torch-mlir/lib/Conversion/TorchToMhlo/Gather.cpp

//===----------------------------------------------------------------------===//
//
// 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/TorchToMhlo/TorchToMhlo.h"

#include "../PassDetail.h"
#include "./MhloLegalizeUtils.h"
#include "./PopulatePatterns.h"
#include "mhlo/IR/hlo_ops.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "torch-mlir/Conversion/Utils/Utils.h"
#include "torch-mlir/Dialect/Torch/IR/TorchDialect.h"
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionOps.h"

using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
using namespace mlir::torch::torch_to_mhlo;

namespace {
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 = input.getType().dyn_cast<RankedTensorType>();
  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 = indices.getType().dyn_cast<RankedTensorType>();
  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 = mhlo::GatherDimensionNumbersAttr::get(
      rewriter.getContext(),
      /*offsetDims=*/offsetDims,
      /*collapsedSliceDims=*/collapsedSliceDims,
      /*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<mhlo::DynamicGatherOp>(loc, outputTy, input, indices,
                                     sliceSizesTensor, dimsAttr)
      .getResult();
}
} // 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.weight();
  auto weightTy = weight.getType().template cast<RankedTensorType>();
  if (!weightTy)
    return op.emitError("only ranked tensor types are supported");

  int64_t padding_idx;
  if (!matchPattern(op.padding_idx(), m_TorchConstantInt(&padding_idx)))
    return rewriter.notifyMatchFailure(
        op, "only constant padding_idx is currently supported");

  bool scale_grad_by_freq;
  if (!matchPattern(op.scale_grad_by_freq(),
                    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.sparse(), 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.indices(), 0, options.dimSizeIndexBits);
  rewriter.replaceOpWithNewOp<mhlo::ConvertOp>(
      op, getTypeConverter()->convertType(op.getType()), output);

  return success();
}

template <>
LogicalResult ConvertAtenOp<AtenIndexSelectOp>::matchAndRewrite(
    AtenIndexSelectOp op, OpAdaptor adaptor,
    ConversionPatternRewriter &rewriter) const {
  auto self = adaptor.self();
  auto selfTy = self.getType().template cast<RankedTensorType>();
  if (!selfTy)
    return op.emitError("only ranked tensor types are supported");
  int64_t dim;
  if (!matchPattern(op.dim(), m_TorchConstantInt(&dim)))
    return rewriter.notifyMatchFailure(
        op, "only constant dim is currently supported");

  Value output = gatherTensorAlongSingleAxis(
      rewriter, op, self, adaptor.index(), dim, options.dimSizeIndexBits);

  rewriter.replaceOpWithNewOp<mhlo::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.self();
  Value index = adaptor.index();
  auto inputType = input.getType().cast<RankedTensorType>();
  auto indexType = index.getType().cast<RankedTensorType>();
  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.dim(), 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.sparse_grad(), m_TorchConstantBool(&sparseGrad))) {
    return rewriter.notifyMatchFailure(
        op, "only constant boolean `sparse_grad` param supported");
  }

  auto options = getOptions();
  auto indexShapeInfo =
      mhlo::getDimSizesOfTensor(rewriter, op, index, options.dimSizeIndexBits);
  if (failed(indexShapeInfo)) {
    return rewriter.notifyMatchFailure(
        op, "failed to get dim sizes of `index` param");
  }
  auto intType = rewriter.getIntegerType(options.dimSizeIndexBits);
  auto one = rewriter.create<arith::ConstantOp>(
      loc, rewriter.getIntegerAttr(intType, 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<mhlo::DynamicReshapeOp>(
          loc, toConcatIndexType, index, toConcatIndexShape));
    } else {
      toConcat.push_back(rewriter.create<mhlo::DynamicIotaOp>(
          loc, toConcatIndexType, toConcatIndexShape,
          rewriter.getI64IntegerAttr(i)));
    }
  }
  auto gatherIndicies = rewriter.create<mhlo::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 = mhlo::GatherDimensionNumbersAttr::get(
      rewriter.getContext(),
      /*offsetDims=*/{},
      /*collapsedSliceDims=*/collapsedDims,
      /*startIndexMap=*/startIndexMap,
      /*indexVecDim=*/indexVecDim);

  rewriter.replaceOpWithNewOp<mhlo::GatherOp>(
      op, input, gatherIndicies, dimsAttr,
      rewriter.getI64TensorAttr(sliceSizes));
  return success();
}

void mlir::torch::torch_to_mhlo::populateGatherOpPatternsAndLegality(
    TypeConverter &typeConverter, RewritePatternSet &patterns,
    ConversionTarget &target, const TorchToMhloOptions &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(AtenIndexSelectOp);
  INSERT_ATENOP_PATTERN(AtenGatherOp);
#undef INSERT_ATENOP_PATTERN
}