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

1072 lines
41 KiB
C++

//===----------------------------------------------------------------------===//
//
// 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 "mlir-hlo/Dialect/mhlo/IR/chlo_ops.h"
#include "mlir-hlo/Dialect/mhlo/IR/hlo_ops.h"
#include "mlir/Dialect/Arithmetic/IR/Arithmetic.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/TorchUpstream.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionOps.h"
#include <iostream>
#include <numeric>
using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
bool skipMultiplyAlpha(Value alphaValue) {
double doubleValue;
auto isFloat = matchPattern(alphaValue, m_TorchConstantFloat(&doubleValue));
int64_t intValue;
auto isInt = matchPattern(alphaValue, m_TorchConstantInt(&intValue));
return ((isFloat && doubleValue == 1.0) || (isInt && intValue == 1.0));
}
// These legalizations are for unary ops with only for floating point datatypes.
// There is no supported quantized integer mode for these.
namespace {
template <typename AtenOpT, typename MhloOpT>
class ConvertAtenUnaryFPOnlyOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value self = adaptor.self();
auto selfTy = self.getType().cast<TensorType>();
if (!selfTy)
return op.emitError("only Tensor types supported in MHLO");
if (selfTy.getElementType().isa<mlir::FloatType>()) {
rewriter.replaceOpWithNewOp<MhloOpT>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
op.getType()),
self);
return success();
} else {
return op.emitError(
"only floating-point datatype legalization supported");
}
}
};
} // namespace
// aten.ones & aten.zeros
// Ref: Error checking based on the Torch to TOSA lowering
namespace {
template <typename AtenOpT, int fillVal>
class ConvertAtenConstPatternOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template dyn_cast<TensorType>();
if (!outType)
return op.emitError("only Tensor types supported in MHLO");
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat())
return op.emitError(
"only floating-point or integer datatype legalization supported");
// FIXME: Handle layout, device and pin_memory. Assume dtype has been
// processed to set output type correctly?
if (!op.layout().getType().template isa<Torch::NoneType>())
return op.emitError("only default layout is supported");
bool pinMemory;
if (!op.pin_memory().getType().template isa<Torch::NoneType>() &&
(!matchPattern(op.pin_memory(), m_TorchConstantBool(&pinMemory)) ||
pinMemory)) {
return op.emitError(
"unsupported pin_memory, should be either None or false");
}
SmallVector<int64_t> shape;
if (!matchPattern(op.size(), m_TorchConstantIntList(shape))) {
return op.emitError("shape must be a list of Scalar constants");
}
int64_t size = 1;
for (auto s : shape)
size *= s;
SmallVector<int32_t> values(size, fillVal);
auto constOp =
mhlo::getConstTensor<int32_t>(rewriter, op, values, shape).value();
rewriter.replaceOpWithNewOp<mhlo::ConvertOp>(op, outType, constOp);
return success();
}
};
} // namespace
// These binary op legalizations are specific to add/sub which have an
// alpha multiplier.
namespace {
template <typename AtenOpT, typename ChloOpT>
class ConvertAtenAddSubOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value lhs = adaptor.self();
RankedTensorType lhsType = lhs.getType().dyn_cast<RankedTensorType>();
Value rhs = adaptor.other();
RankedTensorType rhsType = rhs.getType().dyn_cast<RankedTensorType>();
if (!lhsType)
return op.emitError("only Tensor types supported in MHLO");
TensorType outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat()) {
return op.emitError(
"only floating-point or integer datatype legalization supported");
}
if (!rhsType) {
rhs = mhlo::scalarToMhloTensor(rewriter, op, adaptor.other(), outElemTy);
if (isa<AtenRsubScalarOp>(op)) {
std::swap(lhs, rhs);
}
}
lhs = mhlo::promoteType(rewriter, lhs, outType);
rhs = mhlo::promoteType(rewriter, rhs, outType);
if (!skipMultiplyAlpha(op.alpha())) {
Value alpha =
mhlo::scalarToMhloTensor(rewriter, op, adaptor.alpha(), outElemTy);
DenseIntElementsAttr bcastDimensions;
rhs = rewriter.create<chlo::BroadcastMulOp>(op->getLoc(), rhs, alpha,
bcastDimensions);
}
DenseIntElementsAttr bcastDimensions;
rewriter.replaceOpWithNewOp<ChloOpT>(op, outType, lhs, rhs,
bcastDimensions);
return success();
}
};
} // namespace
// Binary op legalizations for Mul/Div variants.
namespace {
template <typename AtenOpT, typename ChloOpT>
class ConvertAtenMulDivOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value lhs = adaptor.self();
auto lhsType = lhs.getType().dyn_cast<TensorType>();
Value rhs = adaptor.other();
TensorType rhsType = rhs.getType().dyn_cast<TensorType>();
if (!lhsType)
return op.emitError("only Tensor types supported in MHLO");
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat()) {
return op.emitError(
"only floating-point or integer datatype legalization supported");
}
if (std::is_same<AtenOpT, AtenSquareOp>()) {
rhs = lhs;
} else if (!rhsType) {
rhs = mhlo::scalarToMhloTensor(rewriter, op, adaptor.other(), outElemTy);
}
DenseIntElementsAttr bcastDimensions;
lhs = mhlo::promoteType(rewriter, lhs, outType);
rhs = mhlo::promoteType(rewriter, rhs, outType);
auto loc = op.getLoc();
Value result =
rewriter.create<ChloOpT>(loc, outType, lhs, rhs, bcastDimensions);
if (!isa<AtenDivTensorModeOp>(op)) {
rewriter.replaceOp(op, result);
return success();
}
AtenDivTensorModeOp divTensorModeOp =
llvm::dyn_cast<AtenDivTensorModeOp>(op.getOperation());
std::string roundingMode;
if (!matchPattern(divTensorModeOp.rounding_mode(),
m_TorchConstantStr(roundingMode)))
return rewriter.notifyMatchFailure(
op, "only support constant str rounding mode");
if (roundingMode == "trunc") {
// "trunc" - rounds the results of the division towards zero. Equivalent
// to C-style integer division.
auto sign = rewriter.create<mhlo::SignOp>(loc, result);
auto abs = rewriter.create<mhlo::AbsOp>(loc, result);
auto floor = rewriter.create<mhlo::FloorOp>(loc, abs);
result = rewriter.create<mhlo::MulOp>(loc, sign, floor).getResult();
}
if (roundingMode == "floor") {
// "floor" - rounds the results of the division down. Equivalent to
// floor division in Python (the // operator)
result = rewriter.create<mhlo::FloorOp>(loc, result).getResult();
}
rewriter.replaceOp(op, result);
return success();
}
};
} // namespace
// Binary op legalizations for comparator ops.
namespace {
template <typename AtenOpT>
class ConvertAtenCompareOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value lhs = adaptor.self();
Value rhs = adaptor.other();
RankedTensorType lhsTy = lhs.getType().dyn_cast<RankedTensorType>();
RankedTensorType rhsTy = rhs.getType().dyn_cast<RankedTensorType>();
if (!lhsTy)
return op.emitError("only Tensor types supported in MHLO");
RankedTensorType outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<RankedTensorType>();
Type lhsElemTy = lhsTy.getElementType();
if (!lhsElemTy.isIntOrFloat()) {
return op.emitError(
"only floating-point or integer datatype legalization supported");
}
if (!rhsTy) {
rhs = mhlo::scalarToMhloTensor(rewriter, op, adaptor.other(), lhsElemTy);
}
// TODO: what is the PyTorch default type promotion?
rhs = mhlo::promoteType(rewriter, rhs, lhsTy);
mhlo::ComparisonTypeAttr compareTypeAttr;
mhlo::ComparisonDirectionAttr compareDirectionAttr;
if (lhsElemTy.isa<mlir::FloatType>()) {
compareTypeAttr = mhlo::ComparisonTypeAttr::get(
op->getContext(), mhlo::ComparisonType::FLOAT);
} else if (lhsElemTy.isa<mlir::IntegerType>()) {
compareTypeAttr = mhlo::ComparisonTypeAttr::get(
op->getContext(), mhlo::ComparisonType::SIGNED);
}
if (std::is_same<AtenOpT, AtenLtTensorOp>() ||
std::is_same<AtenOpT, AtenLtScalarOp>()) {
compareDirectionAttr = mhlo::ComparisonDirectionAttr::get(
op->getContext(), mhlo::ComparisonDirection::LT);
} else if (std::is_same<AtenOpT, AtenGtTensorOp>() ||
std::is_same<AtenOpT, AtenGtScalarOp>()) {
compareDirectionAttr = mhlo::ComparisonDirectionAttr::get(
op->getContext(), mhlo::ComparisonDirection::GT);
} else if (std::is_same<AtenOpT, AtenEqTensorOp>() ||
std::is_same<AtenOpT, AtenEqScalarOp>()) {
compareDirectionAttr = mhlo::ComparisonDirectionAttr::get(
op->getContext(), mhlo::ComparisonDirection::EQ);
} else if (std::is_same<AtenOpT, AtenNeTensorOp>() ||
std::is_same<AtenOpT, AtenNeScalarOp>()) {
compareDirectionAttr = mhlo::ComparisonDirectionAttr::get(
op->getContext(), mhlo::ComparisonDirection::NE);
}
DenseIntElementsAttr bcastDimensions;
rewriter.replaceOpWithNewOp<chlo::BroadcastCompareOp>(
op, outType, lhs, rhs, bcastDimensions, compareDirectionAttr,
compareTypeAttr);
return success();
}
};
} // namespace
// AtenTransposeIntOp
namespace {
class ConvertAtenTransposeIntOp
: public OpConversionPattern<AtenTransposeIntOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenTransposeIntOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value self = adaptor.self();
int64_t dim0;
if (!matchPattern(op.dim0(), m_TorchConstantInt(&dim0))) {
return rewriter.notifyMatchFailure(op, "dim0 must be constant");
}
int64_t dim1;
if (!matchPattern(op.dim1(), m_TorchConstantInt(&dim1))) {
return rewriter.notifyMatchFailure(op, "dim1 must be constant");
}
auto inType = self.getType().cast<RankedTensorType>();
auto inputRank = inType.getRank();
auto outType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
dim0 = toPositiveDim(dim0, inputRank);
if (!isValidDim(dim0, inputRank)) {
return rewriter.notifyMatchFailure(op, "dim0 out of range");
}
dim1 = toPositiveDim(dim1, inputRank);
if (!isValidDim(dim1, inputRank)) {
return rewriter.notifyMatchFailure(op, "dim1 out of range");
}
SmallVector<int64_t> permValues(inputRank);
std::iota(std::begin(permValues), std::end(permValues), 0);
std::swap(permValues[dim0], permValues[dim1]);
DenseIntElementsAttr permutation = DenseIntElementsAttr::get(
RankedTensorType::get({static_cast<long int>(permValues.size())},
rewriter.getI64Type()),
permValues);
rewriter.replaceOpWithNewOp<mhlo::TransposeOp>(op, outType, self,
permutation);
return success();
}
};
} // namespace
// AtenBroadcastToOp
namespace {
class ConvertAtenBroadcastToOp : public OpConversionPattern<AtenBroadcastToOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenBroadcastToOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value self = adaptor.self();
auto selfTy = self.getType().cast<RankedTensorType>();
auto outType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
#ifdef TORCH_MLIR_ENABLE_MHLO_STATIC_SHAPE
if (selfTy.hasStaticShape()) {
Value bcastOp = mhlo::promoteAndBroadcast(rewriter, self, outType);
rewriter.replaceOp(op, bcastOp);
return success();
}
#endif
SmallVector<Value> shape;
if (!(getListConstructElements(adaptor.size(), shape))) {
return op->emitError("desired shape must be a list of scalar");
}
SmallVector<Value> bcastShapeVec;
int64_t totalRank = shape.size();
int64_t selfRank = selfTy.getRank();
int64_t leadingRank = totalRank - selfRank;
for (int64_t i = 0; i < totalRank; ++i) {
Value dValue = shape[i];
Value newD;
int64_t dInt;
if (!(matchPattern(dValue, m_TorchConstantInt(&dInt)))) {
return op->emitError("element of desired shape must be a scalar");
}
if (i >= leadingRank && dInt == -1) {
newD = rewriter.create<mlir::tensor::DimOp>(op->getLoc(), self,
i - leadingRank);
} else {
dValue = rewriter.create<torch::TorchConversion::ToI64Op>(op->getLoc(),
dValue);
newD = rewriter.create<mlir::arith::IndexCastOp>(
op->getLoc(), rewriter.getIndexType(), dValue);
}
bcastShapeVec.push_back(newD);
}
#ifdef TORCH_MLIR_ENABLE_MHLO_TRUNC_DIMSIZE_TO_I32
for (auto &dsize : bcastShapeVec) {
auto dsizeI64 = rewriter.create<mlir::arith::IndexCastOp>(
op->getLoc(), rewriter.getI64Type(), dsize);
dsize = rewriter.create<arith::TruncIOp>(op->getLoc(),
rewriter.getI32Type(), dsizeI64);
}
#endif
Value bcastShapeTensor = rewriter.create<mlir::tensor::FromElementsOp>(
op->getLoc(), ValueRange{bcastShapeVec});
auto dimensionNumbers =
llvm::to_vector<4>(llvm::seq<int64_t>(leadingRank, totalRank));
rewriter.replaceOpWithNewOp<mhlo::DynamicBroadcastInDimOp>(
op, outType, self, bcastShapeTensor,
rewriter.getI64TensorAttr(dimensionNumbers));
return success();
}
};
} // namespace
// AtenPermuteOp
namespace {
class ConvertAtenPermuteOp : public OpConversionPattern<AtenPermuteOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenPermuteOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value self = adaptor.self();
// Not a ranked tensor type
auto inType = self.getType().dyn_cast<RankedTensorType>();
auto outType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
if (!inType)
return op.emitError("only ranked tensor types with static shapes are "
"currently supported");
SmallVector<int64_t> permValues;
if (!matchPattern(adaptor.dims(), m_TorchConstantIntList(permValues)))
return rewriter.notifyMatchFailure(
op, "only constant dimensions are currently supported");
int64_t inRank = inType.getRank();
for (auto &d : permValues) {
d = toPositiveDim(d, inRank);
if (!isValidDim(d, inRank))
return op.emitError("not all dims are valid");
}
DenseIntElementsAttr permutation = DenseIntElementsAttr::get(
RankedTensorType::get({static_cast<long int>(permValues.size())},
rewriter.getI64Type()),
permValues);
rewriter.replaceOpWithNewOp<mhlo::TransposeOp>(op, outType, self,
permutation);
return success();
}
};
} // namespace
namespace {
template <typename AtenOpT>
class ConvertAtenOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override;
};
} // namespace
// AtenTanhOp
namespace {
template <>
LogicalResult ConvertAtenOp<AtenTanhOp>::matchAndRewrite(
AtenTanhOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.self();
auto selfTy = self.getType().cast<TensorType>();
if (selfTy && selfTy.getElementType().isa<mlir::FloatType>()) {
rewriter.replaceOpWithNewOp<mhlo::TanhOp>(
op, getTypeConverter()->convertType(op.getType()), self);
return success();
} else {
return op.emitError(
"only floating-point datatype legalization currently supported");
}
}
} // namespace
// ValueTensorLiteralOp
namespace {
template <>
LogicalResult ConvertAtenOp<ValueTensorLiteralOp>::matchAndRewrite(
ValueTensorLiteralOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
RankedTensorType resultType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
// Tensors with integer types need to be converted to signless integer
// element type. All tensors with element types other than integer can reuse
// existing elements attribute.
// TODO: what about unsigned integer?
if (auto elements = op.valueAttr().dyn_cast<DenseIntElementsAttr>()) {
Type builtinTensorElemTy = resultType.getElementType();
unsigned bitWidth = builtinTensorElemTy.getIntOrFloatBitWidth();
DenseElementsAttr valueAttr =
elements.mapValues(builtinTensorElemTy, [&](const APInt &v) {
return APInt(bitWidth, v.getSExtValue());
});
rewriter.replaceOpWithNewOp<mhlo::ConstantOp>(op, resultType, valueAttr);
return success();
}
rewriter.replaceOpWithNewOp<mhlo::ConstantOp>(op, resultType,
adaptor.value());
return success();
}
} // namespace
// AtenReciprocalOp
// Reciprocal(x) = Div(1, x)
namespace {
template <>
LogicalResult ConvertAtenOp<AtenReciprocalOp>::matchAndRewrite(
AtenReciprocalOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value input = adaptor.self();
auto inputTy = input.getType().cast<RankedTensorType>();
auto outTy =
getTypeConverter()->convertType(op.getType()).cast<RankedTensorType>();
if (!inputTy.getElementType().isa<mlir::FloatType>()) {
return op.emitError("only floating-point datatype legalization supported "
"for AtenReciprocalOp");
}
Value oneTensor = chlo::getConstantLike(rewriter, op->getLoc(), 1, input);
rewriter.replaceOpWithNewOp<mhlo::DivOp>(op, outTy, oneTensor, input);
return success();
}
} // namespace
// PrimNumToTensorScalarOp
namespace {
template <>
LogicalResult ConvertAtenOp<PrimNumToTensorScalarOp>::matchAndRewrite(
PrimNumToTensorScalarOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
RankedTensorType outputType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
auto outputElemType = outputType.getElementType();
Value mhloTensor =
mhlo::scalarToMhloTensor(rewriter, op, adaptor.a(), outputElemType);
rewriter.replaceOp(op, mhloTensor);
return success();
}
} // namespace
// AtenContiguousOp
// Ref: TosaToTosa.cpp for implementation details
namespace {
template <>
LogicalResult ConvertAtenOp<AtenContiguousOp>::matchAndRewrite(
AtenContiguousOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = adaptor.self().getType().dyn_cast<TensorType>();
if (!selfType)
return op.emitError("only tensor types are currently supported");
// FIXME: memory_format is not handled.
rewriter.replaceOp(op, adaptor.self());
return success();
}
} // namespace
// AtenReluOp
// Relu(x) = Max(0, x)
namespace {
template <>
LogicalResult ConvertAtenOp<AtenReluOp>::matchAndRewrite(
AtenReluOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value lhs = adaptor.self();
auto lhsTy = lhs.getType().cast<RankedTensorType>();
auto lhsElemTy = lhsTy.getElementType();
if (!lhsElemTy.isa<mlir::FloatType>()) {
return op->emitError("only float tensor in relu op is supported");
}
Value zeroTensor;
zeroTensor = chlo::getConstantLike(
rewriter, op->getLoc(),
APFloat::getZero(lhsElemTy.cast<mlir::FloatType>().getFloatSemantics(),
false),
lhs);
rewriter.replaceOpWithNewOp<mhlo::MaxOp>(op, lhs, zeroTensor);
return success();
}
} // namespace
// Convert a Aten::GELU to HLO
// Gelu(x) = x * 1/2 * [1 + erf(x/(sqrt(2)))]
namespace {
template <>
LogicalResult ConvertAtenOp<AtenGeluOp>::matchAndRewrite(
AtenGeluOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op.getLoc();
Value input = adaptor.self();
auto inputTy = input.getType().template dyn_cast<RankedTensorType>();
if (!inputTy) {
return op.emitError("only ranked tensor type is supported.");
}
Value one = chlo::getConstantLike(rewriter, loc, 1.0, input);
Value two = chlo::getConstantLike(rewriter, loc, 2.0, input);
Value half = chlo::getConstantLike(rewriter, loc, 0.5, input);
auto rsqrtTwo = rewriter.create<mlir::mhlo::RsqrtOp>(loc, two);
auto erfElement = rewriter.create<mhlo::MulOp>(loc, input, rsqrtTwo);
auto erf = rewriter.create<mlir::chlo::ErfOp>(loc, erfElement);
auto erfAdd = rewriter.create<mhlo::AddOp>(loc, erf, one);
auto halfMul = rewriter.create<mhlo::MulOp>(loc, erfAdd, half);
rewriter.replaceOpWithNewOp<mhlo::MulOp>(op, input, halfMul);
return success();
}
} // namespace
// AtenErfOp
namespace {
template <>
LogicalResult ConvertAtenOp<AtenErfOp>::matchAndRewrite(
AtenErfOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value input = adaptor.self();
auto inputType = input.getType().cast<TensorType>();
if (!inputType.getElementType().isa<mlir::FloatType>()) {
return rewriter.notifyMatchFailure(op, "only float tensor is supported");
}
rewriter.replaceOpWithNewOp<chlo::ErfOp>(
op, getTypeConverter()->convertType(op.getType()), input);
return success();
}
} // namespace
// AtenBatchNormOp
namespace {
template <>
LogicalResult ConvertAtenOp<AtenBatchNormOp>::matchAndRewrite(
AtenBatchNormOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value input = adaptor.input();
// shape = [N, C, H, W]
auto inputTy = input.getType().cast<RankedTensorType>();
Value weight = adaptor.weight();
Value bias = adaptor.bias();
Value runningMean = adaptor.running_mean();
Value runningVar = adaptor.running_var();
// momentum is ignored
Value momentum = adaptor.momentum();
(void)momentum;
if (inputTy.getRank() <= 2) {
return rewriter.notifyMatchFailure(op,
"input should have rank larger than 2");
}
if (!inputTy.getElementType().template isa<mlir::FloatType>()) {
return op.emitError("only input tensor of float type is supported");
}
auto inputElemTy = inputTy.getElementType().cast<mlir::FloatType>();
Value channelDim = rewriter.create<tensor::DimOp>(op->getLoc(), input, 1);
#ifdef TORCH_MLIR_ENABLE_MHLO_TRUNC_DIMSIZE_TO_I32
auto channelDimI64 = rewriter.create<mlir::arith::IndexCastOp>(
op->getLoc(), rewriter.getI64Type(), channelDim);
channelDim = rewriter.create<arith::TruncIOp>(
op->getLoc(), rewriter.getI32Type(), channelDimI64);
#endif
Value channelShape = rewriter.create<tensor::FromElementsOp>(
op->getLoc(), ValueRange{channelDim});
if (failed(checkNotNone(rewriter, op, weight))) {
weight = mhlo::getConstantOfShape(
rewriter, op->getLoc(), APFloat(inputElemTy.getFloatSemantics(), 1),
channelShape,
RankedTensorType::get({inputTy.getShape()[1]},
inputTy.getElementType()));
}
if (failed(checkNotNone(rewriter, op, bias))) {
bias = mhlo::getConstantOfShape(
rewriter, op->getLoc(), APFloat(inputElemTy.getFloatSemantics(), 0),
channelShape,
RankedTensorType::get({inputTy.getShape()[1]},
inputTy.getElementType()));
}
if (failed(checkNotNone(rewriter, op, runningVar))) {
runningVar = mhlo::getConstantOfShape(
rewriter, op->getLoc(), APFloat(inputElemTy.getFloatSemantics(), 1),
channelShape,
RankedTensorType::get({inputTy.getShape()[1]},
inputTy.getElementType()));
}
if (failed(checkNotNone(rewriter, op, runningMean))) {
runningMean = mhlo::getConstantOfShape(
rewriter, op->getLoc(), APFloat(inputElemTy.getFloatSemantics(), 0),
channelShape,
RankedTensorType::get({inputTy.getShape()[1]},
inputTy.getElementType()));
}
auto weightTy = weight.getType().cast<RankedTensorType>();
auto biasTy = bias.getType().cast<RankedTensorType>();
auto runningMeanTy = runningMean.getType().cast<RankedTensorType>();
auto runningVarTy = runningVar.getType().cast<RankedTensorType>();
if (weightTy.getRank() != 1 || biasTy.getRank() != 1 ||
runningMeanTy.getRank() != 1 || runningVarTy.getRank() != 1) {
return rewriter.notifyMatchFailure(
op, "expect weight, bias, running_mean and running_var to be rank 1");
}
if (!weightTy.getElementType().template isa<mlir::FloatType>() ||
!biasTy.getElementType().template isa<mlir::FloatType>() ||
!runningMeanTy.getElementType().template isa<mlir::FloatType>() ||
!runningVarTy.getElementType().template isa<mlir::FloatType>()) {
return op.emitError("only float weight/bias/runningMean/runningVar tensor "
"of float type is supported");
}
double eps = 0.0;
if (!matchPattern(op.eps(), m_TorchConstantFloat(&eps))) {
return rewriter.notifyMatchFailure(op, "non-float(double) eps unsupported");
}
bool training = false;
if (!matchPattern(op.training(), m_TorchConstantBool(&training))) {
return rewriter.notifyMatchFailure(op, "non-bool training unsupported");
}
// TODO: handle cudnnEnabled parameter. Here, we just ignore it!
bool cudnnEnabled = false;
if (!matchPattern(op.cudnn_enabled(), m_TorchConstantBool(&cudnnEnabled))) {
return rewriter.notifyMatchFailure(op,
"non-bool cudnn_enabled unsupported");
}
if (training) {
Type outputTy = getTypeConverter()->convertType(op.getType());
Type batchMeanOrVarTy =
RankedTensorType::get(weightTy.getShape(), inputTy.getElementType());
auto batchNormTrainingResult = rewriter.create<mhlo::BatchNormTrainingOp>(
op.getLoc(), outputTy, batchMeanOrVarTy, batchMeanOrVarTy, input,
weight, bias, rewriter.getF32FloatAttr(eps),
rewriter.getI64IntegerAttr(1));
rewriter.replaceOp(op, batchNormTrainingResult.getResult(0));
return success();
} else {
Type outputTy = getTypeConverter()->convertType(op.getType());
rewriter.replaceOpWithNewOp<mhlo::BatchNormInferenceOp>(
op, outputTy, input, weight, bias, runningMean, runningVar,
rewriter.getFloatAttr(inputTy.getElementType(), eps),
rewriter.getI64IntegerAttr(1));
return success();
}
}
} // namespace
// AtenNativeLayerNormOp
namespace {
template <>
LogicalResult ConvertAtenOp<AtenNativeLayerNormOp>::matchAndRewrite(
AtenNativeLayerNormOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value input = adaptor.input();
auto inputTy = input.getType().cast<RankedTensorType>();
auto inputShape = inputTy.getShape();
auto inputRank = inputTy.getRank();
Value weight = adaptor.weight();
Value bias = adaptor.bias();
if (!inputTy.hasStaticShape()) {
return op->emitError("dynamic shaped input is not supported");
}
SmallVector<int64_t> normalizedShape;
if (!matchPattern(op.normalized_shape(),
m_TorchConstantIntList(normalizedShape))) {
return rewriter.notifyMatchFailure(
op, "normalized_shape must be a list of const int");
}
double eps = 0;
if (!matchPattern(op.eps(), m_TorchConstantFloat(&eps))) {
return rewriter.notifyMatchFailure(op,
"non const float eps is unsupported");
}
if (failed(checkNotNone(rewriter, op, weight)) ||
failed(checkNotNone(rewriter, op, bias))) {
return op->emitError("none weight or bias is unsupported");
}
auto weightTy = weight.getType().cast<RankedTensorType>();
auto biasTy = bias.getType().cast<RankedTensorType>();
if (!inputTy.getElementType().isa<mlir::FloatType>() ||
!biasTy.getElementType().isa<mlir::FloatType>() ||
!weightTy.getElementType().isa<mlir::FloatType>()) {
return op->emitError("currently only float data type are supported");
}
int64_t normalizedShapeRank = normalizedShape.size();
if (weightTy.getRank() != normalizedShapeRank ||
biasTy.getRank() != normalizedShapeRank ||
inputRank < normalizedShapeRank || normalizedShapeRank < 1) {
return rewriter.notifyMatchFailure(op, "input or weight or bias shape or"
"normalized shape not compatible");
}
for (int64_t i = 1; i <= normalizedShapeRank; i++) {
if (inputShape[inputRank - i] != normalizedShape[normalizedShapeRank - i] ||
weightTy.getShape()[normalizedShapeRank - i] !=
normalizedShape[normalizedShapeRank - i] ||
biasTy.getShape()[normalizedShapeRank - i] !=
normalizedShape[normalizedShapeRank - i]) {
return op.emitError("mismatching contracting dimension");
}
}
// Flatten dims to fit batch_norm operation.
int64_t numFeatureDimSize = 1;
int64_t numEmbeddingDimSize = 1;
for (int64_t i = 0; i < inputRank - normalizedShapeRank; i++) {
numFeatureDimSize *= inputShape[i];
}
for (int64_t i = 0; i < normalizedShapeRank; i++) {
numEmbeddingDimSize *= normalizedShape[i];
}
SmallVector<int64_t> inputFlattenShape{1, numFeatureDimSize,
numEmbeddingDimSize};
SmallVector<int64_t> meanOrVarMhloOutShape{numFeatureDimSize};
auto mhloBatchNormOutTy =
RankedTensorType::get(inputFlattenShape, inputTy.getElementType());
auto mhloBathNormOutMeanOrVarTy =
RankedTensorType::get(meanOrVarMhloOutShape, inputTy.getElementType());
// Reshape input
auto mhloInput = rewriter.create<mhlo::DynamicReshapeOp>(
op->getLoc(), mhloBatchNormOutTy, input,
mhlo::getConstTensor(rewriter, op, llvm::makeArrayRef(inputFlattenShape),
{static_cast<int64_t>(inputFlattenShape.size())})
.value());
// Generate "scale" and "offset" Value for mhlo.BatchNormTrainingOp.
SmallVector<APFloat> zeroConstVec(
numFeatureDimSize, APFloat::getZero(inputTy.getElementType()
.cast<mlir::FloatType>()
.getFloatSemantics()));
SmallVector<APFloat> oneConstVec(
numFeatureDimSize,
APFloat(
inputTy.getElementType().cast<mlir::FloatType>().getFloatSemantics(),
1));
auto oneOrZeroConstType =
RankedTensorType::get({numFeatureDimSize}, inputTy.getElementType());
Value scale = rewriter.create<mhlo::ConstantOp>(
op->getLoc(), oneOrZeroConstType,
DenseElementsAttr::get(oneOrZeroConstType, oneConstVec));
Value offset = rewriter.create<mhlo::ConstantOp>(
op->getLoc(), oneOrZeroConstType,
DenseElementsAttr::get(oneOrZeroConstType, zeroConstVec));
auto batchNormTrainingResult = rewriter.create<mhlo::BatchNormTrainingOp>(
op->getLoc(), mhloBatchNormOutTy, mhloBathNormOutMeanOrVarTy,
mhloBathNormOutMeanOrVarTy, mhloInput, scale, offset,
rewriter.getF32FloatAttr(eps), rewriter.getI64IntegerAttr(1));
// Reshape back
auto outputTy =
getTypeConverter()->convertType(op.getType(0)).cast<RankedTensorType>();
auto outputMeanOrVarTy =
getTypeConverter()->convertType(op.getType(1)).cast<RankedTensorType>();
auto output = rewriter.create<mhlo::DynamicReshapeOp>(
op->getLoc(), outputTy, batchNormTrainingResult.getResult(0),
mhlo::getConstTensor(rewriter, op, outputTy.getShape(),
{static_cast<int64_t>(outputTy.getShape().size())})
.value());
auto mean = rewriter.create<mhlo::DynamicReshapeOp>(
op->getLoc(), outputMeanOrVarTy, batchNormTrainingResult.getResult(1),
mhlo::getConstTensor(
rewriter, op, outputMeanOrVarTy.getShape(),
{static_cast<int64_t>(outputMeanOrVarTy.getShape().size())})
.value());
auto var = rewriter.create<mhlo::DynamicReshapeOp>(
op->getLoc(), outputMeanOrVarTy, batchNormTrainingResult.getResult(2),
mhlo::getConstTensor(
rewriter, op, outputMeanOrVarTy.getShape(),
{static_cast<int64_t>(outputMeanOrVarTy.getShape().size())})
.value());
// Apply affine transform: output x weight + bias [element-wise]
auto bcastedWeight = mhlo::promoteAndBroadcast(rewriter, weight, outputTy);
auto bcastedBias = mhlo::promoteAndBroadcast(rewriter, bias, outputTy);
auto outputMulWeight =
rewriter.create<mhlo::MulOp>(op->getLoc(), output, bcastedWeight);
auto finalOuput =
rewriter.create<mhlo::AddOp>(op->getLoc(), outputMulWeight, bcastedBias);
rewriter.replaceOp(op, {finalOuput, mean, var});
return success();
}
} // namespace
// AtenCatOp
namespace {
template <>
LogicalResult ConvertAtenOp<AtenCatOp>::matchAndRewrite(
AtenCatOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto outType =
getTypeConverter()->convertType(op.getType()).cast<RankedTensorType>();
int64_t dim;
if (!matchPattern(op.dim(), m_TorchConstantInt(&dim))) {
return rewriter.notifyMatchFailure(op,
"only constant dim param is supported");
}
SmallVector<Value> torchTensors;
if (!getListConstructElements(op.tensors(), torchTensors)) {
return rewriter.notifyMatchFailure(
op, "input should comes from a PrimListConstructOp");
}
SmallVector<Value> builtinTensors = getTypeConvertedValues(
rewriter, op->getLoc(), getTypeConverter(), torchTensors);
// Promote type
for (auto &v : builtinTensors) {
v = mhlo::promoteType(rewriter, v, outType);
}
rewriter.replaceOpWithNewOp<mhlo::ConcatenateOp>(
op, ValueRange(builtinTensors), static_cast<uint64_t>(dim));
return success();
}
} // namespace
void mlir::torch::torch_to_mhlo::populateBasicOpPatternsAndLegality(
TypeConverter &typeConverter, RewritePatternSet &patterns,
ConversionTarget &target) {
MLIRContext *context = patterns.getContext();
target.addIllegalOp<AtenTransposeIntOp>();
patterns.add<ConvertAtenTransposeIntOp>(typeConverter, context);
target.addIllegalOp<AtenBroadcastToOp>();
patterns.add<ConvertAtenBroadcastToOp>(typeConverter, context);
target.addIllegalOp<AtenPermuteOp>();
patterns.add<ConvertAtenPermuteOp>(typeConverter, context);
#define INSERT_UNARY_FPONLY_PATTERN(AtenOp, MhloOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenUnaryFPOnlyOp<AtenOp, MhloOp>>(typeConverter, \
context);
INSERT_UNARY_FPONLY_PATTERN(AtenLogOp, mhlo::LogOp);
INSERT_UNARY_FPONLY_PATTERN(AtenExpOp, mhlo::ExpOp);
INSERT_UNARY_FPONLY_PATTERN(AtenCloneOp, mhlo::CopyOp);
INSERT_UNARY_FPONLY_PATTERN(AtenSqrtOp, mhlo::SqrtOp);
INSERT_UNARY_FPONLY_PATTERN(AtenNegOp, mhlo::NegOp);
#undef INSERT_UNARY_FPONLY_PATTERN
#define INSERT_CONSTANT_FILL_PATTERN(AtenOp, fillVal) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenConstPatternOp<AtenOp, fillVal>>(typeConverter, \
context);
INSERT_CONSTANT_FILL_PATTERN(AtenOnesOp, 1);
INSERT_CONSTANT_FILL_PATTERN(AtenZerosOp, 0);
#undef INSERT_CONSTANT_FILL_PATTERN
#define INSERT_BINARY_ADDSUB_PATTERN(AtenOp, ChloOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenAddSubOp<AtenOp, ChloOp>>(typeConverter, context);
INSERT_BINARY_ADDSUB_PATTERN(AtenAddTensorOp, chlo::BroadcastAddOp);
INSERT_BINARY_ADDSUB_PATTERN(AtenAddScalarOp, chlo::BroadcastAddOp);
INSERT_BINARY_ADDSUB_PATTERN(AtenSubTensorOp, chlo::BroadcastSubOp);
INSERT_BINARY_ADDSUB_PATTERN(AtenSubScalarOp, chlo::BroadcastSubOp);
INSERT_BINARY_ADDSUB_PATTERN(AtenRsubScalarOp, chlo::BroadcastSubOp);
#undef INSERT_BINARY_ADDSUB_PATTERN
#define INSERT_BINARY_MULDIV_PATTERN(AtenOp, ChloOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenMulDivOp<AtenOp, ChloOp>>(typeConverter, context);
INSERT_BINARY_MULDIV_PATTERN(AtenMulTensorOp, chlo::BroadcastMulOp);
INSERT_BINARY_MULDIV_PATTERN(AtenMulScalarOp, chlo::BroadcastMulOp);
INSERT_BINARY_MULDIV_PATTERN(AtenDivTensorOp, chlo::BroadcastDivOp);
INSERT_BINARY_MULDIV_PATTERN(AtenDivTensorModeOp, chlo::BroadcastDivOp);
INSERT_BINARY_MULDIV_PATTERN(AtenDivScalarOp, chlo::BroadcastDivOp);
#undef INSERT_BINARY_MULDIV_PATTERN
#define INSERT_BINARY_COMPARE_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenCompareOp<AtenOp>>(typeConverter, context);
INSERT_BINARY_COMPARE_PATTERN(AtenGtTensorOp);
INSERT_BINARY_COMPARE_PATTERN(AtenGtScalarOp);
INSERT_BINARY_COMPARE_PATTERN(AtenLtTensorOp);
INSERT_BINARY_COMPARE_PATTERN(AtenLtScalarOp);
INSERT_BINARY_COMPARE_PATTERN(AtenEqTensorOp);
INSERT_BINARY_COMPARE_PATTERN(AtenEqScalarOp);
INSERT_BINARY_COMPARE_PATTERN(AtenNeTensorOp);
INSERT_BINARY_COMPARE_PATTERN(AtenNeScalarOp);
#undef INSERT_BINARY_COMPARE_PATTERN
#define INSERT_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenOp<AtenOp>>(typeConverter, context);
INSERT_ATENOP_PATTERN(AtenTanhOp);
INSERT_ATENOP_PATTERN(ValueTensorLiteralOp);
INSERT_ATENOP_PATTERN(AtenReciprocalOp);
INSERT_ATENOP_PATTERN(PrimNumToTensorScalarOp);
INSERT_ATENOP_PATTERN(AtenContiguousOp);
INSERT_ATENOP_PATTERN(AtenReluOp);
INSERT_ATENOP_PATTERN(AtenGeluOp);
INSERT_ATENOP_PATTERN(AtenErfOp);
INSERT_ATENOP_PATTERN(AtenCatOp);
INSERT_ATENOP_PATTERN(AtenBatchNormOp);
INSERT_ATENOP_PATTERN(AtenNativeLayerNormOp);
#undef INSERT_ATENOP_PATTERN
}