| 1 | //===- ConvertConv2DToImg2Col.cpp - im2col implementation -----------------===// |
| 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/Affine/Utils.h" |
| 10 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 11 | #include "mlir/Dialect/Complex/IR/Complex.h" |
| 12 | #include "mlir/Dialect/Linalg/IR/Linalg.h" |
| 13 | #include "mlir/Dialect/Linalg/Transforms/Transforms.h" |
| 14 | #include "mlir/Dialect/Tensor/IR/Tensor.h" |
| 15 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 16 | #include "mlir/Dialect/Utils/ReshapeOpsUtils.h" |
| 17 | #include "mlir/Dialect/Utils/StructuredOpsUtils.h" |
| 18 | #include "mlir/IR/AffineExpr.h" |
| 19 | #include "mlir/IR/AffineMap.h" |
| 20 | #include "mlir/IR/Builders.h" |
| 21 | #include "mlir/IR/BuiltinAttributes.h" |
| 22 | #include "mlir/IR/BuiltinTypes.h" |
| 23 | #include "llvm/ADT/SmallVectorExtras.h" |
| 24 | #include <cassert> |
| 25 | #include <utility> |
| 26 | |
| 27 | namespace mlir { |
| 28 | namespace linalg { |
| 29 | static bool hasAllOneValues(DenseIntElementsAttr attr) { |
| 30 | return llvm::all_of( |
| 31 | Range&: attr, P: [](const APInt &element) { return element.getSExtValue() == 1; }); |
| 32 | } |
| 33 | |
| 34 | static Value createAdd(Location loc, Value x, Value y, OpBuilder &builder) { |
| 35 | if (isa<IntegerType>(Val: x.getType())) |
| 36 | return arith::AddIOp::create(builder, location: loc, lhs: x, rhs: y); |
| 37 | if (isa<ComplexType>(Val: x.getType())) |
| 38 | return complex::AddOp::create(builder, location: loc, lhs: x, rhs: y); |
| 39 | return arith::AddFOp::create(builder, location: loc, lhs: x, rhs: y); |
| 40 | } |
| 41 | |
| 42 | static Value createMul(Location loc, Value x, Value y, Type accType, |
| 43 | OpBuilder &builder) { |
| 44 | // Linalg named ops specify signed extend for named ops. |
| 45 | Value xConvert = |
| 46 | convertScalarToDtype(b&: builder, loc, operand: x, toType: accType, /*isUnsignedCast=*/false); |
| 47 | Value yConvert = |
| 48 | convertScalarToDtype(b&: builder, loc, operand: y, toType: accType, /*isUnsignedCast=*/false); |
| 49 | if (isa<ComplexType>(Val: accType)) |
| 50 | return complex::MulOp::create(builder, location: loc, lhs: xConvert, rhs: yConvert); |
| 51 | if (isa<IntegerType>(Val: accType)) |
| 52 | return arith::MulIOp::create(builder, location: loc, lhs: xConvert, rhs: yConvert); |
| 53 | return arith::MulFOp::create(builder, location: loc, lhs: xConvert, rhs: yConvert); |
| 54 | } |
| 55 | |
| 56 | // Generate the affine expression to compute the convolved index |
| 57 | // for the input as `oIndex * stride + fIndex * dilation`, |
| 58 | // where oIndex: output iterator; fIndex: filter iterator. |
| 59 | static AffineExpr getConvolvedExpr(OpBuilder &b, int64_t stride, |
| 60 | int64_t dilation, bool useSymbols = true) { |
| 61 | AffineExpr oExpr, fExpr; |
| 62 | if (useSymbols) |
| 63 | bindSymbols(ctx: b.getContext(), exprs&: oExpr, exprs&: fExpr); |
| 64 | else |
| 65 | bindDims(ctx: b.getContext(), exprs&: oExpr, exprs&: fExpr); |
| 66 | return AffineExpr(stride * oExpr + dilation * fExpr); |
| 67 | } |
| 68 | |
| 69 | // Stores the affine expressions to map the iteration space of the im2col matrix |
| 70 | // to the corresponding indices of the output and filter matrices |
| 71 | struct Im2ColToOperandsExprs { |
| 72 | AffineExpr fhIndex; |
| 73 | AffineExpr fwIndex; |
| 74 | AffineExpr icIndex; |
| 75 | AffineExpr ohIndex; |
| 76 | AffineExpr owIndex; |
| 77 | }; |
| 78 | |
| 79 | // Stores the affine expressions to map the iteration space of the im2col matrix |
| 80 | // to the input matrix indices |
| 81 | struct Im2ColToInputDimsExprs { |
| 82 | AffineExpr bIndex; |
| 83 | AffineExpr hIndex; |
| 84 | AffineExpr wIndex; |
| 85 | AffineExpr cIndex; |
| 86 | }; |
| 87 | |
| 88 | /// Construct the affine expressions that map the indices of the im2col matrix |
| 89 | /// to the corresponding input tensor indices for a 2D convolution with the the |
| 90 | /// provided strides. |
| 91 | /// |
| 92 | /// @param exprs Affine expressions for output and filter indices. |
| 93 | /// @param strides [height, width] stride values for the convolution. |
| 94 | /// @param dilations [height, width] dilation values for the convolution. |
| 95 | /// @param rewriter Pattern rewriter. |
| 96 | /// @return Affine expressions mapping im2col matrix indices to input |
| 97 | /// offsets. |
| 98 | static Im2ColToInputDimsExprs |
| 99 | getIm2ColInputExpressions(Im2ColToOperandsExprs exprs, |
| 100 | ArrayRef<int64_t> strides, |
| 101 | ArrayRef<int64_t> dilations, RewriterBase &rewriter) { |
| 102 | // maps the iteration space of the im2col matrix to (output_y, filter_y) |
| 103 | auto hIndicesMap = AffineMap::inferFromExprList( |
| 104 | exprsList: {ArrayRef{exprs.ohIndex, exprs.fhIndex}}, context: rewriter.getContext())[0]; |
| 105 | // maps the iteration space of the im2col matrix to (output_x, filter_x) |
| 106 | auto wIndicesMap = AffineMap::inferFromExprList( |
| 107 | exprsList: {ArrayRef{exprs.owIndex, exprs.fwIndex}}, context: rewriter.getContext())[0]; |
| 108 | // Compute the input indexing map, to map the indices of the im2col matrix to |
| 109 | // the original input offsets. Each element of the im2col matrix corresponds |
| 110 | // to a pair of (out_element, filter_element). First, we build the expressions |
| 111 | // to compute the input (ix, iy) indices from [out_x/y, filter_x/y] pairs; |
| 112 | // then we compose them with the maps that map the im2col matrix elements to |
| 113 | // the (out_element, filter_element) pairs. |
| 114 | auto bIndexExpr = rewriter.getAffineDimExpr(position: 0U); |
| 115 | auto hIndexExpr = getConvolvedExpr(b&: rewriter, stride: strides[0], dilation: dilations[0], |
| 116 | /*useSymbols*/ false); |
| 117 | hIndexExpr = hIndexExpr.compose(map: hIndicesMap); |
| 118 | auto wIndexExpr = getConvolvedExpr(b&: rewriter, stride: strides[1], dilation: dilations[1], |
| 119 | /*useSymbols*/ false); |
| 120 | wIndexExpr = wIndexExpr.compose(map: wIndicesMap); |
| 121 | auto cIndexExpr = exprs.icIndex; |
| 122 | return {.bIndex: bIndexExpr, .hIndex: hIndexExpr, .wIndex: wIndexExpr, .cIndex: cIndexExpr}; |
| 123 | } |
| 124 | |
| 125 | FailureOr<std::pair<Operation *, Operation *>> |
| 126 | rewriteInIm2Col(RewriterBase &rewriter, linalg::Conv2DNhwcHwcfOp convOp) { |
| 127 | auto inputType = cast<ShapedType>(Val: convOp.getInputs()[0].getType()); |
| 128 | auto filterType = cast<ShapedType>(Val: convOp.getInputs()[1].getType()); |
| 129 | auto outputType = cast<ShapedType>(Val: convOp.getOutputs()[0].getType()); |
| 130 | |
| 131 | if (!convOp.hasPureTensorSemantics()) |
| 132 | return rewriter.notifyMatchFailure( |
| 133 | arg&: convOp, msg: "expected op to have pure tensor semantics" ); |
| 134 | |
| 135 | if (!filterType.hasStaticShape()) |
| 136 | return rewriter.notifyMatchFailure( |
| 137 | arg&: convOp, msg: "expected a static shape for the filter" ); |
| 138 | |
| 139 | if (!inputType.hasStaticShape()) |
| 140 | return rewriter.notifyMatchFailure(arg&: convOp, |
| 141 | msg: "expected a static shape for the input" ); |
| 142 | |
| 143 | MLIRContext *context = rewriter.getContext(); |
| 144 | Value input = convOp.getInputs()[0]; |
| 145 | Value filter = convOp.getInputs()[1]; |
| 146 | Value output = convOp.getOutputs()[0]; |
| 147 | |
| 148 | ArrayRef<int64_t> filterShape = filterType.getShape(); |
| 149 | ArrayRef<int64_t> outputShape = outputType.getShape(); |
| 150 | |
| 151 | int64_t n = outputShape[0]; |
| 152 | int64_t oh = outputShape[1]; |
| 153 | int64_t ow = outputShape[2]; |
| 154 | int64_t oc = outputShape[3]; |
| 155 | int64_t fh = filterShape[0]; |
| 156 | int64_t fw = filterShape[1]; |
| 157 | int64_t ic = filterShape[2]; |
| 158 | |
| 159 | Location loc = convOp.getLoc(); |
| 160 | |
| 161 | assert(isa<RankedTensorType>(filterType) && |
| 162 | "expected filter type to be a ranked tensor" ); |
| 163 | auto tensorFilterType = cast<RankedTensorType>(Val&: filterType); |
| 164 | |
| 165 | // Reshape output and filter to the LHS and result of a (B)MNK matmul. |
| 166 | SmallVector<ReassociationIndices> filterReassocIndices = {{0, 1, 2}, {3}}; |
| 167 | auto reshapedFilterType = |
| 168 | RankedTensorType::get(shape: {fh * fw * ic, oc}, elementType: filterType.getElementType(), |
| 169 | encoding: tensorFilterType.getEncoding()); |
| 170 | Value reshapedFilter = tensor::CollapseShapeOp::create( |
| 171 | builder&: rewriter, location: loc, resultType: reshapedFilterType, src: filter, reassociation: filterReassocIndices); |
| 172 | |
| 173 | SmallVector<ReassociationIndices> outputReassocIndices = {{0}, {1, 2}, {3}}; |
| 174 | RankedTensorType reshapedOutputType = |
| 175 | RankedTensorType::get(shape: {n, oh * ow, oc}, elementType: outputType.getElementType()); |
| 176 | Value reshapedOutput = tensor::CollapseShapeOp::create( |
| 177 | builder&: rewriter, location: loc, resultType: reshapedOutputType, src: output, reassociation: outputReassocIndices); |
| 178 | |
| 179 | SmallVector<int64_t> colTensorShape = {n, oh * ow, fh * fw * ic}; |
| 180 | Value colTensor = tensor::EmptyOp::create(builder&: rewriter, location: loc, staticShape: colTensorShape, |
| 181 | elementType: inputType.getElementType()); |
| 182 | |
| 183 | // Convert the input to a (BMK) column tensor. |
| 184 | auto nloops = colTensorShape.size(); |
| 185 | |
| 186 | auto parallel = utils::IteratorType::parallel; |
| 187 | auto reduction = utils::IteratorType::reduction; |
| 188 | SmallVector<utils::IteratorType> img2colIterators(nloops, parallel); |
| 189 | |
| 190 | // Given an index of the im2col matrix, retrieve the corresponding indices of |
| 191 | // the output and filter matrices |
| 192 | auto mIndicesExprs = |
| 193 | delinearize(linearIndex: rewriter.getAffineDimExpr(position: 1U), strides: ArrayRef<int64_t>{ow, 1}); |
| 194 | auto kIndicesExprs = delinearize(linearIndex: rewriter.getAffineDimExpr(position: 2U), |
| 195 | strides: ArrayRef<int64_t>{fw * ic, ic, 1}); |
| 196 | Im2ColToOperandsExprs i2cToOperExprs; |
| 197 | i2cToOperExprs.fhIndex = kIndicesExprs[0]; |
| 198 | i2cToOperExprs.fwIndex = kIndicesExprs[1]; |
| 199 | i2cToOperExprs.icIndex = kIndicesExprs[2]; |
| 200 | i2cToOperExprs.ohIndex = mIndicesExprs[0]; |
| 201 | i2cToOperExprs.owIndex = mIndicesExprs[1]; |
| 202 | |
| 203 | // im2col[n, oh*ow, fh*fw*ic] = input[n, sh*oh + dh*fh, sw*ow + dw*fw, ic] |
| 204 | Im2ColToInputDimsExprs inExprs = getIm2ColInputExpressions( |
| 205 | exprs: i2cToOperExprs, strides: llvm::to_vector(Range: convOp.getStrides().getValues<int64_t>()), |
| 206 | dilations: llvm::to_vector(Range: convOp.getDilations().getValues<int64_t>()), rewriter); |
| 207 | auto inMap = |
| 208 | AffineMap::inferFromExprList(exprsList: {ArrayRef{inExprs.bIndex, inExprs.hIndex, |
| 209 | inExprs.wIndex, inExprs.cIndex}}, |
| 210 | context: rewriter.getContext())[0]; |
| 211 | |
| 212 | SmallVector<AffineMap> img2colIndexingMaps = { |
| 213 | inMap, AffineMap::getMultiDimIdentityMap(numDims: nloops, context)}; |
| 214 | |
| 215 | auto img2ColTensor = linalg::GenericOp::create( |
| 216 | builder&: rewriter, location: loc, resultTensorTypes: colTensor.getType(), |
| 217 | /*inputs=*/input, /*outputs=*/colTensor, indexingMaps: img2colIndexingMaps, |
| 218 | iteratorTypes: img2colIterators, |
| 219 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 220 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: args[0]); |
| 221 | }); |
| 222 | |
| 223 | // Because the filter does not share the same batch dimension, |
| 224 | // the batch dimension is only used in indexing the input and output. Thus |
| 225 | // we cannot use existing linalg named ops like linalg.batch_matmul. |
| 226 | // i.e. (B x) M x K * K x N = (B x) M x N |
| 227 | AffineExpr bDim, mDim, nDim, kDim; |
| 228 | bindDims(ctx: context, exprs&: bDim, exprs&: mDim, exprs&: nDim, exprs&: kDim); |
| 229 | auto lhsMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {bDim, mDim, kDim}, context); |
| 230 | auto rhsMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {kDim, nDim}, context); |
| 231 | auto resultMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {bDim, mDim, nDim}, context); |
| 232 | SmallVector<utils::IteratorType> genericIterators = {parallel, parallel, |
| 233 | parallel, reduction}; |
| 234 | |
| 235 | auto genericOp = linalg::GenericOp::create( |
| 236 | builder&: rewriter, location: loc, resultTensorTypes: reshapedOutputType, |
| 237 | /*inputs=*/ValueRange{img2ColTensor.getResult(i: 0), reshapedFilter}, |
| 238 | /*outputs=*/ValueRange{reshapedOutput}, |
| 239 | indexingMaps: ArrayRef<AffineMap>{lhsMap, rhsMap, resultMap}, iteratorTypes: genericIterators, |
| 240 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 241 | Value mul = |
| 242 | createMul(loc, x: args[0], y: args[1], accType: args[2].getType(), builder&: nestedBuilder); |
| 243 | Value add = createAdd(loc, x: mul, y: args[2], builder&: nestedBuilder); |
| 244 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: add); |
| 245 | }); |
| 246 | Value result = genericOp.getResults().front(); |
| 247 | |
| 248 | auto reshapedResult = tensor::ExpandShapeOp::create( |
| 249 | builder&: rewriter, location: loc, resultType: outputType, src: result, reassociation: outputReassocIndices); |
| 250 | |
| 251 | rewriter.replaceOp(op: convOp, newValues: ArrayRef<Value>{reshapedResult}); |
| 252 | |
| 253 | return std::make_pair(x: img2ColTensor.getOperation(), |
| 254 | y: reshapedResult.getOperation()); |
| 255 | } |
| 256 | |
| 257 | FailureOr<std::pair<Operation *, Operation *>> |
| 258 | rewriteInIm2Col(RewriterBase &rewriter, |
| 259 | linalg::DepthwiseConv2DNhwcHwcOp convOp) { |
| 260 | auto inputType = cast<RankedTensorType>(Val: convOp.getInputs()[0].getType()); |
| 261 | auto filterType = cast<RankedTensorType>(Val: convOp.getInputs()[1].getType()); |
| 262 | auto outputType = cast<RankedTensorType>(Val: convOp.getOutputs()[0].getType()); |
| 263 | |
| 264 | if (!convOp.hasPureTensorSemantics()) |
| 265 | return rewriter.notifyMatchFailure( |
| 266 | arg&: convOp, msg: "expected op to have pure tensor semantics" ); |
| 267 | |
| 268 | if (!filterType.hasStaticShape()) |
| 269 | return rewriter.notifyMatchFailure( |
| 270 | arg&: convOp, msg: "expected a static shape for the filter" ); |
| 271 | |
| 272 | if (!inputType.hasStaticShape()) |
| 273 | return rewriter.notifyMatchFailure(arg&: convOp, |
| 274 | msg: "expected a static shape for the input" ); |
| 275 | |
| 276 | // TODO: Support dilation. |
| 277 | if (!hasAllOneValues(attr: convOp.getDilations())) |
| 278 | return rewriter.notifyMatchFailure(arg&: convOp, |
| 279 | msg: "expected all ones for dilations" ); |
| 280 | |
| 281 | Location loc = convOp.getLoc(); |
| 282 | |
| 283 | auto transposeOperand = [&](Value operand, ArrayRef<int64_t> indices) { |
| 284 | auto operandTensorType = cast<RankedTensorType>(Val: operand.getType()); |
| 285 | auto nloops = indices.size(); |
| 286 | ArrayRef<int64_t> inputShape = operandTensorType.getShape(); |
| 287 | |
| 288 | SmallVector<AffineExpr> exprs = |
| 289 | llvm::map_to_vector<4>(C&: indices, F: [&](int64_t index) -> AffineExpr { |
| 290 | return rewriter.getAffineDimExpr(position: index); |
| 291 | }); |
| 292 | |
| 293 | SmallVector<int64_t> targetShape = llvm::map_to_vector<4>( |
| 294 | C&: indices, F: [&](int64_t index) -> int64_t { return inputShape[index]; }); |
| 295 | |
| 296 | Value outputTensor = tensor::EmptyOp::create( |
| 297 | builder&: rewriter, location: loc, staticShape: targetShape, elementType: operandTensorType.getElementType()); |
| 298 | |
| 299 | SmallVector<utils::IteratorType> loopAttributeTypes( |
| 300 | nloops, utils::IteratorType::parallel); |
| 301 | |
| 302 | SmallVector<AffineMap> indexingMaps = { |
| 303 | inversePermutation( |
| 304 | map: AffineMap::get(dimCount: nloops, symbolCount: 0, results: exprs, context: rewriter.getContext())), |
| 305 | AffineMap::getMultiDimIdentityMap(numDims: nloops, context: rewriter.getContext())}; |
| 306 | |
| 307 | auto transposedOp = linalg::GenericOp::create( |
| 308 | builder&: rewriter, location: loc, resultTensorTypes: outputTensor.getType(), |
| 309 | /*inputs=*/operand, /*outputs=*/outputTensor, indexingMaps, |
| 310 | iteratorTypes: loopAttributeTypes, |
| 311 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 312 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: args[0]); |
| 313 | }); |
| 314 | |
| 315 | return transposedOp.getResult(i: 0); |
| 316 | }; |
| 317 | |
| 318 | Value input = convOp.getInputs()[0]; |
| 319 | Value filter = convOp.getInputs()[1]; |
| 320 | Value output = convOp.getOutputs()[0]; |
| 321 | |
| 322 | // Transpose input, filter so channels are outermost |
| 323 | Value inputT = transposeOperand(input, {0, 3, 1, 2}); |
| 324 | Value filterT = transposeOperand(filter, {2, 0, 1}); |
| 325 | ArrayRef<int64_t> filterTShape = |
| 326 | cast<RankedTensorType>(Val: filterT.getType()).getShape(); |
| 327 | ArrayRef<int64_t> outputShape = outputType.getShape(); |
| 328 | |
| 329 | int n = outputShape[0]; |
| 330 | int oh = outputShape[1]; |
| 331 | int ow = outputShape[2]; |
| 332 | int c = outputShape[3]; |
| 333 | int fh = filterTShape[1]; |
| 334 | int fw = filterTShape[2]; |
| 335 | |
| 336 | SmallVector<int64_t> colTensorShape = {n, c, oh, ow, fh, fw}; |
| 337 | Value transposedOutputTensor = transposeOperand(output, {0, 3, 1, 2}); |
| 338 | |
| 339 | AffineExpr nDim, cDim, ohDim, owDim, khDim, kwDim; |
| 340 | bindDims(ctx: rewriter.getContext(), exprs&: nDim, exprs&: cDim, exprs&: ohDim, exprs&: owDim, exprs&: khDim, exprs&: kwDim); |
| 341 | |
| 342 | AffineExpr shSym = rewriter.getAffineConstantExpr( |
| 343 | constant: convOp.getStrides().getValues<int64_t>()[0]); |
| 344 | AffineExpr swSym = rewriter.getAffineConstantExpr( |
| 345 | constant: convOp.getStrides().getValues<int64_t>()[1]); |
| 346 | |
| 347 | SmallVector<AffineExpr> inputExprs = {nDim, cDim, ohDim * shSym + khDim, |
| 348 | owDim * swSym + kwDim}; |
| 349 | |
| 350 | auto nloops = colTensorShape.size(); |
| 351 | |
| 352 | SmallVector<utils::IteratorType> loopAttributeTypes( |
| 353 | nloops, utils::IteratorType::parallel); |
| 354 | |
| 355 | SmallVector<AffineMap> indexingMaps = { |
| 356 | AffineMap::get(dimCount: nloops, symbolCount: 0, results: inputExprs, context: rewriter.getContext()), |
| 357 | AffineMap::getMultiDimIdentityMap(numDims: nloops, context: rewriter.getContext())}; |
| 358 | |
| 359 | Value colTensor = tensor::EmptyOp::create(builder&: rewriter, location: loc, staticShape: colTensorShape, |
| 360 | elementType: inputType.getElementType()); |
| 361 | |
| 362 | auto img2ColTensor = linalg::GenericOp::create( |
| 363 | builder&: rewriter, location: loc, resultTensorTypes: colTensor.getType(), |
| 364 | /*inputs=*/inputT, /*outputs=*/colTensor, indexingMaps, |
| 365 | iteratorTypes: loopAttributeTypes, |
| 366 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 367 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: args[0]); |
| 368 | }); |
| 369 | |
| 370 | SmallVector<ReassociationIndices> img2ColTensorReassocIndices = { |
| 371 | {0, 1}, {2, 3}, {4, 5}}; |
| 372 | SmallVector<ReassociationIndices> filterReassociationIndice = {{0}, {1, 2}}; |
| 373 | SmallVector<ReassociationIndices> outputReassociationIndice = {{0, 1}, |
| 374 | {2, 3}}; |
| 375 | |
| 376 | auto reshapedImg2ColTensorType = RankedTensorType::get( |
| 377 | shape: {n * c, oh * ow, fh * fw}, elementType: inputType.getElementType()); |
| 378 | auto reshapedFilterTensorType = |
| 379 | RankedTensorType::get(shape: {c, fh * fw}, elementType: filterType.getElementType()); |
| 380 | auto reshapedOutputTensorType = |
| 381 | RankedTensorType::get(shape: {n * c, oh * ow}, elementType: outputType.getElementType()); |
| 382 | |
| 383 | Value reshapedImg2ColTensor = tensor::CollapseShapeOp::create( |
| 384 | builder&: rewriter, location: loc, resultType: reshapedImg2ColTensorType, src: img2ColTensor.getResult(i: 0), |
| 385 | reassociation: img2ColTensorReassocIndices); |
| 386 | Value reshapedFilterTensor = |
| 387 | tensor::CollapseShapeOp::create(builder&: rewriter, location: loc, resultType: reshapedFilterTensorType, |
| 388 | src: filterT, reassociation: filterReassociationIndice); |
| 389 | Value reshapedoutputTensor = tensor::CollapseShapeOp::create( |
| 390 | builder&: rewriter, location: loc, resultType: reshapedOutputTensorType, src: transposedOutputTensor, |
| 391 | reassociation: outputReassociationIndice); |
| 392 | |
| 393 | auto batchMatVecResult = linalg::BatchMatvecOp::create( |
| 394 | builder&: rewriter, location: loc, resultTensorTypes: TypeRange{reshapedoutputTensor.getType()}, |
| 395 | inputs: ValueRange{reshapedImg2ColTensor, reshapedFilterTensor}, |
| 396 | outputs: ValueRange{reshapedoutputTensor}); |
| 397 | |
| 398 | SmallVector<ReassociationIndices> batchMatVecReassociationIndice = {{0, 1}, |
| 399 | {2, 3}}; |
| 400 | |
| 401 | auto batchMatVecResultReshaped = tensor::ExpandShapeOp::create( |
| 402 | builder&: rewriter, location: loc, resultType: transposedOutputTensor.getType(), |
| 403 | src: batchMatVecResult.getResult(i: 0), reassociation: batchMatVecReassociationIndice); |
| 404 | |
| 405 | Value transposedResult = |
| 406 | transposeOperand(batchMatVecResultReshaped, {0, 2, 3, 1}); |
| 407 | |
| 408 | rewriter.replaceOp(op: convOp, newValues: ArrayRef<Value>{transposedResult}); |
| 409 | return std::make_pair(x: img2ColTensor.getOperation(), |
| 410 | y: transposedResult.getDefiningOp()); |
| 411 | } |
| 412 | |
| 413 | FailureOr<std::pair<Operation *, Operation *>> |
| 414 | rewriteInIm2Col(RewriterBase &rewriter, linalg::Conv2DNchwFchwOp convOp) { |
| 415 | auto inputType = cast<ShapedType>(Val: convOp.getInputs()[0].getType()); |
| 416 | auto filterType = cast<ShapedType>(Val: convOp.getInputs()[1].getType()); |
| 417 | auto outputType = cast<ShapedType>(Val: convOp.getOutputs()[0].getType()); |
| 418 | |
| 419 | if (!convOp.hasPureTensorSemantics()) |
| 420 | return rewriter.notifyMatchFailure( |
| 421 | arg&: convOp, msg: "expected op to have pure tensor semantics" ); |
| 422 | |
| 423 | if (!filterType.hasStaticShape()) |
| 424 | return rewriter.notifyMatchFailure( |
| 425 | arg&: convOp, msg: "expected a static shape for the filter" ); |
| 426 | |
| 427 | if (!inputType.hasStaticShape()) |
| 428 | return rewriter.notifyMatchFailure(arg&: convOp, |
| 429 | msg: "expected a static shape for the input" ); |
| 430 | |
| 431 | Value input = convOp.getInputs()[0]; |
| 432 | Value filter = convOp.getInputs()[1]; |
| 433 | Value output = convOp.getOutputs()[0]; |
| 434 | |
| 435 | auto filterShape = filterType.getShape(); |
| 436 | auto outputShape = outputType.getShape(); |
| 437 | |
| 438 | int64_t n = outputShape[0]; |
| 439 | int64_t oc = outputShape[1]; |
| 440 | int64_t oh = outputShape[2]; |
| 441 | int64_t ow = outputShape[3]; |
| 442 | int64_t ic = filterShape[1]; |
| 443 | int64_t fh = filterShape[2]; |
| 444 | int64_t fw = filterShape[3]; |
| 445 | |
| 446 | auto loc = convOp.getLoc(); |
| 447 | MLIRContext *context = rewriter.getContext(); |
| 448 | |
| 449 | assert(isa<RankedTensorType>(filterType) && |
| 450 | "expected filter type to be a ranked tensor" ); |
| 451 | auto tensorFilterType = cast<RankedTensorType>(Val&: filterType); |
| 452 | |
| 453 | SmallVector<ReassociationIndices> filterReassocIndices = {{0}, {1, 2, 3}}; |
| 454 | auto reshapedFilterType = |
| 455 | RankedTensorType::get(shape: {oc, ic * fh * fw}, elementType: inputType.getElementType(), |
| 456 | encoding: tensorFilterType.getEncoding()); |
| 457 | Value reshapedFilter = tensor::CollapseShapeOp::create( |
| 458 | builder&: rewriter, location: loc, resultType: reshapedFilterType, src: filter, reassociation: filterReassocIndices); |
| 459 | |
| 460 | SmallVector<ReassociationIndices> outputReassocIndices = {{0}, {1}, {2, 3}}; |
| 461 | auto reshapedOutputType = |
| 462 | RankedTensorType::get(shape: {n, oc, oh * ow}, elementType: outputType.getElementType()); |
| 463 | Value reshapedOutput = tensor::CollapseShapeOp::create( |
| 464 | builder&: rewriter, location: loc, resultType: reshapedOutputType, src: output, reassociation: outputReassocIndices); |
| 465 | |
| 466 | // Convert the input to a (BKN) tensor. |
| 467 | SmallVector<int64_t, 4> colTensorShape = {n, ic * fh * fw, oh * ow}; |
| 468 | Value colTensor = tensor::EmptyOp::create(builder&: rewriter, location: loc, staticShape: colTensorShape, |
| 469 | elementType: inputType.getElementType()); |
| 470 | |
| 471 | auto nloops = colTensorShape.size(); |
| 472 | |
| 473 | auto parallel = utils::IteratorType::parallel; |
| 474 | auto reduction = utils::IteratorType::reduction; |
| 475 | SmallVector<utils::IteratorType, 3> img2colIterators(nloops, parallel); |
| 476 | |
| 477 | // Recover the original iteration indices from the problem/input sizes: |
| 478 | // given an index of the im2col matrix, retrieve the corresponding indices of |
| 479 | // the output and filter matrices |
| 480 | auto kIndicesExprs = delinearize(linearIndex: rewriter.getAffineDimExpr(position: 1U), |
| 481 | strides: ArrayRef<int64_t>{fh * fw, fw, 1}); |
| 482 | auto mIndicesExprs = |
| 483 | delinearize(linearIndex: rewriter.getAffineDimExpr(position: 2U), strides: ArrayRef<int64_t>{ow, 1}); |
| 484 | Im2ColToOperandsExprs i2cToOperExprs; |
| 485 | i2cToOperExprs.icIndex = kIndicesExprs[0]; |
| 486 | i2cToOperExprs.fhIndex = kIndicesExprs[1]; |
| 487 | i2cToOperExprs.fwIndex = kIndicesExprs[2]; |
| 488 | i2cToOperExprs.ohIndex = mIndicesExprs[0]; |
| 489 | i2cToOperExprs.owIndex = mIndicesExprs[1]; |
| 490 | Im2ColToInputDimsExprs inExprs = getIm2ColInputExpressions( |
| 491 | exprs: i2cToOperExprs, strides: llvm::to_vector(Range: convOp.getStrides().getValues<int64_t>()), |
| 492 | dilations: llvm::to_vector(Range: convOp.getDilations().getValues<int64_t>()), rewriter); |
| 493 | auto inMap = |
| 494 | AffineMap::inferFromExprList(exprsList: {ArrayRef{inExprs.bIndex, inExprs.cIndex, |
| 495 | inExprs.hIndex, inExprs.wIndex}}, |
| 496 | context: rewriter.getContext())[0]; |
| 497 | // im2col[n, ic*fh*fw, oh*ow] = input[n, ic, sh*oh + dh*fh, sw*ow + dw*fw] |
| 498 | SmallVector<AffineMap> img2colIndexingMaps = { |
| 499 | inMap, AffineMap::getMultiDimIdentityMap(numDims: nloops, context)}; |
| 500 | |
| 501 | auto img2ColTensor = linalg::GenericOp::create( |
| 502 | builder&: rewriter, location: loc, resultTensorTypes: colTensor.getType(), |
| 503 | /*inputs=*/input, /*outputs=*/colTensor, indexingMaps: img2colIndexingMaps, |
| 504 | iteratorTypes: img2colIterators, |
| 505 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 506 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: args[0]); |
| 507 | }); |
| 508 | |
| 509 | // Because the filter does not share the same batch dimension, |
| 510 | // the batch dimension is only used in indexing the input and output. Thus |
| 511 | // we cannot use existing linalg named ops like linalg.batch_matmul. |
| 512 | // i.e. M x K * (B x) K x N = (B x) M x N |
| 513 | AffineExpr bDim, mDim, nDim, kDim; |
| 514 | bindDims(ctx: context, exprs&: bDim, exprs&: mDim, exprs&: nDim, exprs&: kDim); |
| 515 | auto lhsMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {mDim, kDim}, context); |
| 516 | auto rhsMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {bDim, kDim, nDim}, context); |
| 517 | auto resultMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {bDim, mDim, nDim}, context); |
| 518 | SmallVector<utils::IteratorType> genericIterators = {parallel, parallel, |
| 519 | parallel, reduction}; |
| 520 | auto genericOp = linalg::GenericOp::create( |
| 521 | builder&: rewriter, location: loc, resultTensorTypes: reshapedOutputType, |
| 522 | /*inputs=*/ValueRange{reshapedFilter, img2ColTensor.getResult(i: 0)}, |
| 523 | /*outputs=*/ValueRange{reshapedOutput}, |
| 524 | indexingMaps: ArrayRef<AffineMap>{lhsMap, rhsMap, resultMap}, iteratorTypes: genericIterators, |
| 525 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 526 | Value mul = |
| 527 | createMul(loc, x: args[0], y: args[1], accType: args[2].getType(), builder&: nestedBuilder); |
| 528 | Value add = createAdd(loc, x: mul, y: args[2], builder&: nestedBuilder); |
| 529 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: add); |
| 530 | }); |
| 531 | Value result = genericOp.getResults().front(); |
| 532 | |
| 533 | auto reshapedResult = tensor::ExpandShapeOp::create( |
| 534 | builder&: rewriter, location: loc, resultType: outputType, src: result, reassociation: outputReassocIndices); |
| 535 | |
| 536 | rewriter.replaceOp(op: convOp, newValues: ArrayRef<Value>{reshapedResult}); |
| 537 | |
| 538 | return std::make_pair(x: img2ColTensor.getOperation(), |
| 539 | y: reshapedResult.getOperation()); |
| 540 | } |
| 541 | |
| 542 | FailureOr<std::pair<Operation *, Operation *>> |
| 543 | rewriteInIm2Col(RewriterBase &rewriter, linalg::Conv2DNhwcFhwcOp convOp) { |
| 544 | auto inputType = cast<ShapedType>(Val: convOp.getInputs()[0].getType()); |
| 545 | auto filterType = cast<ShapedType>(Val: convOp.getInputs()[1].getType()); |
| 546 | auto outputType = cast<ShapedType>(Val: convOp.getOutputs()[0].getType()); |
| 547 | |
| 548 | if (!convOp.hasPureTensorSemantics()) |
| 549 | return rewriter.notifyMatchFailure( |
| 550 | arg&: convOp, msg: "expected op to have pure tensor semantics" ); |
| 551 | |
| 552 | if (!filterType.hasStaticShape()) |
| 553 | return rewriter.notifyMatchFailure( |
| 554 | arg&: convOp, msg: "expected a static shape for the filter" ); |
| 555 | |
| 556 | if (!inputType.hasStaticShape()) |
| 557 | return rewriter.notifyMatchFailure(arg&: convOp, |
| 558 | msg: "expected a static shape for the input" ); |
| 559 | |
| 560 | MLIRContext *context = rewriter.getContext(); |
| 561 | Value input = convOp.getInputs()[0]; |
| 562 | Value filter = convOp.getInputs()[1]; |
| 563 | Value output = convOp.getOutputs()[0]; |
| 564 | |
| 565 | ArrayRef<int64_t> filterShape = filterType.getShape(); |
| 566 | ArrayRef<int64_t> outputShape = outputType.getShape(); |
| 567 | |
| 568 | int64_t n = outputShape[0]; |
| 569 | int64_t oh = outputShape[1]; |
| 570 | int64_t ow = outputShape[2]; |
| 571 | int64_t oc = outputShape[3]; |
| 572 | int64_t fh = filterShape[1]; |
| 573 | int64_t fw = filterShape[2]; |
| 574 | int64_t ic = filterShape[3]; |
| 575 | |
| 576 | Location loc = convOp.getLoc(); |
| 577 | |
| 578 | assert(isa<RankedTensorType>(filterType) && |
| 579 | "expected filter type to be a ranked tensor" ); |
| 580 | auto tensorFilterType = cast<RankedTensorType>(Val&: filterType); |
| 581 | |
| 582 | // Reshape output and filter to the LHS and result of a "row-wise" matrix |
| 583 | // multiplication. |
| 584 | SmallVector<ReassociationIndices> filterReassocIndices = {{0}, {1, 2, 3}}; |
| 585 | auto reshapedFilterType = |
| 586 | RankedTensorType::get(shape: {oc, fh * fw * ic}, elementType: filterType.getElementType(), |
| 587 | encoding: tensorFilterType.getEncoding()); |
| 588 | Value reshapedFilter = tensor::CollapseShapeOp::create( |
| 589 | builder&: rewriter, location: loc, resultType: reshapedFilterType, src: filter, reassociation: filterReassocIndices); |
| 590 | |
| 591 | SmallVector<ReassociationIndices> outputReassocIndices = {{0}, {1, 2}, {3}}; |
| 592 | RankedTensorType reshapedOutputType = |
| 593 | RankedTensorType::get(shape: {n, oh * ow, oc}, elementType: outputType.getElementType()); |
| 594 | Value reshapedOutput = tensor::CollapseShapeOp::create( |
| 595 | builder&: rewriter, location: loc, resultType: reshapedOutputType, src: output, reassociation: outputReassocIndices); |
| 596 | |
| 597 | // Shape of the Toeplitz matrix produced by Im2col. |
| 598 | SmallVector<int64_t> colTensorShape = {n, oh * ow, fh * fw * ic}; |
| 599 | Value colTensor = tensor::EmptyOp::create(builder&: rewriter, location: loc, staticShape: colTensorShape, |
| 600 | elementType: inputType.getElementType()); |
| 601 | |
| 602 | // Convert the input to a (BMK) column tensor. |
| 603 | auto nloops = colTensorShape.size(); |
| 604 | |
| 605 | auto parallel = utils::IteratorType::parallel; |
| 606 | auto reduction = utils::IteratorType::reduction; |
| 607 | SmallVector<utils::IteratorType> img2colIterators(nloops, parallel); |
| 608 | |
| 609 | // Given an index of the im2col matrix, retrieve the corresponding indices of |
| 610 | // the output and filter matrices |
| 611 | auto mIndicesExprs = |
| 612 | delinearize(linearIndex: rewriter.getAffineDimExpr(position: 1U), strides: ArrayRef<int64_t>{ow, 1}); |
| 613 | auto kIndicesExprs = delinearize(linearIndex: rewriter.getAffineDimExpr(position: 2U), |
| 614 | strides: ArrayRef<int64_t>{fw * ic, ic, 1}); |
| 615 | Im2ColToOperandsExprs i2cToOperExprs; |
| 616 | i2cToOperExprs.fhIndex = kIndicesExprs[0]; |
| 617 | i2cToOperExprs.fwIndex = kIndicesExprs[1]; |
| 618 | i2cToOperExprs.icIndex = kIndicesExprs[2]; |
| 619 | i2cToOperExprs.ohIndex = mIndicesExprs[0]; |
| 620 | i2cToOperExprs.owIndex = mIndicesExprs[1]; |
| 621 | |
| 622 | // im2col[n, oh*ow, fh*fw*ic] = input[n, sh*oh + dh*fh, sw*ow + dw*fw, ic] |
| 623 | Im2ColToInputDimsExprs inExprs = getIm2ColInputExpressions( |
| 624 | exprs: i2cToOperExprs, strides: llvm::to_vector(Range: convOp.getStrides().getValues<int64_t>()), |
| 625 | dilations: llvm::to_vector(Range: convOp.getDilations().getValues<int64_t>()), rewriter); |
| 626 | auto inMap = |
| 627 | AffineMap::inferFromExprList(exprsList: {ArrayRef{inExprs.bIndex, inExprs.hIndex, |
| 628 | inExprs.wIndex, inExprs.cIndex}}, |
| 629 | context: rewriter.getContext())[0]; |
| 630 | SmallVector<AffineMap> img2colIndexingMaps = { |
| 631 | inMap, AffineMap::getMultiDimIdentityMap(numDims: nloops, context)}; |
| 632 | |
| 633 | auto img2ColTensor = linalg::GenericOp::create( |
| 634 | builder&: rewriter, location: loc, resultTensorTypes: colTensor.getType(), |
| 635 | /*inputs=*/input, /*outputs=*/colTensor, indexingMaps: img2colIndexingMaps, |
| 636 | iteratorTypes: img2colIterators, |
| 637 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 638 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: args[0]); |
| 639 | }); |
| 640 | |
| 641 | // Because we didn't transpose the filters we don't actually have a batched |
| 642 | // matrix multiply. Instead, we have an operation consisting of "row-wise" dot |
| 643 | // products. |
| 644 | AffineExpr bDim, mDim, nDim, kDim; |
| 645 | bindDims(ctx: context, exprs&: bDim, exprs&: mDim, exprs&: nDim, exprs&: kDim); |
| 646 | auto lhsMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {bDim, mDim, kDim}, context); |
| 647 | auto rhsMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {nDim, kDim}, context); |
| 648 | auto resultMap = AffineMap::get(dimCount: 4, symbolCount: 0, results: {bDim, mDim, nDim}, context); |
| 649 | SmallVector<utils::IteratorType> genericIterators = {parallel, parallel, |
| 650 | parallel, reduction}; |
| 651 | |
| 652 | auto genericOp = linalg::GenericOp::create( |
| 653 | builder&: rewriter, location: loc, resultTensorTypes: reshapedOutputType, |
| 654 | /*inputs=*/ValueRange{img2ColTensor.getResult(i: 0), reshapedFilter}, |
| 655 | /*outputs=*/ValueRange{reshapedOutput}, |
| 656 | indexingMaps: ArrayRef<AffineMap>{lhsMap, rhsMap, resultMap}, iteratorTypes: genericIterators, |
| 657 | odsArg5: [&](OpBuilder &nestedBuilder, Location nestedLoc, ValueRange args) { |
| 658 | Value mul = |
| 659 | createMul(loc, x: args[0], y: args[1], accType: args[2].getType(), builder&: nestedBuilder); |
| 660 | Value add = createAdd(loc, x: mul, y: args[2], builder&: nestedBuilder); |
| 661 | linalg::YieldOp::create(builder&: nestedBuilder, location: nestedLoc, values: add); |
| 662 | }); |
| 663 | Value result = genericOp.getResults().front(); |
| 664 | |
| 665 | auto reshapedResult = tensor::ExpandShapeOp::create( |
| 666 | builder&: rewriter, location: loc, resultType: outputType, src: result, reassociation: outputReassocIndices); |
| 667 | |
| 668 | rewriter.replaceOp(op: convOp, newValues: ArrayRef<Value>{reshapedResult}); |
| 669 | |
| 670 | return std::make_pair(x: img2ColTensor.getOperation(), |
| 671 | y: reshapedResult.getOperation()); |
| 672 | } |
| 673 | |
| 674 | namespace { |
| 675 | |
| 676 | class ConvertConv2DNhwcHwcf final |
| 677 | : public OpRewritePattern<linalg::Conv2DNhwcHwcfOp> { |
| 678 | public: |
| 679 | using Base::Base; |
| 680 | |
| 681 | LogicalResult matchAndRewrite(linalg::Conv2DNhwcHwcfOp convOp, |
| 682 | PatternRewriter &rewriter) const override { |
| 683 | if (failed(Result: rewriteInIm2Col(rewriter, convOp))) |
| 684 | return failure(); |
| 685 | return success(); |
| 686 | } |
| 687 | }; |
| 688 | |
| 689 | class ConvertDepthwiseConv2DNhwcHwc final |
| 690 | : public OpRewritePattern<linalg::DepthwiseConv2DNhwcHwcOp> { |
| 691 | public: |
| 692 | using OpRewritePattern<linalg::DepthwiseConv2DNhwcHwcOp>::OpRewritePattern; |
| 693 | |
| 694 | LogicalResult matchAndRewrite(linalg::DepthwiseConv2DNhwcHwcOp convOp, |
| 695 | PatternRewriter &rewriter) const override { |
| 696 | if (failed(Result: rewriteInIm2Col(rewriter, convOp))) |
| 697 | return failure(); |
| 698 | return success(); |
| 699 | } |
| 700 | }; |
| 701 | |
| 702 | class ConvertConv2DNchwFchw final |
| 703 | : public OpRewritePattern<linalg::Conv2DNchwFchwOp> { |
| 704 | public: |
| 705 | using Base::Base; |
| 706 | |
| 707 | LogicalResult matchAndRewrite(linalg::Conv2DNchwFchwOp convOp, |
| 708 | PatternRewriter &rewriter) const override { |
| 709 | if (failed(Result: rewriteInIm2Col(rewriter, convOp))) |
| 710 | return failure(); |
| 711 | return success(); |
| 712 | } |
| 713 | }; |
| 714 | |
| 715 | class ConvertConv2DNhwcFhwc final |
| 716 | : public OpRewritePattern<linalg::Conv2DNhwcFhwcOp> { |
| 717 | public: |
| 718 | using Base::Base; |
| 719 | |
| 720 | LogicalResult matchAndRewrite(linalg::Conv2DNhwcFhwcOp convOp, |
| 721 | PatternRewriter &rewriter) const override { |
| 722 | if (failed(Result: rewriteInIm2Col(rewriter, convOp))) |
| 723 | return failure(); |
| 724 | return success(); |
| 725 | } |
| 726 | }; |
| 727 | } // end anonymous namespace |
| 728 | |
| 729 | void populateConvertConv2DToImg2ColPatterns(RewritePatternSet &patterns) { |
| 730 | MLIRContext *context = patterns.getContext(); |
| 731 | patterns.insert<ConvertConv2DNhwcHwcf, ConvertDepthwiseConv2DNhwcHwc, |
| 732 | ConvertConv2DNchwFchw, ConvertConv2DNhwcFhwc>(arg&: context); |
| 733 | } |
| 734 | } // end namespace linalg |
| 735 | } // end namespace mlir |
| 736 | |