
| 290 | return failure(); |
| 291 | if (ShapedType::isDynamic(dValue: tileSize)) |
| 292 | return failure(); |
| 293 | std::optional<int64_t> cstHigh = getConstantIntValue(ofr: high); |
| 294 | if (!cstHigh) |
| 295 | return failure(); |
| 296 | int64_t paddingSize = outerShapeWithoutTranspose[pos] * tileSize - |
| 297 | unpackedType.getDimSize(idx: pos); |
| 298 | // Do not fold the op if it requires artificial padding. |
| 299 | if (paddingSize + cstHigh.value() >= tileSize) |
| 300 | return failure(); |
| 301 | } |
| 302 | |
| 303 | rewriter.replaceOpWithNewOp<PackOp>( |
| 304 | op: packOp, args: padOp.getSource(), args: packOp.getDest(), args: packOp.getInnerDimsPos(), |
| 305 | args: packOp.getMixedTiles(), args&: constantPaddingValue, |
| 306 | args: packOp.getOuterDimsPerm()); |
| 307 | return success(); |
| 308 | } |
| 309 | |
| 310 | private: |
| 311 | ControlFoldIntoPackUnpackFn controlFn; |
| 312 | }; |
| 313 | |
| 314 | /// Fold a `unpack` -> `extract_slice` into the `unpack` since it already |
| 315 | /// has extract_slice semantics. |
| 316 | struct |
| 317 | : public OpRewritePattern<tensor::ExtractSliceOp> { |
| 318 | public: |
| 319 | FoldUnpackWithExtractSliceOp(MLIRContext *context, |
| 320 | ControlFoldIntoPackUnpackFn controlFn) |
| 321 | : OpRewritePattern<tensor::ExtractSliceOp>(context), |
| 322 | controlFn(std::move(controlFn)) {} |
| 323 | |
| 324 | LogicalResult matchAndRewrite(tensor::ExtractSliceOp sliceOp, |
| 325 | PatternRewriter &rewriter) const override { |
| 326 | auto unpackOp = sliceOp.getSource().getDefiningOp<UnPackOp>(); |
| 327 | if (!unpackOp) |
| 328 | return failure(); |
| 329 | |
| 330 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 331 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 332 | // recommended, see #225650 for details. |
| 333 | if (!unpackOp.hasPureTensorSemantics()) |
| 334 | return failure(); |
| 335 | |
| 336 | // User controlled folding function. |
| 337 | if (controlFn && !controlFn(&sliceOp.getSourceMutable())) |
| 338 | return failure(); |
| 339 | |
| 340 | if (!unpackOp.canFoldSliceOp(sliceOp)) |
| 341 | return failure(); |
| 342 | |
| 343 | // Create a new empty output tensor. |
| 344 | Type elementType = unpackOp.getDestType().getElementType(); |
| 345 | Value output = tensor::EmptyOp::create( |
| 346 | builder&: rewriter, location: sliceOp.getLoc(), sizes: sliceOp.getMixedSizes(), elementType); |
| 347 | rewriter.replaceOpWithNewOp<UnPackOp>( |
| 348 | op: sliceOp, args: unpackOp.getSource(), args&: output, args: unpackOp.getInnerDimsPos(), |
| 349 | args: unpackOp.getMixedTiles(), args: unpackOp.getOuterDimsPerm()); |
| 350 | return success(); |
| 351 | } |
| 352 | |
| 353 | private: |
| 354 | ControlFoldIntoPackUnpackFn ; |
| 355 | }; |
| 356 | |
| 357 | // Applies 'permutation' on 'inVec' and stores the result in resVec. |
| 358 | // 'inVec' may be empty, in that case it's one-to-one mapping with permutation. |
| 359 | // `rank` sets the boundary for permutation i.e., the permutation dim can't be |
| 360 | // greater than the rank specified. If it's so then return false. |
| 361 | // For e.g., permutation {1, 0, 3, 2} with rank 2 is allowed since the values in |
| 362 | // permutation[:rank] doesn't exceed rank, whereas, permutation {1, 3, 0, 2} is |
| 363 | // not allowed since `3` exceeds the value of the rank in the given range. |
| 364 | static bool checkAndPermute(ArrayRef<int64_t> permutation, |
| 365 | ArrayRef<int64_t> inVec, |
| 366 | SmallVectorImpl<int64_t> &resVec, int64_t rank) { |
| 367 | |
| 368 | for (unsigned int i = 0; i < rank; ++i) { |
| 369 | int64_t remappedPosition = permutation[i]; |
| 370 | if (remappedPosition >= rank) |
| 371 | return false; |
| 372 | if (!inVec.empty()) |
| 373 | remappedPosition = inVec[remappedPosition]; |
| 374 | resVec.push_back(Elt: remappedPosition); |
| 375 | } |
| 376 | |
| 377 | return true; |
| 378 | } |
| 379 | |
| 380 | /// Fold 'pack' -> 'transpose' into 'pack' since 'pack' already has transpose |
| 381 | /// semantics. |
| 382 | struct FoldProducerPackWithConsumerLinalgTransposeOp |
| 383 | : public OpInterfaceRewritePattern<linalg::LinalgOp> { |
| 384 | |
| 385 | public: |
| 386 | FoldProducerPackWithConsumerLinalgTransposeOp( |
| 387 | MLIRContext *context, ControlFoldIntoPackUnpackFn controlFn) |
| 388 | : OpInterfaceRewritePattern<linalg::LinalgOp>(context), |
| 389 | controlFn(std::move(controlFn)) {} |
| 390 | |
| 391 | LogicalResult matchAndRewrite(linalg::LinalgOp linalgOp, |
| 392 | PatternRewriter &rewriter) const override { |
| 393 | auto packOp = linalgOp->getOperand(idx: 0).getDefiningOp<PackOp>(); |
| 394 | |
| 395 | if (!packOp) |
| 396 | return failure(); |
| 397 | |
| 398 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 399 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 400 | // recommended, see #225650 for details. |
| 401 | if (!packOp.hasPureTensorSemantics()) |
| 402 | return failure(); |
| 403 | |
| 404 | // User controlled folding function. |
| 405 | if (controlFn && !controlFn(&linalgOp->getOpOperand(idx: 0))) |
| 406 | return failure(); |
| 407 | |
| 408 | FailureOr<SmallVector<int64_t>> maybePerm = |
| 409 | getTransposeOpPermutation(linalgOp); |
| 410 | if (failed(Result: maybePerm)) |
| 411 | return failure(); |
| 412 | |
| 413 | auto innerDimsPos = packOp.getInnerDimsPos(); |
| 414 | auto mixedInnerTiles = packOp.getMixedTiles(); |
| 415 | auto outerDimsPerm = packOp.getOuterDimsPerm(); |
| 416 | const auto &transposePerm = maybePerm.value(); |
| 417 | SmallVector<int64_t> newOuterDimsPermVec; |
| 418 | SmallVector<int64_t> newInnerDimsPosVec; |
| 419 | SmallVector<OpFoldResult> newMixedInnerTilesVec; |
| 420 | int64_t srcRank = packOp.getSourceRank(); |
| 421 | |
| 422 | if (!checkAndPermute(permutation: transposePerm, inVec: outerDimsPerm, resVec&: newOuterDimsPermVec, |
| 423 | rank: srcRank)) |
| 424 | return rewriter.notifyMatchFailure( |
| 425 | arg&: linalgOp, |
| 426 | msg: "Cannot fold in tensor.pack if a tile dimension was transposed " |
| 427 | "with a non-tile dimension in linalg.transpose." ); |
| 428 | |
| 429 | // Process transpose operation for tiled inner dimensions |
| 430 | for (unsigned int i = srcRank; i < transposePerm.size(); ++i) { |
| 431 | int64_t remappedPosition = transposePerm[i] - srcRank; |
| 432 | newMixedInnerTilesVec.push_back(Elt: mixedInnerTiles[remappedPosition]); |
| 433 | newInnerDimsPosVec.push_back(Elt: innerDimsPos[remappedPosition]); |
| 434 | } |
| 435 | |
| 436 | Value output = packOp.createDestinationTensor( |
| 437 | b&: rewriter, loc: linalgOp.getLoc(), source: packOp.getSource(), innerTileSizes: newMixedInnerTilesVec, |
| 438 | innerDimsPos: newInnerDimsPosVec, outerDimsPerm: newOuterDimsPermVec); |
| 439 | |
| 440 | rewriter.replaceOpWithNewOp<PackOp>( |
| 441 | op: linalgOp, args: packOp.getSource(), args&: output, args&: newInnerDimsPosVec, |
| 442 | args&: newMixedInnerTilesVec, args: packOp.getPaddingValue(), args&: newOuterDimsPermVec); |
| 443 | |
| 444 | return success(); |
| 445 | } |
| 446 | |
| 447 | private: |
| 448 | ControlFoldIntoPackUnpackFn controlFn; |
| 449 | }; |
| 450 | |
| 451 | /// Fold 'transpose' -> 'pack' into 'pack' since 'pack' already has transpose |
| 452 | /// semantics. |
| 453 | struct FoldConsumerPackWithProducerLinalgTransposeOp |
| 454 | : public OpRewritePattern<PackOp> { |
| 455 | |
| 456 | public: |
| 457 | FoldConsumerPackWithProducerLinalgTransposeOp( |
| 458 | MLIRContext *context, ControlFoldIntoPackUnpackFn controlFn) |
| 459 | : OpRewritePattern<PackOp>(context), controlFn(std::move(controlFn)) {} |
| 460 | |
| 461 | LogicalResult matchAndRewrite(PackOp packOp, |
| 462 | PatternRewriter &rewriter) const override { |
| 463 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 464 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 465 | // recommended, see #225650 for details. |
| 466 | if (!packOp.hasPureTensorSemantics()) |
| 467 | return failure(); |
| 468 | |
| 469 | auto linalgOp = packOp.getSource().getDefiningOp<linalg::LinalgOp>(); |
| 470 | if (!linalgOp) |
| 471 | return failure(); |
| 472 | |
| 473 | // User controlled folding function. |
| 474 | if (controlFn && !controlFn(&packOp.getSourceMutable())) |
| 475 | return failure(); |
| 476 | |
| 477 | FailureOr<SmallVector<int64_t>> maybePerm = |
| 478 | getTransposeOpPermutation(linalgOp); |
| 479 | if (failed(Result: maybePerm)) |
| 480 | return failure(); |
| 481 | |
| 482 | auto transposePermutation = maybePerm.value(); |
| 483 | auto outerDimsPerm = packOp.getOuterDimsPerm(); |
| 484 | auto innerDimsPos = packOp.getInnerDimsPos(); |
| 485 | SmallVector<int64_t> newInnerDimsPosVec; |
| 486 | SmallVector<int64_t> newOuterDimsPermVec = |
| 487 | llvm::to_vector(Range&: transposePermutation); |
| 488 | |
| 489 | if (!outerDimsPerm.empty()) |
| 490 | applyPermutationToVector(inVec&: newOuterDimsPermVec, permutation: outerDimsPerm); |
| 491 | |
| 492 | // Can't use applyPermutationToVector for newInnerDimsPosVec since input and |
| 493 | // permutation rank won't necessarily be equal in all cases. |
| 494 | for (auto dim : innerDimsPos) |
| 495 | newInnerDimsPosVec.push_back(Elt: transposePermutation[dim]); |
| 496 | |
| 497 | Value output = packOp.createDestinationTensor( |
| 498 | b&: rewriter, loc: packOp.getLoc(), source: linalgOp->getOperand(idx: 0), |
| 499 | innerTileSizes: packOp.getMixedTiles(), innerDimsPos: newInnerDimsPosVec, outerDimsPerm: newOuterDimsPermVec); |
| 500 | |
| 501 | rewriter.replaceOpWithNewOp<PackOp>( |
| 502 | op: packOp, args: linalgOp->getOperand(idx: 0), args&: output, args&: newInnerDimsPosVec, |
| 503 | args: packOp.getMixedTiles(), args: packOp.getPaddingValue(), args&: newOuterDimsPermVec); |
| 504 | |
| 505 | return success(); |
| 506 | } |
| 507 | |
| 508 | private: |
| 509 | ControlFoldIntoPackUnpackFn controlFn; |
| 510 | }; |
| 511 | |
| 512 | /// Fold 'unpack' -> 'transpose' into 'unpack' since 'unpack' already has |
| 513 | /// transpose semantics. |
| 514 | struct FoldProducerUnPackWithConsumerLinalgTransposeOp |
| 515 | : public OpInterfaceRewritePattern<linalg::LinalgOp> { |
| 516 | |
| 517 | public: |
| 518 | FoldProducerUnPackWithConsumerLinalgTransposeOp( |
| 519 | MLIRContext *context, ControlFoldIntoPackUnpackFn controlFn) |
| 520 | : OpInterfaceRewritePattern<linalg::LinalgOp>(context), |
| 521 | controlFn(std::move(controlFn)) {} |
| 522 | |
| 523 | LogicalResult matchAndRewrite(linalg::LinalgOp linalgOp, |
| 524 | PatternRewriter &rewriter) const override { |
| 525 | auto unPackOp = linalgOp->getOperand(idx: 0).getDefiningOp<UnPackOp>(); |
| 526 | |
| 527 | if (!unPackOp) |
| 528 | return failure(); |
| 529 | |
| 530 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 531 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 532 | // recommended, see #225650 for details. |
| 533 | if (!unPackOp.hasPureTensorSemantics()) |
| 534 | return failure(); |
| 535 | |
| 536 | // User controlled folding function. |
| 537 | if (controlFn && !controlFn(&linalgOp->getOpOperand(idx: 0))) |
| 538 | return failure(); |
| 539 | |
| 540 | FailureOr<SmallVector<int64_t>> maybePerm = |
| 541 | getTransposeOpPermutation(linalgOp); |
| 542 | if (failed(Result: maybePerm)) |
| 543 | return failure(); |
| 544 | |
| 545 | auto outerDimsPerm = unPackOp.getOuterDimsPerm(); |
| 546 | auto innerDimsPos = unPackOp.getInnerDimsPos(); |
| 547 | SmallVector<int64_t> newInnerDimsPosVec; |
| 548 | SmallVector<int64_t> newOuterDimsPermVec = |
| 549 | invertPermutationVector(permutation: maybePerm.value()); |
| 550 | |
| 551 | // Can't use applyPermutationToVector for newInnerDimsPosVec since input and |
| 552 | // permutation rank won't necessarily be equal in all cases. |
| 553 | for (auto dim : innerDimsPos) |
| 554 | newInnerDimsPosVec.push_back(Elt: newOuterDimsPermVec[dim]); |
| 555 | |
| 556 | if (!outerDimsPerm.empty()) |
| 557 | applyPermutationToVector(inVec&: newOuterDimsPermVec, permutation: outerDimsPerm); |
| 558 | |
| 559 | // Reuse the destination of the transpose op. |
| 560 | rewriter.replaceOpWithNewOp<UnPackOp>( |
| 561 | op: linalgOp, args: unPackOp.getSource(), args: linalgOp.getDpsInits()[0], |
| 562 | args&: newInnerDimsPosVec, args: unPackOp.getMixedTiles(), args&: newOuterDimsPermVec); |
| 563 | |
| 564 | return success(); |
| 565 | } |
| 566 | |
| 567 | private: |
| 568 | ControlFoldIntoPackUnpackFn controlFn; |
| 569 | }; |
| 570 | |
| 571 | /// Fold 'transpose' -> 'unpack' into 'unpack' since 'unpack' already has |
| 572 | /// transpose semantics. |
| 573 | struct FoldConsumerUnPackWithProducerLinalgTransposeOp |
| 574 | : public OpRewritePattern<UnPackOp> { |
| 575 | using OpRewritePattern<UnPackOp>::OpRewritePattern; |
| 576 | |
| 577 | public: |
| 578 | FoldConsumerUnPackWithProducerLinalgTransposeOp( |
| 579 | MLIRContext *context, ControlFoldIntoPackUnpackFn controlFn) |
| 580 | : OpRewritePattern<UnPackOp>(context), controlFn(std::move(controlFn)) {} |
| 581 | |
| 582 | LogicalResult matchAndRewrite(UnPackOp unPackOp, |
| 583 | PatternRewriter &rewriter) const override { |
| 584 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 585 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 586 | // recommended, see #225650 for details. |
| 587 | if (!unPackOp.hasPureTensorSemantics()) |
| 588 | return failure(); |
| 589 | |
| 590 | auto linalgOp = unPackOp.getSource().getDefiningOp<linalg::LinalgOp>(); |
| 591 | if (!linalgOp) |
| 592 | return failure(); |
| 593 | |
| 594 | // User controlled folding function. |
| 595 | if (controlFn && !controlFn(&unPackOp.getSourceMutable())) |
| 596 | return failure(); |
| 597 | |
| 598 | FailureOr<SmallVector<int64_t>> maybePerm = |
| 599 | getTransposeOpPermutation(linalgOp); |
| 600 | if (failed(Result: maybePerm)) |
| 601 | return failure(); |
| 602 | |
| 603 | SmallVector<SmallVector<OpFoldResult>> unpackOpResultDims; |
| 604 | if (failed(Result: reifyResultShapes(b&: rewriter, op: unPackOp, reifiedReturnShapes&: unpackOpResultDims))) { |
| 605 | return failure(); |
| 606 | } |
| 607 | |
| 608 | SmallVector<int64_t> inverseTransposePerm = |
| 609 | invertPermutationVector(permutation: maybePerm.value()); |
| 610 | auto outerDimsPerm = unPackOp.getOuterDimsPerm(); |
| 611 | auto innerDimsPos = unPackOp.getInnerDimsPos(); |
| 612 | int64_t destRank = unPackOp.getSourceRank() - innerDimsPos.size(); |
| 613 | auto mixedInnerTilesVec = unPackOp.getMixedTiles(); |
| 614 | SmallVector<int64_t> newOuterDimsPermVec; |
| 615 | SmallVector<int64_t> newInnerDimsPosVec; |
| 616 | SmallVector<OpFoldResult> newMixedInnerTilesVec; |
| 617 | if (!checkAndPermute(permutation: inverseTransposePerm, inVec: outerDimsPerm, |
| 618 | resVec&: newOuterDimsPermVec, rank: destRank)) |
| 619 | return rewriter.notifyMatchFailure( |
| 620 | arg&: unPackOp, |
| 621 | msg: "Cannot fold in tensor.unpack if a tile dimension was transposed " |
| 622 | "with a non-tile dimension in linalg.transpose." ); |
| 623 | |
| 624 | // Process transpose operation for tiled inner dimensions |
| 625 | for (unsigned int i = destRank; i < inverseTransposePerm.size(); ++i) { |
| 626 | int64_t remappedPosition = inverseTransposePerm[i] - destRank; |
| 627 | newMixedInnerTilesVec.push_back(Elt: mixedInnerTilesVec[remappedPosition]); |
| 628 | newInnerDimsPosVec.push_back(Elt: innerDimsPos[remappedPosition]); |
| 629 | } |
| 630 | |
| 631 | auto elemType = |
| 632 | cast<ShapedType>(Val: unPackOp->getResultTypes()[0]).getElementType(); |
| 633 | Value output = tensor::EmptyOp::create(builder&: rewriter, location: unPackOp->getLoc(), |
| 634 | sizes: unpackOpResultDims[0], elementType: elemType); |
| 635 | |
| 636 | rewriter.replaceOpWithNewOp<UnPackOp>( |
| 637 | op: unPackOp, args: linalgOp->getOperand(idx: 0), args&: output, args&: newInnerDimsPosVec, |
| 638 | args&: newMixedInnerTilesVec, args&: newOuterDimsPermVec); |
| 639 | |
| 640 | return success(); |
| 641 | } |
| 642 | |
| 643 | private: |
| 644 | ControlFoldIntoPackUnpackFn controlFn; |
| 645 | }; |
| 646 | |
| 647 | /// tensor.empty does not define any tensor contents, so an unpadded pack |
| 648 | /// can be folded away. |
| 649 | struct FoldEmptyTensorWithPackOp : public OpRewritePattern<PackOp> { |
| 650 | using OpRewritePattern<PackOp>::OpRewritePattern; |
| 651 | |
| 652 | LogicalResult matchAndRewrite(PackOp packOp, |
| 653 | PatternRewriter &rewriter) const override { |
| 654 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 655 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 656 | // recommended, see #225650 for details. |
| 657 | if (!packOp.hasPureTensorSemantics()) |
| 658 | return failure(); |
| 659 | |
| 660 | // Check for tensor.empty source. |
| 661 | auto emptyOp = packOp.getSource().getDefiningOp<tensor::EmptyOp>(); |
| 662 | if (!emptyOp) |
| 663 | return failure(); |
| 664 | |
| 665 | // Check for padding. |
| 666 | // Packing with padding cannot be simply removed. |
| 667 | if (packOp.getPaddingValue()) |
| 668 | return rewriter.notifyMatchFailure(arg&: packOp, msg: "expects no padding value" ); |
| 669 | |
| 670 | // Replace the pack directly with its destination. |
| 671 | rewriter.replaceOp(op: packOp, newValues: packOp.getDest()); |
| 672 | |
| 673 | return success(); |
| 674 | } |
| 675 | }; |
| 676 | |
| 677 | /// tensor.empty does not define any tensor contents, so an unpack |
| 678 | /// can be folded away. |
| 679 | struct FoldEmptyTensorWithUnPackOp : public OpRewritePattern<UnPackOp> { |
| 680 | using OpRewritePattern<UnPackOp>::OpRewritePattern; |
| 681 | |
| 682 | LogicalResult matchAndRewrite(UnPackOp unPackOp, |
| 683 | PatternRewriter &rewriter) const override { |
| 684 | // Pack/unpack memref transformations are unsupported. The memref forms |
| 685 | // are mainly for bufferization and scalar lowering. Other uses are not |
| 686 | // recommended, see #225650 for details. |
| 687 | if (!unPackOp.hasPureTensorSemantics()) |
| 688 | return failure(); |
| 689 | |
| 690 | // Check for tensor.empty source. |
| 691 | auto emptyOp = unPackOp.getSource().getDefiningOp<tensor::EmptyOp>(); |
| 692 | if (!emptyOp) |
| 693 | return failure(); |
| 694 | |
| 695 | // Replace the unpack directly with its destination. |
| 696 | rewriter.replaceOp(op: unPackOp, newValues: unPackOp.getDest()); |
| 697 | |
| 698 | return success(); |
| 699 | } |
| 700 | }; |
| 701 | |
| 702 | } // namespace |
| 703 | |
| 704 | void populateFoldIntoPackAndUnpackPatterns( |
| 705 | RewritePatternSet &patterns, const ControlFoldIntoPackUnpackFn &controlFn) { |
| 706 | patterns.insert<FoldUnpackWithExtractSliceOp, FoldPadWithPackOp, |
| 707 | FoldProducerPackWithConsumerLinalgTransposeOp, |
| 708 | FoldConsumerPackWithProducerLinalgTransposeOp, |
| 709 | FoldConsumerUnPackWithProducerLinalgTransposeOp, |
| 710 | FoldProducerUnPackWithConsumerLinalgTransposeOp>( |
| 711 | arg: patterns.getContext(), args: controlFn); |
| 712 | } |
| 713 | |
| 714 | void populateSimplifyPackAndUnpackPatterns(RewritePatternSet &patterns) { |
| 715 | patterns.add<SimplifyPackToExpandShape, SimplifyUnPackToCollapseShape>( |
| 716 | arg: patterns.getContext()); |
| 717 | } |
| 718 | |
| 719 | void populateFoldPackUnpackIntoTensorEmptyPatterns( |
| 720 | RewritePatternSet &patterns) { |
| 721 | patterns.add<FoldEmptyTensorWithPackOp, FoldEmptyTensorWithUnPackOp>( |
| 722 | arg: patterns.getContext()); |
| 723 | } |
| 724 | |
| 725 | } // namespace linalg |
| 726 | } // namespace mlir |
| 727 | |