torch-mlir/lib/Conversion/TorchToStablehlo/GatherScatter.cpp

1470 lines
60 KiB
C++
Raw Normal View History

//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// Also available under a BSD-style license. See LICENSE.
//
//===----------------------------------------------------------------------===//
#include "torch-mlir/Conversion/TorchToStablehlo/TorchToStablehlo.h"
#include "../PassDetail.h"
#include "PopulatePatterns.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Complex/IR/Complex.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "stablehlo/dialect/ChloOps.h"
#include "stablehlo/dialect/StablehloOps.h"
#include "torch-mlir/Conversion/TorchToStablehlo/StablehloLegalizeUtils.h"
#include "torch-mlir/Conversion/Utils/Utils.h"
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
#include "torch-mlir/Dialect/Torch/IR/TorchTypes.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
using namespace mlir::torch::torch_to_stablehlo;
namespace {
static Value createInitialValueForGatherScatterOp(Operation *op,
RankedTensorType constType,
PatternRewriter &rewriter) {
if (!constType.hasStaticShape()) {
return nullptr;
}
auto elementTy = constType.getElementType();
if (isa<AtenEmbeddingBagPaddingIdxOp>(op)) {
if (isa<mlir::FloatType>(elementTy)) {
auto constAttr = DenseElementsAttr::get(
constType, {APFloat::getZero(
cast<mlir::FloatType>(elementTy).getFloatSemantics(),
/*negative=*/false)});
return rewriter.create<stablehlo::ConstantOp>(op->getLoc(), constType,
constAttr);
} else if (isa<mlir::IntegerType>(elementTy) &&
elementTy.getIntOrFloatBitWidth() != 8) {
auto constAttr = DenseElementsAttr::get(
constType, {APInt::getZero(elementTy.getIntOrFloatBitWidth())});
return rewriter.create<stablehlo::ConstantOp>(op->getLoc(), constType,
constAttr);
}
}
op->emitError("unimplemented lowering in "
"createInitialValueForGatherScatterOp");
return nullptr;
}
Value gatherTensorAlongSingleAxis(PatternRewriter &rewriter, Operation *op,
Value input, Value indices, int64_t axis,
size_t dimSizeIndexBits) {
auto loc = op->getLoc();
Type intType = rewriter.getIntegerType(dimSizeIndexBits);
Value one = rewriter.create<arith::ConstantOp>(
loc, rewriter.getIntegerAttr(intType, 1));
// sliceSizes
auto inputRankTy = dyn_cast<RankedTensorType>(input.getType());
auto inputRank = inputRankTy.getRank();
SmallVector<Value, 4> sliceSizes;
sliceSizes.reserve(inputRank);
for (int64_t r = 0; r < inputRank; ++r) {
if (r == axis) {
sliceSizes.push_back(one);
} else {
sliceSizes.push_back(rewriter.create<arith::IndexCastOp>(
loc, intType, rewriter.create<tensor::DimOp>(loc, input, r)));
}
}
auto sliceSizesTensor =
rewriter.create<tensor::FromElementsOp>(loc, sliceSizes);
// offsetDims
SmallVector<int64_t, 4> offsetDims;
offsetDims.reserve(inputRank);
for (int64_t r = 0; r < axis; ++r) {
offsetDims.push_back(r);
}
auto indicesRankTy = dyn_cast<RankedTensorType>(indices.getType());
auto indicesRank = indicesRankTy.getRank();
for (int64_t r = axis + 1; r < inputRank; ++r) {
offsetDims.push_back(r + indicesRank - 1);
}
// collapsedSliceDims
SmallVector<int64_t, 4> collapsedSliceDims(1, axis);
// startIndexMap
SmallVector<int64_t, 4> startIndexMap(1, axis);
// indexVecDim
int64_t indexVecDim = indicesRank;
auto dimsAttr = stablehlo::GatherDimensionNumbersAttr::get(
rewriter.getContext(),
/*offsetDims=*/offsetDims,
/*collapsedSliceDims=*/collapsedSliceDims,
/*operandBatchingDims=*/{},
/*startIndicesBatchingDims=*/{},
/*startIndexMap=*/startIndexMap,
/*indexVecDim=*/indexVecDim);
// outputShape = input.shape[:axis] + indices.shape +
// input.shape[axis + 1:]
auto inputShape = inputRankTy.getShape();
auto indicesShape = indicesRankTy.getShape();
SmallVector<int64_t, 4> outputShape(inputShape.begin(),
inputShape.begin() + axis);
outputShape.insert(outputShape.end(), indicesShape.begin(),
indicesShape.end());
outputShape.insert(outputShape.end(), inputShape.begin() + axis + 1,
inputShape.end());
// create output tensor type
auto outputTy =
RankedTensorType::get(outputShape, inputRankTy.getElementType());
return rewriter
.create<stablehlo::DynamicGatherOp>(loc, outputTy, input, indices,
sliceSizesTensor, dimsAttr)
.getResult();
}
template <typename OpTy, typename OpAdaptor>
LogicalResult prepareArgumentsForSlicingOp(OpTy op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
SmallVector<Value> &resultShape,
SmallVector<Value> &offsets,
SmallVector<Value> &strides) {
Location loc = op.getLoc();
auto input = adaptor.getSelf();
RankedTensorType inputType = cast<RankedTensorType>(input.getType());
Value zero = rewriter.create<arith::ConstantIndexOp>(loc, 0);
Value one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return op->emitError("unimplemented: dim is not constant");
int64_t inputRank = inputType.getRank();
dim = toPositiveDim(dim, inputRank);
if (!isValidDim(dim, inputRank))
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
SmallVector<Value> inputShape = getTensorSizes(rewriter, loc, input);
Value dimSize = inputShape[dim];
Value torchTypeStart = op.getStart();
Value torchTypeEnd = op.getEnd();
Value builtinTypeStart = adaptor.getStart();
Value builtinTypeEnd = adaptor.getEnd();
if (isa<OptionalType>(torchTypeStart.getType()) ||
isa<OptionalType>(torchTypeEnd.getType()))
return rewriter.notifyMatchFailure(op, "unimplemented optional type arg");
int64_t step;
if (!matchPattern(op.getStep(), m_TorchConstantInt(&step))) {
if (!isa<Torch::NoneType>(op.getStep().getType()))
return op->emitError("unimplemented: step is not constant");
step = 1;
}
Value start = toPositiveValidDim(rewriter, loc, torchTypeStart,
builtinTypeStart, zero, dimSize);
Value end = toPositiveValidDim(rewriter, loc, torchTypeEnd, builtinTypeEnd,
dimSize, dimSize);
// end >= start ? end : start
Value endSgeStart = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::sge, end, start);
end = rewriter.create<arith::SelectOp>(loc, endSgeStart, end, start);
Value stepIndex = rewriter.create<arith::ConstantIndexOp>(loc, step);
// Slice logic: resultSize = floordiv(end - start + step - 1, step)
resultShape = getTensorSizes(rewriter, loc, input);
Value len = rewriter.create<arith::SubIOp>(loc, end, start);
Value resultSize = rewriter.create<arith::AddIOp>(loc, len, stepIndex);
resultSize = rewriter.create<arith::SubIOp>(loc, resultSize, one);
resultSize = rewriter.create<arith::FloorDivSIOp>(loc, resultSize, stepIndex);
resultShape[dim] = resultSize;
strides.resize(inputType.getRank(), one);
offsets.resize(inputType.getRank(), zero);
offsets[dim] = start;
strides[dim] = rewriter.create<arith::MulIOp>(loc, strides[dim], stepIndex);
return success();
}
} // namespace
namespace {
// A helper function used to generate stablehlo's ScatterIndices or
// GatherIndices from torch's indices, usually appear in torch ops, like
// aten.index.Tensor or aten.input_put A usage example is as follow: Input: [[1,
// 2, 3],
// [4, 5, 6],
// [7, 8, 9]]
// Indices[0]: [[0, 0, 0],
// [2, 2, 0]]
// Indices[1]: [[2],
// [1]]
// Step 1: broadcast indices tensors
// Indices[0]: [[0, 0, 0],
// [2, 2, 0]]
// Indices[1]: [[2, 2, 2],
// [1, 1, 1]]
// Step 2: concat index tensors at a unsqueezed -1 dimension.
// Indices: [[[0, 2], [0, 2], [0, 2]],
// [[2, 1], [2, 1], [0, 1]]]
FailureOr<Value> broadcastAndConcatIndices(Operation *op,
ConversionPatternRewriter &rewriter,
SmallVector<Value> indexTensors,
llvm::ArrayRef<int64_t> inputShape,
size_t dimSizeIndexBits,
int &maxIndexRank) {
// Step 1: broadcast indices tensors
SmallVector<int64_t> indicesShape;
SmallVector<int64_t> expandShape;
SmallVector<int64_t> concatShape;
bool allIndexStaticShape = true;
Value bcastSizeTensor;
// concat index tensor into to indices tensor for concat
for (size_t i = 0; i < indexTensors.size(); i++) {
auto indexTensor = indexTensors[i];
auto indexTensorType = cast<RankedTensorType>(indexTensor.getType());
for (int64_t size : makeShapeTorchCompatible(indexTensorType.getShape())) {
if (size == kUnknownSize)
allIndexStaticShape = false;
}
maxIndexRank = std::max(maxIndexRank, (int)indexTensorType.getRank());
}
if (!allIndexStaticShape) {
auto bcastSizeTensorInfo = hlo::getBroadcastResultShape(
rewriter, op, indexTensors, dimSizeIndexBits);
if (failed(bcastSizeTensorInfo)) {
return failure();
}
bcastSizeTensor = *bcastSizeTensorInfo;
}
for (int i = 0; i < maxIndexRank; i++) {
indicesShape.push_back(inputShape[i]);
expandShape.push_back(inputShape[i]);
concatShape.push_back(inputShape[i]);
}
expandShape.push_back(1);
concatShape.push_back(indexTensors.size());
SmallVector<Value> broadcastedIndices;
2024-05-16 15:33:23 +08:00
Type indexElemTy = rewriter.getI64Type();
RankedTensorType bcastIndexType =
RankedTensorType::get(indicesShape, indexElemTy);
for (auto indexTensor : indexTensors) {
Value bcastVal;
RankedTensorType reshapeType =
RankedTensorType::get(expandShape, indexElemTy);
if (allIndexStaticShape) {
bcastVal = hlo::promoteAndBroadcast(rewriter, indexTensor, bcastIndexType,
std::nullopt);
bcastVal = rewriter.create<stablehlo::ReshapeOp>(op->getLoc(),
reshapeType, bcastVal);
} else {
bcastVal = hlo::promoteAndBroadcast(rewriter, indexTensor, bcastIndexType,
bcastSizeTensor);
auto bcastValShapeTensorVec =
*hlo::getDimSizesOfTensor(rewriter, op, bcastVal, dimSizeIndexBits);
bcastValShapeTensorVec.push_back(rewriter.create<mlir::arith::ConstantOp>(
op->getLoc(), rewriter.getIntegerAttr(
rewriter.getIntegerType(dimSizeIndexBits), 1)));
Value bcastValShapeTensor = rewriter
.create<tensor::FromElementsOp>(
op->getLoc(), bcastValShapeTensorVec)
.getResult();
bcastVal = rewriter.create<stablehlo::DynamicReshapeOp>(
op->getLoc(), reshapeType, bcastVal, bcastValShapeTensor);
}
broadcastedIndices.push_back(bcastVal);
}
// Step 2: concat index tensors at a unsqueezed -1 dimension.
Value finalIndexTensor = broadcastedIndices[0];
if (broadcastedIndices.size() > 1) {
RankedTensorType concatTy = RankedTensorType::get(concatShape, indexElemTy);
finalIndexTensor = rewriter.create<stablehlo::ConcatenateOp>(
op->getLoc(), concatTy, ValueRange(broadcastedIndices),
concatShape.size() - 1);
}
return finalIndexTensor;
}
} // namespace
// Ref:
// https://pytorch.org/docs/stable/generated/torch.nn.functional.embedding.html
// padding_idx (int, optional)
// If specified, the entries at padding_idx do not contribute to the
// gradient; therefore, the embedding vector at padding_idx is not updated
// during training, i.e. it remains as a fixed “pad”.
// scale_grad_by_freq (boolean, optional)
// If given, this will scale gradients by the inverse of frequency of the
// words in the mini-batch. Default False.
// sparse (bool, optional)
// If True, gradient w.r.t. weight matrix will be a sparse tensor.
template <>
LogicalResult ConvertAtenOp<AtenEmbeddingOp>::matchAndRewrite(
AtenEmbeddingOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto weight = adaptor.getWeight();
auto weightTy = cast<RankedTensorType>(weight.getType());
if (!weightTy)
return op.emitError("only ranked tensor types are supported");
int64_t padding_idx;
if (!matchPattern(op.getPaddingIdx(), m_TorchConstantInt(&padding_idx)))
return rewriter.notifyMatchFailure(
op, "only constant padding_idx is currently supported");
bool scale_grad_by_freq;
if (!matchPattern(op.getScaleGradByFreq(),
m_TorchConstantBool(&scale_grad_by_freq)))
return rewriter.notifyMatchFailure(
op, "only constant scale_grad_by_freq is currently supported");
if (scale_grad_by_freq)
return rewriter.notifyMatchFailure(
op, "scale gradients is currently not supported");
bool sparse;
if (!matchPattern(op.getSparse(), m_TorchConstantBool(&sparse)))
return rewriter.notifyMatchFailure(
op, "only constant sparse is currently supported");
if (sparse)
return rewriter.notifyMatchFailure(
op, "sparse gradients is currently not supported");
Value output = gatherTensorAlongSingleAxis(
rewriter, op, weight, adaptor.getIndices(), 0, options.dimSizeIndexBits);
rewriter.replaceOpWithNewOp<stablehlo::ConvertOp>(
op, getTypeConverter()->convertType(op.getType()), output);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenEmbeddingBagPaddingIdxOp>::matchAndRewrite(
AtenEmbeddingBagPaddingIdxOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op->getLoc();
Value weight = adaptor.getWeight();
Value indices = adaptor.getIndices();
Value offsets = adaptor.getOffsets();
auto weightTy = cast<RankedTensorType>(weight.getType());
if (weightTy && weightTy.hasStaticShape() && weightTy.getRank() != 2)
return rewriter.notifyMatchFailure(
op, "weight must be rank 2 tensor with static shapes");
auto indicesTy = cast<RankedTensorType>(indices.getType());
if (indicesTy && indicesTy.hasStaticShape() && indicesTy.getRank() != 1)
return rewriter.notifyMatchFailure(
op, "indices must be a vector with static shapes");
auto offsetsTy = cast<RankedTensorType>(offsets.getType());
if (offsetsTy && offsetsTy.getRank() != 1 && offsetsTy.hasStaticShape() &&
offsetsTy.getShape()[0] == 1)
return rewriter.notifyMatchFailure(
op, "offsets must be a vector with static shape equal to 1");
if (!isa<Torch::NoneType>(op.getPaddingIdx().getType()))
return rewriter.notifyMatchFailure(
op, "Unimplemented: padding_idx should be none");
if (!isa<Torch::NoneType>(op.getPerSampleWeights().getType()))
return rewriter.notifyMatchFailure(
op, "Unimplemented: per_sample_weights should be none");
bool includeLastOffset;
if (!matchPattern(op.getIncludeLastOffset(),
m_TorchConstantBool(&includeLastOffset))) {
return rewriter.notifyMatchFailure(
op, "include_last_offset is expected to be a constant boolean value.");
}
if (includeLastOffset)
return rewriter.notifyMatchFailure(
op, "include_last_offset is currently not supported");
bool scaleGradByFreq;
if (!matchPattern(op.getScaleGradByFreq(),
m_TorchConstantBool(&scaleGradByFreq)))
return rewriter.notifyMatchFailure(
op, "only constant scale_grad_by_freq is currently supported");
if (scaleGradByFreq)
return rewriter.notifyMatchFailure(
op, "scale gradients is currently not supported");
bool sparse;
if (!matchPattern(op.getSparse(), m_TorchConstantBool(&sparse)))
return rewriter.notifyMatchFailure(
op, "only constant sparse is currently supported");
if (sparse)
return rewriter.notifyMatchFailure(
op, "sparse gradients is currently not supported");
int64_t modeInt;
if (!matchPattern(op.getMode(), m_TorchConstantInt(&modeInt))) {
return rewriter.notifyMatchFailure(
op, "mode is expected to be a constant integer value.");
}
if (modeInt != torch_upstream::EmbeddingBagMode::MODE_SUM) {
return rewriter.notifyMatchFailure(op,
"Unimplemented: Mean and Max mode are "
"not supported yet for EmbeddingBag.");
}
const auto &options =
ConvertAtenOp<AtenEmbeddingBagPaddingIdxOp>::getOptions();
auto weightDimSizes =
*hlo::getDimSizesOfTensor(rewriter, op, weight, options.dimSizeIndexBits);
auto indicesDimSizes = *hlo::getDimSizesOfTensor(rewriter, op, indices,
options.dimSizeIndexBits);
auto offsetsDimSizes = *hlo::getDimSizesOfTensor(rewriter, op, offsets,
options.dimSizeIndexBits);
Value gatherOutput = gatherTensorAlongSingleAxis(
rewriter, op, weight, indices, 0, options.dimSizeIndexBits);
Type elementTy = weightTy.getElementType();
auto constType = RankedTensorType::get({}, elementTy);
Value initValue =
createInitialValueForGatherScatterOp(op, constType, rewriter);
if (!initValue)
return failure();
auto stablehloReduceOp = rewriter.create<stablehlo::ReduceOp>(
Bump stablehlo to openxla/stablehlo@fd52182f76cadb82f2064fe5fc49a4fb4347a826 (#2821) With the recent LLVM integrate and changes from https://github.com/llvm/llvm-project/pull/78260, we hit this build error in Stablehlo (which is quite old). ``` external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1020:14: error: no member named 'startRootUpdate' in 'mlir::PatternRewriter' rewriter.startRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1026:16: error: no member named 'finalizeRootUpdate' in 'mlir::PatternRewriter' rewriter.finalizeRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1029:16: error: no member named 'cancelRootUpdate' in 'mlir::PatternRewriter' rewriter.cancelRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1108:14: error: no member named 'updateRootInPlace' in 'mlir::PatternRewriter' rewriter.updateRootInPlace(op->getParentOp(), [&]() { return; }); ~~~~~~~~ ^ 4 errors generated. Target @torch-mlir//:torch-mlir-opt failed to build ``` I'm still puzzled as to how this didn't fail with the CMake merge gating CI (do we not test Stablehlo builds/tests?). In any case, bumping our submodule to https://github.com/openxla/stablehlo/pull/1918 fixes it. It exposes a new failing lit test in TorchToStablehlo though, that I have looped stablehlo developers into ([here](https://discord.com/channels/999073994483433573/999074539138990131/1201235845391331419)). ``` bazel run @torch-mlir//test/Conversion:TorchToStablehlo/scatter.mlir.test ...external/torch-mlir/test/Conversion/TorchToStablehlo/scatter.mlir within split at <stdin>:1 offset :33:8: error: unexpected error: Expects non-empty reduction block for type inference %0 = torch.aten.scatter.src %arg0, %int0, %arg1, %arg2 : !torch.vtensor<[?,?],si64>, !torch.int, !torch.vtensor<[?,?],si64>, !torch.vtensor<[?,?],si64> -> !torch.vtensor<[?,?],si64> ^ LLVM ERROR: Failed to infer result type(s). ``` Bazel CI: https://github.com/sjain-stanford/torch-mlir/actions/runs/7732673480/job/21083102228
2024-02-01 06:21:17 +08:00
op.getLoc(), gatherOutput, initValue, rewriter.getDenseI64ArrayAttr({0}),
elementTy);
Region &region = stablehloReduceOp.getBody();
Block &block = region.emplaceBlock();
auto blockArgumentTy = RankedTensorType::get({}, elementTy);
block.addArgument(blockArgumentTy, op->getLoc());
block.addArgument(blockArgumentTy, op->getLoc());
auto *firstArgument = block.args_begin();
auto secondArgument = block.args_rbegin();
{
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPointToStart(&block);
Value addResult = rewriter.create<stablehlo::AddOp>(
op->getLoc(), blockArgumentTy, *firstArgument, *secondArgument);
rewriter.create<stablehlo::ReturnOp>(op->getLoc(), addResult);
}
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
auto outShapeInfo = hlo::getDimIndexOfTensor(rewriter, op, weight);
if (failed(outShapeInfo)) {
return rewriter.notifyMatchFailure(
op, "failed to get dimension sizes of the input");
}
auto outShapeVec = *outShapeInfo;
auto one = rewriter.create<mlir::arith::ConstantOp>(
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
op->getLoc(), rewriter.getIntegerAttr(rewriter.getIndexType(), 1));
outShapeVec[0] = one;
auto outShapeTensor =
rewriter.create<mlir::tensor::FromElementsOp>(op->getLoc(), outShapeVec);
auto resultA = rewriter.create<stablehlo::DynamicReshapeOp>(
loc, getTypeConverter()->convertType(op.getType(0)),
stablehloReduceOp.getResult(0), outShapeTensor);
RankedTensorType resultType = cast<RankedTensorType>(
getTypeConverter()->convertType(op->getResult(1).getType()));
Value resultB =
createInitialValueForGatherScatterOp(op, resultType, rewriter);
if (!resultB)
return failure();
resultType = cast<RankedTensorType>(
getTypeConverter()->convertType(op->getResult(2).getType()));
Value resultC =
createInitialValueForGatherScatterOp(op, resultType, rewriter);
if (!resultC)
return failure();
resultType = cast<RankedTensorType>(
getTypeConverter()->convertType(op->getResult(3).getType()));
Value resultD =
createInitialValueForGatherScatterOp(op, resultType, rewriter);
if (!resultD)
return failure();
rewriter.replaceOp(op, {resultA, resultB, resultC, resultD});
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenIndexSelectOp>::matchAndRewrite(
AtenIndexSelectOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return op.emitError("only ranked tensor types are supported");
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(
op, "only constant dim is currently supported");
int64_t inputRank = selfTy.getRank();
dim = toPositiveDim(dim, inputRank);
if (!isValidDim(dim, inputRank))
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
Value output = gatherTensorAlongSingleAxis(
rewriter, op, self, adaptor.getIndex(), dim, options.dimSizeIndexBits);
rewriter.replaceOpWithNewOp<stablehlo::ConvertOp>(
op, getTypeConverter()->convertType(op.getType()), output);
return success();
}
// AtenGatherOp
template <>
LogicalResult ConvertAtenOp<AtenGatherOp>::matchAndRewrite(
AtenGatherOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op->getLoc();
Value input = adaptor.getSelf();
Value index = adaptor.getIndex();
auto inputType = cast<RankedTensorType>(input.getType());
auto indexType = cast<RankedTensorType>(index.getType());
auto indexElemType = indexType.getElementType();
if (indexType.getRank() != inputType.getRank()) {
return op.emitError("`index` and `input` param should have the same rank");
}
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim))) {
return rewriter.notifyMatchFailure(
op, "only constant int `dim` param supported");
}
dim = toPositiveDim(dim, inputType.getRank());
if (!isValidDim(dim, inputType.getRank())) {
return rewriter.notifyMatchFailure(op, "invalid `dim` param detected");
}
bool sparseGrad = false;
if (!matchPattern(op.getSparseGrad(), m_TorchConstantBool(&sparseGrad))) {
return rewriter.notifyMatchFailure(
op, "only constant boolean `sparse_grad` param supported");
}
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
auto indexShapeInfo = hlo::getDimIndexOfTensor(rewriter, op, index);
if (failed(indexShapeInfo)) {
return rewriter.notifyMatchFailure(
op, "failed to get dim sizes of `index` param");
}
auto one = rewriter.create<arith::ConstantOp>(
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
loc, rewriter.getIntegerAttr(rewriter.getIndexType(), 1));
auto toConcatIndexShapeValueVec = *indexShapeInfo;
toConcatIndexShapeValueVec.push_back(one);
auto toConcatIndexShape =
rewriter.create<tensor::FromElementsOp>(loc, toConcatIndexShapeValueVec);
auto indexShape = indexType.getShape();
SmallVector<int64_t> toConcatIndexShapeVec(indexShape.begin(),
indexShape.end());
toConcatIndexShapeVec.push_back(1);
RankedTensorType toConcatIndexType =
RankedTensorType::get(toConcatIndexShapeVec, indexElemType);
SmallVector<Value> toConcat;
for (int64_t i = 0; i < inputType.getRank(); ++i) {
if (i == dim) {
toConcat.push_back(rewriter.create<stablehlo::DynamicReshapeOp>(
loc, toConcatIndexType, index, toConcatIndexShape));
} else {
toConcat.push_back(rewriter.create<stablehlo::DynamicIotaOp>(
loc, toConcatIndexType, toConcatIndexShape,
rewriter.getI64IntegerAttr(i)));
}
}
auto gatherIndicies = rewriter.create<stablehlo::ConcatenateOp>(
loc, toConcat, static_cast<uint64_t>(inputType.getRank()));
SmallVector<int64_t> sliceSizes(inputType.getRank(), 1);
int64_t indexVecDim = inputType.getRank();
SmallVector<int64_t> collapsedDims;
SmallVector<int64_t> startIndexMap;
for (int64_t i = 0; i < inputType.getRank(); ++i) {
collapsedDims.push_back(i);
startIndexMap.push_back(i);
}
auto dimsAttr = stablehlo::GatherDimensionNumbersAttr::get(
rewriter.getContext(),
/*offsetDims=*/{},
/*collapsedSliceDims=*/collapsedDims,
/*operandBatchingDims=*/{},
/*startIndicesBatchingDims=*/{},
/*startIndexMap=*/startIndexMap,
/*indexVecDim=*/indexVecDim);
rewriter.replaceOpWithNewOp<stablehlo::GatherOp>(
op, input, gatherIndicies, dimsAttr,
Bump stablehlo to openxla/stablehlo@fd52182f76cadb82f2064fe5fc49a4fb4347a826 (#2821) With the recent LLVM integrate and changes from https://github.com/llvm/llvm-project/pull/78260, we hit this build error in Stablehlo (which is quite old). ``` external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1020:14: error: no member named 'startRootUpdate' in 'mlir::PatternRewriter' rewriter.startRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1026:16: error: no member named 'finalizeRootUpdate' in 'mlir::PatternRewriter' rewriter.finalizeRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1029:16: error: no member named 'cancelRootUpdate' in 'mlir::PatternRewriter' rewriter.cancelRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1108:14: error: no member named 'updateRootInPlace' in 'mlir::PatternRewriter' rewriter.updateRootInPlace(op->getParentOp(), [&]() { return; }); ~~~~~~~~ ^ 4 errors generated. Target @torch-mlir//:torch-mlir-opt failed to build ``` I'm still puzzled as to how this didn't fail with the CMake merge gating CI (do we not test Stablehlo builds/tests?). In any case, bumping our submodule to https://github.com/openxla/stablehlo/pull/1918 fixes it. It exposes a new failing lit test in TorchToStablehlo though, that I have looped stablehlo developers into ([here](https://discord.com/channels/999073994483433573/999074539138990131/1201235845391331419)). ``` bazel run @torch-mlir//test/Conversion:TorchToStablehlo/scatter.mlir.test ...external/torch-mlir/test/Conversion/TorchToStablehlo/scatter.mlir within split at <stdin>:1 offset :33:8: error: unexpected error: Expects non-empty reduction block for type inference %0 = torch.aten.scatter.src %arg0, %int0, %arg1, %arg2 : !torch.vtensor<[?,?],si64>, !torch.int, !torch.vtensor<[?,?],si64>, !torch.vtensor<[?,?],si64> -> !torch.vtensor<[?,?],si64> ^ LLVM ERROR: Failed to infer result type(s). ``` Bazel CI: https://github.com/sjain-stanford/torch-mlir/actions/runs/7732673480/job/21083102228
2024-02-01 06:21:17 +08:00
rewriter.getDenseI64ArrayAttr(sliceSizes));
return success();
}
// AtenSliceScatterOp
template <>
LogicalResult ConvertAtenOp<AtenSliceScatterOp>::matchAndRewrite(
AtenSliceScatterOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op.getLoc();
const TypeConverter *typeConverter = getTypeConverter();
auto input = adaptor.getSelf();
RankedTensorType inputType = cast<RankedTensorType>(input.getType());
RankedTensorType resultType = cast<RankedTensorType>(
typeConverter->convertType(op->getResult(0).getType()));
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim))) {
return op->emitError("unimplemented: dim is not constant");
}
int64_t inputRank = inputType.getRank();
dim = toPositiveDim(dim, inputRank);
if (!isValidDim(dim, inputRank)) {
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
}
auto inputShape = inputType.getShape();
auto dimSize = inputShape[dim];
int64_t step;
if (!matchPattern(op.getStep(), m_TorchConstantInt(&step))) {
return op->emitError("unimplemented: step is not constant");
}
int64_t start;
if (!matchPattern(op.getStart(), m_TorchConstantInt(&start))) {
return op->emitError("unimplemented: start is not constant");
} else if (ShapedType::isDynamic(dimSize) and start < 0) {
return op->emitError("unimplemented: not support dynamic dimSize when "
"start smaller than 0.");
}
start = start >= 0 ? start : dimSize + start;
int64_t end;
if (!matchPattern(op.getEnd(), m_TorchConstantInt(&end))) {
return op->emitError("unimplemented: end is not constant");
} else if (ShapedType::isDynamic(dimSize) and end < 0) {
return op->emitError(
"unimplemented: not support dynamic dimSize when end smaller than 0.");
}
end = end >= 0 ? end : dimSize + end;
int64_t size = 0;
std::vector<int64_t> indicesVec;
for (int64_t i = start; i < end; i += step) {
indicesVec.push_back(i);
++size;
}
ArrayRef<int64_t> indices(indicesVec);
std::vector<int64_t> tmp_shape = {size, 1};
ArrayRef<int64_t> shape(tmp_shape);
RankedTensorType constType =
RankedTensorType::get(shape, rewriter.getIntegerType(64));
auto constAttr = DenseElementsAttr::get(
RankedTensorType::get(shape, rewriter.getIntegerType(64)), indices);
auto const_op =
rewriter.create<stablehlo::ConstantOp>(loc, constType, constAttr);
Value scatterIndices = const_op.getResult();
SmallVector<int64_t> updateWindowDims;
for (int64_t i = 0; i < inputType.getRank(); ++i) {
if (i == dim) {
continue;
}
updateWindowDims.push_back(i);
}
auto scatterArgs = stablehlo::ScatterDimensionNumbersAttr::get(
rewriter.getContext(),
/*updateWindowDims=*/updateWindowDims,
/*insertedWindowDims=*/{dim},
/*inputBatchingDims=*/{},
/*scatterIndicesBatchingDims=*/{},
/*scatterDimsToOperandDim=*/{dim},
/*indexVectorDim=*/1);
Value src = adaptor.getSrc();
auto scatterOp = rewriter.create<stablehlo::ScatterOp>(
loc, resultType, input, scatterIndices, src, scatterArgs, false, false);
Block &block = scatterOp.getUpdateComputation().emplaceBlock();
auto blockArgumentType =
RankedTensorType::get({}, inputType.getElementType());
block.addArgument(blockArgumentType, loc);
block.addArgument(blockArgumentType, loc);
auto *lhs = block.args_begin();
auto *rhs = std::next(lhs);
{
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPointToStart(&block);
rewriter.create<stablehlo::ReturnOp>(loc, *rhs);
}
rewriter.replaceOp(op, scatterOp.getResults());
return success();
}
template <typename AtenOpT, int reduceType>
class ConvertAtenScatterOp : public ConvertAtenOp<AtenOpT> {
public:
using ConvertAtenOp<AtenOpT>::ConvertAtenOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Value input = adaptor.getSelf();
Value index = adaptor.getIndex();
Value src = adaptor.getSrc();
auto inputType = cast<RankedTensorType>(input.getType());
auto indexType = cast<RankedTensorType>(index.getType());
auto srcType = cast<RankedTensorType>(src.getType());
auto indexElemType = indexType.getElementType();
if (indexType.getRank() != inputType.getRank() ||
inputType.getRank() != srcType.getRank()) {
return op.emitError(
"`index`, `input` and `src` param should have the same rank");
}
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim))) {
return rewriter.notifyMatchFailure(
op, "only constant int `dim` param supported");
}
dim = toPositiveDim(dim, inputType.getRank());
if (!isValidDim(dim, inputType.getRank())) {
return rewriter.notifyMatchFailure(op, "invalid `dim` param detected");
}
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
auto indexShapeInfo = hlo::getDimIndexOfTensor(rewriter, op, index);
if (failed(indexShapeInfo)) {
return rewriter.notifyMatchFailure(
op, "failed to get dim sizes of `index` param");
}
// slice src tensor to have the same shape bound of index tensor in the
// leading dimensions. PyTorch has guaranteed that src tensor size will not
// be smaller than that of index tensor. REF:
// https://pytorch.org/docs/stable/generated/torch.Tensor.scatter_.html#torch.Tensor.scatter_
auto zero = rewriter.create<arith::ConstantOp>(
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
loc, rewriter.getIntegerAttr(rewriter.getIndexType(), 0));
auto one = rewriter.create<arith::ConstantOp>(
[Stablehlo] use index type as dim size, avoid to generate index_cast (#3526) For example, the original IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %0 = arith.index_cast %dim : index to i64 %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %1 = arith.index_cast %dim_0 : index to i64 %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %2 = arith.index_cast %dim_1 : index to i64 %from_elements = tensor.from_elements %0, %1, %2 : tensor<3xi64> %3 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xi64>) -> tensor<?x?x?xf32> %4 = stablehlo.dot_general %arg0, %3, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %4 : tensor<?x?x?xf32> } } ``` After using IndexType, the IR is: ``` module attributes {torch.debug_module_name = "Matmul3D"} { func.func @forward(%arg0: tensor<?x?x?xf32>, %arg1: tensor<?x?x?xf32>) -> tensor<?x?x?xf32> { %c0 = arith.constant 0 : index %c1 = arith.constant 1 : index %c2 = arith.constant 2 : index %dim = tensor.dim %arg1, %c0 : tensor<?x?x?xf32> %dim_0 = tensor.dim %arg1, %c1 : tensor<?x?x?xf32> %dim_1 = tensor.dim %arg1, %c2 : tensor<?x?x?xf32> %from_elements = tensor.from_elements %dim, %dim_0, %dim_1 : tensor<3xindex> %0 = stablehlo.dynamic_broadcast_in_dim %arg1, %from_elements, dims = [0, 1, 2] : (tensor<?x?x?xf32>, tensor<3xindex>) -> tensor<?x?x?xf32> %1 = stablehlo.dot_general %arg0, %0, batching_dims = [0] x [0], contracting_dims = [2] x [1] : (tensor<?x?x?xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32> return %1 : tensor<?x?x?xf32> } } ``` The benefits of using IndexType on shape tensor: * simplify the IR, avoid to generate `arith.index_cast` * let backend compiler have a chance to decide the index width of shape tensor * let stablehlo backend have a chance to serialize dynamic shape IR by [shape_legalize_to_stablehlo](https://github.com/openxla/stablehlo/blob/main/stablehlo/tests/shape_legalize_to_stablehlo.mlir)
2024-07-07 18:03:03 +08:00
loc, rewriter.getIntegerAttr(rewriter.getIndexType(), 1));
SmallVector<Value> sliceIndicies(srcType.getRank(), zero);
SmallVector<Value> sliceStrides(srcType.getRank(), one);
auto sliceIndiciesValue =
rewriter.create<tensor::FromElementsOp>(loc, sliceIndicies);
auto sliceStridesValue =
rewriter.create<tensor::FromElementsOp>(loc, sliceStrides);
auto sliceLimitIndiciesValue =
rewriter.create<tensor::FromElementsOp>(loc, *indexShapeInfo);
auto newSrcType =
RankedTensorType::get(indexType.getShape(), srcType.getElementType());
src = rewriter.create<stablehlo::RealDynamicSliceOp>(
loc, newSrcType, src, sliceIndiciesValue, sliceLimitIndiciesValue,
sliceStridesValue);
// generate scatter indicies for stablehlo::Scatter op.
auto toConcatIndexShapeValueVec = *indexShapeInfo;
toConcatIndexShapeValueVec.push_back(one);
auto toConcatIndexShape = rewriter.create<tensor::FromElementsOp>(
loc, toConcatIndexShapeValueVec);
auto indexShape = indexType.getShape();
SmallVector<int64_t> toConcatIndexShapeVec(indexShape.begin(),
indexShape.end());
toConcatIndexShapeVec.push_back(1);
RankedTensorType toConcatIndexType =
RankedTensorType::get(toConcatIndexShapeVec, indexElemType);
SmallVector<Value> toConcat;
for (int64_t i = 0; i < inputType.getRank(); ++i) {
if (i == dim) {
toConcat.push_back(rewriter.create<stablehlo::DynamicReshapeOp>(
loc, toConcatIndexType, index, toConcatIndexShape));
} else {
toConcat.push_back(rewriter.create<stablehlo::DynamicIotaOp>(
loc, toConcatIndexType, toConcatIndexShape,
rewriter.getI64IntegerAttr(i)));
}
}
auto scatterIndicies = rewriter.create<stablehlo::ConcatenateOp>(
loc, toConcat, static_cast<uint64_t>(inputType.getRank()));
SmallVector<int64_t> sliceSizes(inputType.getRank(), 1);
// generate ScatterDimensionNumbers for stablehlo::Scatter op.
int64_t indexVecDim = inputType.getRank();
SmallVector<int64_t> scatterDimOperandDimMap;
SmallVector<int64_t> insertedWindowDims;
for (int64_t i = 0; i < inputType.getRank(); ++i) {
scatterDimOperandDimMap.push_back(i);
insertedWindowDims.push_back(i);
}
auto scatterDimensionNumbers = stablehlo::ScatterDimensionNumbersAttr::get(
rewriter.getContext(),
/*updateWindowDims=*/{},
/*insertedWindowDims=*/insertedWindowDims,
/*inputBatchingDims=*/{},
/*scatterIndicesBatchingDims=*/{},
/*scatterDimsToOperandDim=*/scatterDimOperandDimMap,
/*indexVectorDim=*/indexVecDim);
auto stablehloScatterOp = rewriter.create<stablehlo::ScatterOp>(
loc, inputType, input, scatterIndicies, src, scatterDimensionNumbers,
false, false);
// config update computation function: just return the element from src.
Block &block = stablehloScatterOp.getUpdateComputation().emplaceBlock();
// add block arguments
auto blockArgumentType =
RankedTensorType::get({}, inputType.getElementType());
block.addArgument(blockArgumentType, loc);
block.addArgument(blockArgumentType, loc);
auto *lhsArg = block.args_begin();
auto *rhsArg = std::next(lhsArg);
{
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPointToStart(&block);
if (reduceType == 0) {
rewriter.create<stablehlo::ReturnOp>(loc, *rhsArg);
} else if (reduceType == 1) {
Value res = rewriter.create<stablehlo::AddOp>(loc, blockArgumentType,
*lhsArg, *rhsArg);
rewriter.create<stablehlo::ReturnOp>(loc, res);
}
}
rewriter.replaceOp(op, stablehloScatterOp.getResults());
return success();
}
};
// AtenIndexTensorOp
// Convert to StableHlo::GatherOp.
template <>
LogicalResult ConvertAtenOp<AtenIndexTensorHackedTwinOp>::matchAndRewrite(
AtenIndexTensorHackedTwinOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op->getLoc();
Value input = adaptor.getSelf();
auto inputTensorType = cast<RankedTensorType>(input.getType());
auto outType =
cast<RankedTensorType>(getTypeConverter()->convertType(op.getType()));
auto outShape = outType.getShape();
Value indexList = op.getIndices();
SmallVector<Value> indicesTorchType;
if (!getListConstructElements(indexList, indicesTorchType))
return op.emitError(
"unimplemented: the tensor list is not from list construct");
auto indexTensors = getTypeConvertedValues(rewriter, loc, getTypeConverter(),
indicesTorchType);
int maxIndexRank = -1;
auto gatherIndicesInfo =
broadcastAndConcatIndices(op, rewriter, indexTensors, outShape,
options.dimSizeIndexBits, maxIndexRank);
if (failed(gatherIndicesInfo)) {
return rewriter.notifyMatchFailure(
op, "failed to generate broadcasted indices");
}
auto gatherIndices = *gatherIndicesInfo;
int64_t numIndicesDim = indexTensors.size();
int64_t indexVecDim = maxIndexRank;
SmallVector<int64_t> offsetDims;
SmallVector<int64_t> collapsedDims;
SmallVector<int64_t> startIndexMap;
for (int64_t i = 0; i < numIndicesDim; ++i) {
collapsedDims.push_back(i);
startIndexMap.push_back(i);
}
for (int64_t i = numIndicesDim; i < inputTensorType.getRank(); i++) {
offsetDims.push_back(i + maxIndexRank - numIndicesDim);
}
auto dimsAttr = stablehlo::GatherDimensionNumbersAttr::get(
rewriter.getContext(),
/*offsetDims=*/offsetDims,
/*collapsedSliceDims=*/collapsedDims,
/*operandBatchingDims=*/{},
/*startIndicesBatchingDims=*/{},
/*startIndexMap=*/startIndexMap,
/*indexVecDim=*/indexVecDim);
SmallVector<int64_t> sliceSizes;
auto inputShape = makeShapeTorchCompatible(inputTensorType.getShape());
for (int64_t i = 0; i < inputTensorType.getRank(); ++i) {
if (i < numIndicesDim) {
sliceSizes.push_back(1);
} else {
sliceSizes.push_back(inputShape[i]);
}
}
rewriter.replaceOpWithNewOp<stablehlo::GatherOp>(
op, outType, input, gatherIndices, dimsAttr,
Bump stablehlo to openxla/stablehlo@fd52182f76cadb82f2064fe5fc49a4fb4347a826 (#2821) With the recent LLVM integrate and changes from https://github.com/llvm/llvm-project/pull/78260, we hit this build error in Stablehlo (which is quite old). ``` external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1020:14: error: no member named 'startRootUpdate' in 'mlir::PatternRewriter' rewriter.startRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1026:16: error: no member named 'finalizeRootUpdate' in 'mlir::PatternRewriter' rewriter.finalizeRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1029:16: error: no member named 'cancelRootUpdate' in 'mlir::PatternRewriter' rewriter.cancelRootUpdate(op); ~~~~~~~~ ^ external/stablehlo/stablehlo/transforms/StablehloRefineShapes.cpp:1108:14: error: no member named 'updateRootInPlace' in 'mlir::PatternRewriter' rewriter.updateRootInPlace(op->getParentOp(), [&]() { return; }); ~~~~~~~~ ^ 4 errors generated. Target @torch-mlir//:torch-mlir-opt failed to build ``` I'm still puzzled as to how this didn't fail with the CMake merge gating CI (do we not test Stablehlo builds/tests?). In any case, bumping our submodule to https://github.com/openxla/stablehlo/pull/1918 fixes it. It exposes a new failing lit test in TorchToStablehlo though, that I have looped stablehlo developers into ([here](https://discord.com/channels/999073994483433573/999074539138990131/1201235845391331419)). ``` bazel run @torch-mlir//test/Conversion:TorchToStablehlo/scatter.mlir.test ...external/torch-mlir/test/Conversion/TorchToStablehlo/scatter.mlir within split at <stdin>:1 offset :33:8: error: unexpected error: Expects non-empty reduction block for type inference %0 = torch.aten.scatter.src %arg0, %int0, %arg1, %arg2 : !torch.vtensor<[?,?],si64>, !torch.int, !torch.vtensor<[?,?],si64>, !torch.vtensor<[?,?],si64> -> !torch.vtensor<[?,?],si64> ^ LLVM ERROR: Failed to infer result type(s). ``` Bazel CI: https://github.com/sjain-stanford/torch-mlir/actions/runs/7732673480/job/21083102228
2024-02-01 06:21:17 +08:00
rewriter.getDenseI64ArrayAttr(sliceSizes));
return success();
}
// AtenIndexPutHackedTwinOP
// Convert to stablehlo::ScatterOp
template <>
LogicalResult ConvertAtenOp<AtenIndexPutHackedTwinOp>::matchAndRewrite(
AtenIndexPutHackedTwinOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op->getLoc();
Value input = adaptor.getSelf();
Value values = adaptor.getValues();
auto outType =
cast<RankedTensorType>(getTypeConverter()->convertType(op.getType()));
auto inputType = cast<RankedTensorType>(input.getType());
auto valuesType = cast<RankedTensorType>(values.getType());
int64_t valueRank = valuesType.getRank();
auto valuesShape = valuesType.getShape();
bool accumulate;
if (!matchPattern(op.getAccumulate(), m_TorchConstantBool(&accumulate))) {
return rewriter.notifyMatchFailure(op,
"accumulate should be a constant bool");
}
Value indexList = op.getIndices();
SmallVector<Value> indicesTorchType;
if (!getListConstructElements(indexList, indicesTorchType))
return op.emitError(
"unimplemented: the tensor list is not from list construct");
int64_t indexCnt = indicesTorchType.size();
auto indexTensors = getTypeConvertedValues(rewriter, loc, getTypeConverter(),
indicesTorchType);
int maxIndexRank = -1;
auto scatterIndicesInfo =
broadcastAndConcatIndices(op, rewriter, indexTensors, valuesShape,
options.dimSizeIndexBits, maxIndexRank);
if (failed(scatterIndicesInfo)) {
return rewriter.notifyMatchFailure(
op, "failed to generate broadcasted indices");
}
auto scatterIndices = *scatterIndicesInfo;
// create stablehlo::ScatterOp
int64_t indexVecDim = maxIndexRank;
SmallVector<int64_t> scatterDimOperandDimMap;
SmallVector<int64_t> insertedWindowDims;
SmallVector<int64_t> updateWindowDims;
for (int64_t i = 0; i < indexCnt; ++i) {
scatterDimOperandDimMap.push_back(i);
insertedWindowDims.push_back(i);
}
for (int64_t i = maxIndexRank; i < valueRank; ++i) {
updateWindowDims.push_back(i);
}
auto scatterDimensionNumbers = stablehlo::ScatterDimensionNumbersAttr::get(
rewriter.getContext(),
/*updateWindowDims=*/updateWindowDims,
/*insertedWindowDims=*/insertedWindowDims,
/*inputBatchingDims=*/{},
/*scatterIndicesBatchingDims=*/{},
/*scatterDimsToOperandDim=*/scatterDimOperandDimMap,
/*indexVectorDim=*/indexVecDim);
auto stablehloScatterOp = rewriter.create<stablehlo::ScatterOp>(
loc, outType, input, scatterIndices, values, scatterDimensionNumbers,
false, false);
// configure update computation function.
Block &block = stablehloScatterOp.getUpdateComputation().emplaceBlock();
// add block arguments
auto blockArgumentType =
RankedTensorType::get({}, inputType.getElementType());
block.addArgument(blockArgumentType, loc);
block.addArgument(blockArgumentType, loc);
auto *lhsArg = block.args_begin();
auto *rhsArg = std::next(lhsArg);
{
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPointToStart(&block);
if (!accumulate) {
rewriter.create<stablehlo::ReturnOp>(loc, *rhsArg);
} else {
Value out = rewriter.create<stablehlo::AddOp>(loc, blockArgumentType,
*lhsArg, *rhsArg);
rewriter.create<stablehlo::ReturnOp>(loc, out);
}
}
rewriter.replaceOp(op, stablehloScatterOp.getResults());
return success();
}
// AtenGridSamplerOp
// See
// https://github.com/pytorch/pytorch/blob/ec58f1f74ebcec744d2ab90ad34abd09c1018e92/torch/_decomp/decompositions.py#L3923-L4086
namespace {
template <typename T>
static Value getConstantLike(OpBuilder &b, Location loc, T constant,
Value val) {
Type ty = getElementTypeOrSelf(val.getType());
auto getAttr = [&]() -> Attribute {
if (isa<mlir::IntegerType>(ty))
return b.getIntegerAttr(ty, constant);
if (isa<mlir::FloatType>(ty))
return b.getFloatAttr(ty, constant);
if (auto complexTy = dyn_cast<mlir::ComplexType>(ty))
return complex::NumberAttr::get(complexTy, constant, 0);
llvm_unreachable("unhandled element type");
};
return b.create<mlir::chlo::ConstantLikeOp>(loc, cast<TypedAttr>(getAttr()),
val);
}
template <typename T>
static Value getConstTensor(ConversionPatternRewriter &rewriter, Operation *op,
ArrayRef<T> values, ArrayRef<int64_t> shape,
Type ty) {
Location loc = op->getLoc();
RankedTensorType valueType = RankedTensorType::get(shape, ty);
auto valueAttr = DenseElementsAttr::get(valueType, values);
return rewriter.create<stablehlo::ConstantOp>(loc, valueType, valueAttr);
}
template <typename T>
static Value getConstScalarTensor(ConversionPatternRewriter &rewriter,
Operation *op, T value, Type ty) {
return getConstTensor(rewriter, op, ArrayRef<T>{value}, {}, ty);
}
// Helper function to lower AtenGridSamplerOp.
static Value unnormalize(ConversionPatternRewriter &rewriter, Operation *op,
Value coords, int64_t size, Type elemTy,
bool alignCorners) {
Location loc = op->getLoc();
APFloat pointFive(cast<mlir::FloatType>(elemTy).getFloatSemantics(), "0.5");
APFloat sizeFloat =
APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), size);
APFloat one = APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), 1);
APFloat zero = APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), 0);
// double mul = alignCorners ? (size * 0.5 - 0.5) : (size * 0.5);
// double ofs = size * 0.5 - 0.5;
APFloat mul =
alignCorners ? sizeFloat * pointFive - pointFive : sizeFloat * pointFive;
APFloat ofs = sizeFloat * pointFive - pointFive;
Value constMul = getConstScalarTensor(rewriter, op, mul, elemTy);
Value constOfs = getConstScalarTensor(rewriter, op, ofs, elemTy);
// use chlo::BroadcastMulOp to multiply constMul with coords.
DenseI64ArrayAttr bcastDimensions;
Value mulResult = rewriter.create<chlo::BroadcastMulOp>(loc, coords, constMul,
bcastDimensions);
// use chlo::BroadcastAddOp to add constOfs to mulResult.
Value result = rewriter.create<chlo::BroadcastAddOp>(loc, mulResult, constOfs,
bcastDimensions);
return result;
}
static Value computeCoordinates(ConversionPatternRewriter &rewriter,
Operation *op, Value coords, int64_t size,
Type elemTy, int64_t padding_mode) {
// TODO: add support for padding_mode 1 and 2.
return coords;
}
static Value computeSourceIndex(ConversionPatternRewriter &rewriter,
Operation *op, Value coords, int64_t size,
Type elemTy, int64_t padding_mode,
bool alignCorners) {
Value coordsUn =
unnormalize(rewriter, op, coords, size, elemTy, alignCorners);
return computeCoordinates(rewriter, op, coordsUn, size, elemTy, padding_mode);
}
// def in_bounds_cond(xs: Tensor, ys: Tensor) -> Tensor:
// return torch.logical_and(
// 0 <= xs, torch.logical_and(xs < iW, torch.logical_and(0 <= ys, ys
// < iH))
// )
static Value inBoundsCond(ConversionPatternRewriter &rewriter, Operation *op,
Value xs, Value ys, int64_t ih, int64_t iw,
Type elemTy) {
Location loc = op->getLoc();
APFloat zeroFloat =
APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), 0);
Value zero = getConstScalarTensor(rewriter, op, zeroFloat, elemTy);
APFloat iwFloat =
APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), iw);
APFloat ihFloat =
APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), ih);
Value iwFloatValue = getConstScalarTensor(rewriter, op, iwFloat, elemTy);
Value ihFloatValue = getConstScalarTensor(rewriter, op, ihFloat, elemTy);
chlo::ComparisonTypeAttr compareTypeAttr = chlo::ComparisonTypeAttr::get(
rewriter.getContext(), chlo::ComparisonType::FLOAT);
chlo::ComparisonDirectionAttr compareLTAttr =
chlo::ComparisonDirectionAttr::get(rewriter.getContext(),
chlo::ComparisonDirection::LT);
chlo::ComparisonDirectionAttr compareGEAttr =
chlo::ComparisonDirectionAttr::get(rewriter.getContext(),
chlo::ComparisonDirection::GE);
DenseI64ArrayAttr bcastDimensions;
Value cond1 = rewriter.create<chlo::BroadcastCompareOp>(
loc, xs, zero, bcastDimensions, compareGEAttr, compareTypeAttr);
Value cond2 = rewriter.create<chlo::BroadcastCompareOp>(
loc, xs, iwFloatValue, bcastDimensions, compareLTAttr, compareTypeAttr);
Value cond3 = rewriter.create<chlo::BroadcastCompareOp>(
loc, ys, zero, bcastDimensions, compareGEAttr, compareTypeAttr);
Value cond4 = rewriter.create<chlo::BroadcastCompareOp>(
loc, ys, ihFloatValue, bcastDimensions, compareLTAttr, compareTypeAttr);
Value cond5 =
rewriter.create<chlo::BroadcastAndOp>(loc, cond1, cond2, bcastDimensions);
Value cond6 =
rewriter.create<chlo::BroadcastAndOp>(loc, cond3, cond4, bcastDimensions);
return rewriter.create<chlo::BroadcastAndOp>(loc, cond5, cond6,
bcastDimensions);
}
// def clip(xs: Tensor, ys: Tensor, ws: Tensor) -> TensorSequenceType:
// cond = in_bounds_cond(xs, ys)
// # To clip to inside valid coordinates, we map the coordinates
// # to (x, y) = (0, 0) and also set the weight to 0
// # We also change the shape of the tensor to the appropriate one for
// # broadcasting with N_idx, C_idx for the purposes of advanced
// indexing c = C if _expand_grid else 1
// return tuple(
// torch.where(cond, t, 0).view(N, c, oH, oW)
// for t in (xs.to(dtype=torch.int64), ys.to(dtype=torch.int64), ws)
// )
SmallVector<Value> clip(ConversionPatternRewriter &rewriter, Operation *op,
Value xs, Value ys, Value ws, int64_t N, int64_t oH,
int64_t oW, int64_t iH, int64_t iW, Type elemTy) {
Location loc = op->getLoc();
auto indexElemTy = rewriter.getI64Type();
auto indexTy = RankedTensorType::get(mlir::ArrayRef<int64_t>{1}, indexElemTy);
Value zeroIntValue = rewriter.create<stablehlo::ConstantOp>(
loc, indexTy, DenseIntElementsAttr::get(indexTy, ArrayRef<int64_t>{0}));
APFloat zeroAPFloat =
APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), 0);
Value zeroFloatValue =
getConstScalarTensor(rewriter, op, zeroAPFloat, elemTy);
Value cond = inBoundsCond(rewriter, op, xs, ys, iH, iW, elemTy);
Value xsInt = rewriter.create<stablehlo::ConvertOp>(loc, xs, indexElemTy);
Value ysInt = rewriter.create<stablehlo::ConvertOp>(loc, ys, indexElemTy);
Value selectXs = rewriter.create<chlo::BroadcastSelectOp>(
loc, ArrayRef<Value>{cond, xsInt, zeroIntValue});
Value selectYs = rewriter.create<chlo::BroadcastSelectOp>(
loc, ArrayRef<Value>{cond, ysInt, zeroIntValue});
Value selectWs = rewriter.create<chlo::BroadcastSelectOp>(
loc, ArrayRef<Value>{cond, ws, zeroFloatValue});
SmallVector<int64_t> sizes = {N, 1, oH, oW};
Value reshapedXs = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(sizes, indexElemTy), selectXs);
Value reshapedYs = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(sizes, indexElemTy), selectYs);
Value reshapedWs = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(sizes, elemTy), selectWs);
return SmallVector<Value>{reshapedXs, reshapedYs, reshapedWs};
}
Value getSummand(ConversionPatternRewriter &rewriter, Operation *op,
Value input, Value ix, Value iy, Value w, int64_t N,
int64_t oH, int64_t oW, int64_t iH, int64_t iW, Value Nidx,
Value CIdx, RankedTensorType outType, Type elemTy,
size_t dimSizeIndexBits) {
Location loc = op->getLoc();
auto inputTensorType = cast<RankedTensorType>(input.getType());
SmallVector<Value> clipValues =
clip(rewriter, op, ix, iy, w, N, oH, oW, iH, iW, elemTy);
Value idxX = clipValues[0];
Value idxY = clipValues[1];
Value idxW = clipValues[2];
SmallVector<Value> indexTensors{Nidx, CIdx, idxY, idxX};
int maxIndexRank = -1;
auto gatherIndicesInfo = broadcastAndConcatIndices(
input.getDefiningOp(), rewriter, indexTensors, outType.getShape(),
dimSizeIndexBits, maxIndexRank);
auto gatherIndices = *gatherIndicesInfo;
int64_t numIndicesDim = indexTensors.size();
int64_t indexVecDim = maxIndexRank;
SmallVector<int64_t> offsetDims;
SmallVector<int64_t> collapsedDims;
SmallVector<int64_t> startIndexMap;
for (int64_t i = 0; i < numIndicesDim; ++i) {
collapsedDims.push_back(i);
startIndexMap.push_back(i);
}
for (int64_t i = numIndicesDim; i < inputTensorType.getRank(); i++) {
offsetDims.push_back(i + maxIndexRank - numIndicesDim);
}
auto dimsAttr = stablehlo::GatherDimensionNumbersAttr::get(
rewriter.getContext(),
/*offsetDims=*/offsetDims,
/*collapsedSliceDims=*/collapsedDims,
/*operandBatchingDims=*/{},
/*startIndicesBatchingDims=*/{},
/*startIndexMap=*/startIndexMap,
/*indexVecDim=*/indexVecDim);
SmallVector<int64_t> sliceSizes;
auto inputShape = makeShapeTorchCompatible(inputTensorType.getShape());
for (int64_t i = 0; i < inputTensorType.getRank(); ++i) {
if (i < numIndicesDim) {
sliceSizes.push_back(1);
} else {
sliceSizes.push_back(inputShape[i]);
}
}
Value gather = rewriter.create<stablehlo::GatherOp>(
loc, input, gatherIndices, dimsAttr,
rewriter.getDenseI64ArrayAttr(sliceSizes));
// use chlo::BroadcastMulOp to multiply idxW with gather.
DenseI64ArrayAttr bcastDimensions;
return rewriter.create<chlo::BroadcastMulOp>(loc, gather, idxW,
bcastDimensions);
}
} // namespace
template <>
LogicalResult ConvertAtenOp<AtenGridSamplerOp>::matchAndRewrite(
AtenGridSamplerOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op->getLoc();
Value input = adaptor.getInput();
Value grid = adaptor.getGrid();
int64_t interpolationMode;
if (!matchPattern(op.getInterpolationMode(),
m_TorchConstantInt(&interpolationMode)))
return rewriter.notifyMatchFailure(
op, "interpolation_mode must be an integer constant");
int64_t paddingMode;
if (!matchPattern(op.getPaddingMode(), m_TorchConstantInt(&paddingMode)))
return rewriter.notifyMatchFailure(
op, "padding_mode must be an integer constant");
if (interpolationMode != 0 && interpolationMode != 1)
return rewriter.notifyMatchFailure(
op, "only support interpolation_mode = 0 (bilinear) or 1(nearest)");
if (paddingMode != 0)
return rewriter.notifyMatchFailure(op,
"only support paddingMode = 0 (Zero)");
bool alignCorners = false;
if (!matchPattern(op.getAlignCorners(), m_TorchConstantBool(&alignCorners)))
return rewriter.notifyMatchFailure(
op, "alignCorners must be a boolean constant");
RankedTensorType inputTy = cast<RankedTensorType>(input.getType());
RankedTensorType gridTy = cast<RankedTensorType>(grid.getType());
RankedTensorType outTy =
cast<RankedTensorType>(getTypeConverter()->convertType(op.getType()));
Type elemTy = inputTy.getElementType();
if (inputTy.getRank() != 4)
return rewriter.notifyMatchFailure(op, "input must be a 4D tensor");
if (gridTy.getRank() != 4)
return rewriter.notifyMatchFailure(op, "grid must be a 4D tensor");
auto inputSize = inputTy.getShape();
auto gridSize = gridTy.getShape();
int64_t N = inputSize[0];
int64_t C = inputSize[1];
int64_t iH = inputSize[2];
int64_t iW = inputSize[3];
int64_t oH = gridSize[1];
int64_t oW = gridSize[2];
// grid is a 4D tensor with shape (N, oH, oW, 2)
Type indexElemTy = rewriter.getI64Type();
RankedTensorType indexTy =
RankedTensorType::get(mlir::ArrayRef<int64_t>{1}, indexElemTy);
Value constN = rewriter.create<stablehlo::ConstantOp>(
loc, indexTy, DenseIntElementsAttr::get(indexTy, {N}));
Value constC = rewriter.create<stablehlo::ConstantOp>(
loc, indexTy, DenseIntElementsAttr::get(indexTy, {C}));
APFloat one = APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), 1);
APFloat zero = APFloat(cast<mlir::FloatType>(elemTy).getFloatSemantics(), 0);
Value constOneFloat = getConstScalarTensor(rewriter, op, one, elemTy);
auto NidxFlatten = rewriter.create<stablehlo::DynamicIotaOp>(
loc, RankedTensorType::get(mlir::ArrayRef<int64_t>{N}, indexElemTy),
constN, 0);
auto CidxFlatten = rewriter.create<stablehlo::DynamicIotaOp>(
loc, RankedTensorType::get(mlir::ArrayRef<int64_t>{C}, indexElemTy),
constC, 0);
// Reshape NidxFlatten to 4D tensor (N, 1, 1, 1)
auto NidxSizes = mlir::SmallVector<int64_t>{N, 1, 1, 1};
auto Nidx = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(NidxSizes, indexElemTy), NidxFlatten);
// Reshape CidxFlatten to 4D tensor (1, C, 1, 1)
auto CidxSizes = mlir::SmallVector<int64_t>{1, C, 1, 1};
auto Cidx = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(CidxSizes, indexElemTy), CidxFlatten);
llvm::SmallVector<int64_t> stride(4, 1);
auto gridX = rewriter.create<stablehlo::SliceOp>(
loc,
RankedTensorType::get(mlir::SmallVector<int64_t>{N, oH, oW, 1},
gridTy.getElementType()),
grid, mlir::SmallVector<int64_t>{0, 0, 0, 0},
mlir::SmallVector<int64_t>{N, oH, oW, 1}, stride);
auto gridY = rewriter.create<stablehlo::SliceOp>(
loc,
RankedTensorType::get(mlir::SmallVector<int64_t>{N, oH, oW, 1},
gridTy.getElementType()),
grid, mlir::SmallVector<int64_t>{0, 0, 0, 1},
mlir::SmallVector<int64_t>{N, oH, oW, 2}, stride);
// squeeze last dimension
auto gridXshape = mlir::SmallVector<int64_t>{N, oH, oW};
auto gridXReshape = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(gridXshape, gridTy.getElementType()), gridX);
auto gridYReshape = rewriter.create<stablehlo::ReshapeOp>(
loc, RankedTensorType::get(gridXshape, gridTy.getElementType()), gridY);
if (interpolationMode == 0) {
Value ix = computeSourceIndex(rewriter, op, gridXReshape, iW, elemTy,
paddingMode, alignCorners);
Value iy = computeSourceIndex(rewriter, op, gridYReshape, iH, elemTy,
paddingMode, alignCorners);
Value ix_nw = rewriter.create<stablehlo::FloorOp>(loc, ix);
Value iy_nw = rewriter.create<stablehlo::FloorOp>(loc, iy);
DenseI64ArrayAttr bcastDimensions;
Value ix_ne = rewriter.create<chlo::BroadcastAddOp>(
loc, ix_nw, constOneFloat, bcastDimensions);
Value iy_ne = iy_nw;
Value ix_sw = ix_nw;
Value iy_sw = rewriter.create<chlo::BroadcastAddOp>(
loc, iy_nw, constOneFloat, bcastDimensions);
Value ix_se = ix_ne;
Value iy_se = iy_sw;
// w_nw = (ix_se - ix) * (iy_se - iy)
// w_ne = (ix - ix_sw) * (iy_sw - iy)
// w_sw = (ix_ne - ix) * (iy - iy_ne)
// w_se = (ix - ix_nw) * (iy - iy_nw)
Value w_nw = rewriter.create<chlo::BroadcastMulOp>(
loc,
rewriter.create<chlo::BroadcastSubOp>(loc, ix_se, ix, bcastDimensions),
rewriter.create<chlo::BroadcastSubOp>(loc, iy_se, iy, bcastDimensions),
bcastDimensions);
Value w_ne = rewriter.create<chlo::BroadcastMulOp>(
loc,
rewriter.create<chlo::BroadcastSubOp>(loc, ix, ix_sw, bcastDimensions),
rewriter.create<chlo::BroadcastSubOp>(loc, iy_sw, iy, bcastDimensions),
bcastDimensions);
Value w_sw = rewriter.create<chlo::BroadcastMulOp>(
loc,
rewriter.create<chlo::BroadcastSubOp>(loc, ix_ne, ix, bcastDimensions),
rewriter.create<chlo::BroadcastSubOp>(loc, iy, iy_ne, bcastDimensions),
bcastDimensions);
Value w_se = rewriter.create<chlo::BroadcastMulOp>(
loc,
rewriter.create<chlo::BroadcastSubOp>(loc, ix, ix_nw, bcastDimensions),
rewriter.create<chlo::BroadcastSubOp>(loc, iy, iy_nw, bcastDimensions),
bcastDimensions);
Value summand_nw =
getSummand(rewriter, op, input, ix_nw, iy_nw, w_nw, N, oH, oW, iH, iW,
Nidx, Cidx, outTy, elemTy, options.dimSizeIndexBits);
Value summand_ne =
getSummand(rewriter, op, input, ix_ne, iy_ne, w_ne, N, oH, oW, iH, iW,
Nidx, Cidx, outTy, elemTy, options.dimSizeIndexBits);
Value summand_sw =
getSummand(rewriter, op, input, ix_sw, iy_sw, w_sw, N, oH, oW, iH, iW,
Nidx, Cidx, outTy, elemTy, options.dimSizeIndexBits);
Value summand_se =
getSummand(rewriter, op, input, ix_se, iy_se, w_se, N, oH, oW, iH, iW,
Nidx, Cidx, outTy, elemTy, options.dimSizeIndexBits);
// summand_nw + summand_ne + summand_sw + summand_se
Value sum = rewriter.create<stablehlo::AddOp>(loc, summand_nw, summand_ne);
sum = rewriter.create<stablehlo::AddOp>(loc, sum, summand_sw);
sum = rewriter.create<stablehlo::AddOp>(loc, sum, summand_se);
rewriter.replaceOp(op, sum);
} else if (interpolationMode == 1) {
Value ix = computeSourceIndex(rewriter, op, gridXReshape, iW, elemTy,
paddingMode, alignCorners);
Value iy = computeSourceIndex(rewriter, op, gridYReshape, iH, elemTy,
paddingMode, alignCorners);
Value ix_round = rewriter.create<stablehlo::RoundOp>(loc, ix);
Value iy_round = rewriter.create<stablehlo::RoundOp>(loc, iy);
Value oneTensor = getConstantLike(rewriter, loc, 1.0, ix_round);
Value summand = getSummand(rewriter, op, input, ix_round, iy_round,
oneTensor, N, oH, oW, iH, iW, Nidx, Cidx, outTy,
elemTy, options.dimSizeIndexBits);
rewriter.replaceOp(op, summand);
}
return success();
}
void mlir::torch::torch_to_stablehlo::
populateGatherScatterOpPatternsAndLegality(
TypeConverter &typeConverter, RewritePatternSet &patterns,
ConversionTarget &target, const TorchToStablehloOptions &options) {
MLIRContext *context = patterns.getContext();
#define INSERT_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenOp<AtenOp>>(typeConverter, context, options)
INSERT_ATENOP_PATTERN(AtenEmbeddingOp);
INSERT_ATENOP_PATTERN(AtenEmbeddingBagPaddingIdxOp);
INSERT_ATENOP_PATTERN(AtenIndexSelectOp);
INSERT_ATENOP_PATTERN(AtenGatherOp);
INSERT_ATENOP_PATTERN(AtenSliceScatterOp);
INSERT_ATENOP_PATTERN(AtenIndexTensorHackedTwinOp);
INSERT_ATENOP_PATTERN(AtenIndexPutHackedTwinOp);
INSERT_ATENOP_PATTERN(AtenGridSamplerOp);
#undef INSERT_ATENOP_PATTERN
#define INSERT_ATEN_SCATTER_PATTERN(AtenOp, reduceType) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenScatterOp<AtenOp, reduceType>>(typeConverter, \
context, options)
INSERT_ATEN_SCATTER_PATTERN(AtenScatterSrcOp, 0); // 0 for None reduce op
INSERT_ATEN_SCATTER_PATTERN(AtenScatterAddOp, 1); // 1 for Add reduce op
#undef INSERT_ATEN_SCATTER_PATTERN
}