llvm
/
torch-mlir
mirror of https://github.com/llvm/torch-mlir
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
torch-mlir/lib/Conversion/TorchToTosa/TorchToTosa.cpp

5370 lines
210 KiB
C++
Raw Blame 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/TorchToTosa/TorchToTosa.h"
#include "../PassDetail.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Tosa/IR/TosaOps.h"
#include "mlir/IR/Matchers.h"
#include "mlir/Support/LogicalResult.h"
#include "mlir/Transforms/DialectConversion.h"
#include "torch-mlir/Conversion/TorchToTosa/TosaLegalizeCommon.h"
#include "torch-mlir/Conversion/TorchToTosa/TosaLegalizeUtils.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/IR/TorchTypes.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "torch-mlir/Dialect/TorchConversion/IR/TorchConversionDialect.h"
#include "torch-mlir/Dialect/TorchConversion/Transforms/BackendTypeConversion.h"
#include <optional>
using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
namespace {
// These legalizations are for unary ops with only for floating point datatypes.
// There is no supported quantized integer mode for these.
template <typename AtenOpT, typename TosaOpT>
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.getSelf();
auto selfTy = cast<TensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
if (selfTy.getElementType().isa<mlir::FloatType>()) {
rewriter.replaceOpWithNewOp<TosaOpT>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
op.getType()),
self);
return success();
} else {
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization supported");
}
}
};
// These unary op legalizations are identical for floating-point
// or quantized types
template <typename AtenOpT, typename TosaOpT>
class ConvertAtenUnaryOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
rewriter.replaceOpWithNewOp<TosaOpT>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
op.getType()),
adaptor.getSelf());
return success();
}
};
// These binary op legalizations are identical for floating-point
// or quantized types
template <typename AtenOpT, typename TosaOpT>
class ConvertAtenBinaryOp : 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.getSelf();
auto lhsTy = cast<TensorType>(lhs.getType());
Value rhs = adaptor.getOther();
auto rhsTy = cast<TensorType>(rhs.getType());
if (!lhsTy || !rhsTy)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
auto outTy = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
auto binaryOp =
tosa::createBinaryOpAndCast<TosaOpT>(rewriter, op, outTy, lhs, rhs);
rewriter.replaceOp(op, binaryOp.getResult());
return success();
}
};
template <typename T>
static bool isInValidRange(bool isFloat, const double &doubleValue, bool isInt,
const int64_t &intValue) {
if (isFloat) {
// Do a round-trip check here instead of numeric limits due to
// compiler warnings around double <-> int conversion.
return (doubleValue == static_cast<double>(static_cast<T>(doubleValue)));
} else {
assert(isInt);
return (intValue >= static_cast<int64_t>(std::numeric_limits<T>::min())) &&
(intValue <= static_cast<int64_t>(std::numeric_limits<T>::max()));
}
return true;
}
// FIXME: This will eventually go into a Tosa*Utils file.
LogicalResult torchScalarToTosaTensor(ConversionPatternRewriter &rewriter,
Operation *op, Value torchScalarValue,
Value &tosaTensor, Type dtype,
llvm::ArrayRef<int64_t> dshape) {
// Retrieve a const float or int value but create the out Tensor with dtype.
double doubleValue;
auto isFloat =
matchPattern(torchScalarValue, m_TorchConstantFloat(&doubleValue));
int64_t intValue;
auto isInt = matchPattern(torchScalarValue, m_TorchConstantInt(&intValue));
if (!isFloat && !isInt)
return rewriter.notifyMatchFailure(op,
"Unable to extract the scalar constant");
if (isa<mlir::FloatType>(dtype)) {
tosaTensor = tosa::getConstTensor<float>(rewriter, op,
(isFloat ? doubleValue : intValue),
dshape, dtype)
.value();
} else if (auto intType = dyn_cast<mlir::IntegerType>(dtype)) {
auto w = intType.getWidth();
if (w != 1 && w != 32 && w != 64)
return rewriter.notifyMatchFailure(op, [&](Diagnostic &diag) {
diag << "Unsupported integer type: " << intType;
});
if (w == 1) {
if (!isInValidRange<bool>(isFloat, doubleValue, isInt, intValue)) {
return rewriter.notifyMatchFailure(
op, "Supplied value of scalar constant exceeds limits "
"of destination type");
}
bool d = isFloat ? static_cast<bool>(doubleValue)
: static_cast<bool>(intValue);
tosaTensor =
tosa::getConstTensor<bool>(rewriter, op, {d}, dshape).value();
} else if (w == 32) {
if (!isInValidRange<int32_t>(isFloat, doubleValue, isInt, intValue)) {
return rewriter.notifyMatchFailure(
op, "Supplied value of scalar constant exceeds limits "
"of destination type");
}
int32_t d = isFloat ? static_cast<int32_t>(doubleValue)
: static_cast<int32_t>(intValue);
tosaTensor =
tosa::getConstTensor<int32_t>(rewriter, op, {d}, dshape).value();
} else if (w == 64) {
if (!isInValidRange<int64_t>(isFloat, doubleValue, isInt, intValue)) {
return rewriter.notifyMatchFailure(
op, "Supplied value of scalar constant exceeds limits "
"of destination type");
}
int64_t d = (isFloat ? static_cast<int64_t>(doubleValue) : intValue);
tosaTensor =
tosa::getConstTensor<int64_t>(rewriter, op, {d}, dshape).value();
}
} else {
return rewriter.notifyMatchFailure(op, "Usupported element type");
}
return success();
}
LogicalResult torchAlphaToTosaTensor(ConversionPatternRewriter &rewriter,
Operation *op, Value alphaScalar,
Value &alphaTensor, Type dtype,
bool checkForUnity) {
if (succeeded(torchScalarToTosaTensor(rewriter, op, alphaScalar, alphaTensor,
dtype, {})))
return success();
// `alpha` has not been specified.
int64_t alphaValue;
if (!matchPattern(alphaScalar, m_TorchConstantInt(&alphaValue)))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"alpha in TOSA operation");
// When no alpha has been specified, this must be 1.
if (checkForUnity && alphaValue != 1)
return rewriter.notifyMatchFailure(op,
"Unsupported integer value for alpha");
alphaTensor = tosa::getConstTensor<float>(
rewriter, op, {static_cast<float>(alphaValue)}, {}, dtype)
.value();
return success();
}
// These binary op legalizations are specific to add/sub which have an
// alpha multiplier.
template <typename AtenOpT, typename TosaOpT>
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 {
// left : tensor: tensor<i32/i64/f32>
// right : scalar: i32/i64/f32
// tensor: tensor<i32/i64/f32>
// alpha : scalar: i32/i64/f32
// output: tensor: tensor<i32/i64/f32>
Value lhs = adaptor.getSelf();
auto lhsType = dyn_cast<TensorType>(lhs.getType());
Value rhs = adaptor.getOther();
auto rhsType = dyn_cast<TensorType>(rhs.getType());
if (!lhsType)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
if (auto lhsElemTy = dyn_cast<IntegerType>(lhsType.getElementType())) {
if (lhsElemTy.getWidth() > 64)
return rewriter.notifyMatchFailure(
op, "Integers with widths greater than 64 are not supported");
}
// Get output type: tensor<i32/i64/f32>
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
Type rhsAlphaMulElemType;
if (isa<mlir::FloatType>(outElemTy)) {
rhsAlphaMulElemType = outElemTy;
} else {
// if output type is 64, input type should also be 32
rhsAlphaMulElemType = rewriter.getIntegerType(32);
}
// if right is scalar, rhgType==None, which need to be manually cast to
// TensorType else right is tensor, rhsType==tensor<i32/i64/f32>
Value rhsAsTensor;
if (!rhsType) {
if (failed(torchScalarToTosaTensor(rewriter, op, op.getOther(),
rhsAsTensor, rhsAlphaMulElemType, {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA operation");
} else if (rhsType.getElementType() != rhsAlphaMulElemType) {
// right is tensor, rhsType == tensor<i32/i64/f32>
// right must be cast to same type as the alpha, so MulOp success
rhs = rewriter.create<tosa::CastOp>(
op->getLoc(),
RankedTensorType::get(rhsType.getShape(), rhsAlphaMulElemType), rhs);
// reinitialize right value type to tensor<i32/f32>
rhsType = dyn_cast<TensorType>(rhs.getType());
}
auto rhsTensor = rhsType ? rhs : rhsAsTensor;
// Handle scalar value alpha.
// It should be either f32/i32
Value alphaTensor;
if (failed(torchAlphaToTosaTensor(rewriter, op.getOperation(),
op.getAlpha(), alphaTensor,
rhsAlphaMulElemType,
/*checkForUnity=*/false))) {
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"alpha in conversion to TOSA operation");
}
auto mulAlphaOp = tosa::createMulOpAndCast(
rewriter, op,
rhsType ? rhsType : RankedTensorType::get({}, rhsAlphaMulElemType),
rhsTensor, alphaTensor, /*shift=*/0);
if (outElemTy.isInteger(64)) {
// Tosa doesn't support 64-bit elementwise addition and subtraction.
// if outElemTy tensor<i64>, mulTensor must be tensor<i32>,
// left value could be tensor<f32/i32/i64> type, cast left value to
// tensor<i32> type
auto addOrSubi64Op = tosa::createBinaryOpAndCast<TosaOpT>(
rewriter, op,
RankedTensorType::get(outType.getShape(), rhsAlphaMulElemType), lhs,
mulAlphaOp);
// cast tensor<i32> back to tensor<i64>
rewriter.replaceOpWithNewOp<tosa::CastOp>(op, outType, addOrSubi64Op);
return success();
}
auto binaryOp = tosa::createBinaryOpAndCast<TosaOpT>(rewriter, op, outType,
lhs, mulAlphaOp);
rewriter.replaceOp(op, binaryOp.getResult());
return success();
}
}; // namespace
// Binary op legalizations for comparator ops.
template <typename AtenOpT, typename TosaOpT>
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.getSelf();
auto lhsTy = dyn_cast<TensorType>(lhs.getType());
Value rhs = adaptor.getOther();
auto rhsTy = dyn_cast<TensorType>(rhs.getType());
if (!lhsTy)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
auto lhsElemTy = lhsTy.getElementType();
if (!lhsElemTy.isIntOrFloat())
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
// For bitwise operators, only integer datatype legalization is supported
constexpr bool isBitwiseOp =
std::is_same<AtenOpT, AtenBitwiseAndTensorOp>() ||
std::is_same<AtenOpT, AtenBitwiseOrTensorOp>() ||
std::is_same<AtenOpT, AtenBitwiseXorTensorOp>();
if (isa<mlir::FloatType>(lhsElemTy) && isBitwiseOp) {
return rewriter.notifyMatchFailure(op,
"For bitwise operators, only integer "
"datatype legalization is supported");
}
Value rhsAsTensor;
if (!rhsTy) {
if (failed(torchScalarToTosaTensor(rewriter, op, op.getOther(),
rhsAsTensor, lhsElemTy, {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA operation");
}
auto rhsTensor = rhsTy ? rhs : rhsAsTensor;
// There is no Lesser operator in TOSA.
constexpr auto swapLhsRhs = (std::is_same<AtenOpT, AtenLtTensorOp>() ||
std::is_same<AtenOpT, AtenLtScalarOp>());
// Promote lhs and rhs dtypes for bitwise operators.
TensorType resultTy = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
if (isBitwiseOp) {
lhs = tosa::promoteType(rewriter, lhs, resultTy);
rhsTensor = tosa::promoteType(rewriter, rhsTensor, resultTy);
}
auto resultOp = rewriter.create<TosaOpT>(op.getLoc(), resultTy,
(swapLhsRhs ? rhsTensor : lhs),
(swapLhsRhs ? lhs : rhsTensor));
// There is no NE operator in TOSA.
if constexpr (std::is_same<AtenOpT, AtenNeTensorOp>() ||
std::is_same<AtenOpT, AtenNeScalarOp>()) {
rewriter.replaceOpWithNewOp<tosa::LogicalNotOp>(op, resultTy,
resultOp.getResult());
}
else {
rewriter.replaceOp(op, resultOp.getResult());
}
return success();
}
};
// Binary op legalizations for Mul variants.
template <typename AtenOpT>
class ConvertAtenMulOp : 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.getSelf();
auto lhsType = dyn_cast<TensorType>(lhs.getType());
if (!lhsType)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat())
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
Value rhsTensor;
if constexpr (std::is_same<AtenOpT, AtenSquareOp>()) {
rhsTensor = lhs;
} else {
Value rhsAsTensor;
Value rhs = adaptor.getOther();
auto rhsType = dyn_cast<TensorType>(rhs.getType());
if (!rhsType) {
if (failed(torchScalarToTosaTensor(rewriter, op, op.getOther(),
rhsAsTensor, outElemTy, {}))) {
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA operation");
}
}
rhsTensor = rhsType ? rhs : rhsAsTensor;
}
if (isa<mlir::FloatType>(outElemTy) || isa<mlir::IntegerType>(outElemTy)) {
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
auto mulOp = tosa::createMulOpAndCast(rewriter, op, outType, lhs,
rhsTensor, /*shift=*/0);
rewriter.replaceOp(op, mulOp.getResult());
return success();
}
// Quantized multiplication may need to rescale inputs.
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype "
"legalization currently supported");
}
};
template <typename AtenOpT>
class ConvertAtenDivOp : 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.getSelf();
auto lhsTy = dyn_cast<TensorType>(lhs.getType());
Value rhs = adaptor.getOther();
auto rhsTy = dyn_cast<TensorType>(rhs.getType());
if (!lhsTy)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
auto lhsElemTy = lhsTy.getElementType();
if (!lhsElemTy.isIntOrFloat())
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
Value rhsAsTensor;
if (!rhsTy) {
if (failed(torchScalarToTosaTensor(rewriter, op, op.getOther(),
rhsAsTensor, lhsElemTy, {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA operation");
}
auto rhsTensor = rhsTy ? rhs : rhsAsTensor;
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<TensorType>();
// auto result;
Value result;
if (isa<mlir::FloatType>(outType.getElementType())) {
// The input to the reciprocal is an integer sometimes, and we may need to
// promote it to a floating point. Per TOSA specification, the input types
// can only be floating point for tosa::ReciprocalOp.
Value rhsCasted = tosa::promoteType(rewriter, rhsTensor, outType);
auto rcpOp = rewriter.create<tosa::ReciprocalOp>(
op->getLoc(), rhsCasted.getType(), rhsCasted);
result = tosa::createMulOpAndCast(rewriter, op, outType, lhs,
rcpOp.getResult(), /*shift=*/0)
.getResult();
} else {
// The output type can be different than the input types (e.g. dividing an
// int tensor results in a floating point tensor).
result = tosa::createBinaryOpAndCast<tosa::IntDivOp>(
rewriter, op, outType, lhs, rhsTensor)
.getResult();
}
rewriter.replaceOp(op, {result});
return success();
}
};
// This defines a template to construct ops whose legalizations are
// specialized.
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;
};
template <>
LogicalResult ConvertAtenOp<AtenTanhOp>::matchAndRewrite(
AtenTanhOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<TensorType>(self.getType());
if (selfTy && selfTy.getElementType().isa<mlir::FloatType>()) {
rewriter.replaceOpWithNewOp<tosa::TanhOp>(
op, getTypeConverter()->convertType(op.getType()), self);
return success();
}
// Sigmoid legalization in TOSA for quantized element-type uses specialized
// tosa.table construct.
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization currently supported");
}
template <>
LogicalResult ConvertAtenOp<AtenSigmoidOp>::matchAndRewrite(
AtenSigmoidOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<TensorType>(self.getType());
if (selfTy && selfTy.getElementType().isa<mlir::FloatType>()) {
rewriter.replaceOpWithNewOp<tosa::SigmoidOp>(
op, getTypeConverter()->convertType(op.getType()), self);
return success();
}
// Sigmoid legalization in TOSA for quantized element-type uses
// specialized tosa.table construct.
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization currently supported");
}
template <>
LogicalResult ConvertAtenOp<AtenReluOp>::matchAndRewrite(
AtenReluOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<TensorType>(self.getType());
// Maps to tosa.clamp which has both int and fp limits.
int64_t clampMin = 0;
Value clampIn = self;
if (!selfTy) {
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
}
// Rescale the clampIn for quantized types. TBD
if (!selfTy.getElementType().isa<mlir::FloatType>()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization currently supported");
}
rewriter.replaceOpWithNewOp<tosa::ClampOp>(
op, getTypeConverter()->convertType(op.getType()), clampIn,
rewriter.getI64IntegerAttr(clampMin),
rewriter.getI64IntegerAttr(std::numeric_limits<int32_t>::max()),
rewriter.getF32FloatAttr(0.0f),
rewriter.getF32FloatAttr(std::numeric_limits<float>::max()));
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenLeakyReluOp>::matchAndRewrite(
AtenLeakyReluOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<TensorType>(self.getType());
if (!selfTy.getElementType().isa<mlir::FloatType>()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization currently supported");
}
Value alphaScalar = op.getNegativeSlope();
Value alphaTensor;
if (failed(torchScalarToTosaTensor(rewriter, op.getOperation(), alphaScalar,
alphaTensor, selfTy.getElementType(), {})))
return rewriter.notifyMatchFailure(
op, "Negative slope needs to be a scalar constant for conversion to "
"TOSA LeakyReLU operation");
auto zero =
tosa::getConstTensor<float>(rewriter, op, 0, {}, selfTy.getElementType())
.value();
auto cond = rewriter.create<tosa::GreaterEqualOp>(
op->getLoc(),
RankedTensorType::get(selfTy.getShape(), rewriter.getIntegerType(1)),
self, zero);
auto mulTensor = rewriter.create<tosa::MulOp>(
op->getLoc(), getTypeConverter()->convertType(op.getType()), self,
alphaTensor, /*shift=*/0);
rewriter.replaceOpWithNewOp<tosa::SelectOp>(
op, getTypeConverter()->convertType(op.getType()), cond, self, mulTensor);
return success();
}
using ReductionConvFunc = std::optional<Value> (*)(PatternRewriter &,
Operation *,
RankedTensorType, Value,
ElementsAttr, bool);
// They all constitute a common form invoking the appropriate
// converion function in TosaLegalizeCommon.cpp
template <typename AtenOpT, ReductionConvFunc ConversionFuncT>
class ConvertAtenReductionOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
// Each variant must implement corresponding parameter parsing options
virtual LogicalResult readReduceDimsAndKeepDims(
AtenOpT op, OpAdaptor adaptor, ConversionPatternRewriter &rewriter,
ElementsAttr &reduceDimsAttr, bool &keepDims) const {
return rewriter.notifyMatchFailure(
op, "Unimplemented reduce_dims and keep_dims parsing function");
}
// Common rewriter for all reduction ops, calls the specific implementation of
// readReduceDimsAndKeepDims() needed for the op variant.
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value self = adaptor.getSelf();
auto selfTy = cast<TensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
auto outputTy = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<RankedTensorType>();
if (!outputTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor type outputs permitted for reduce_mean");
ElementsAttr reduceDimsAttr;
bool keepDims;
if (failed(readReduceDimsAndKeepDims(op, adaptor, rewriter, reduceDimsAttr,
keepDims)))
return failure();
std::optional<Value> result =
ConversionFuncT(rewriter, op, outputTy, self, reduceDimsAttr, keepDims);
if (!result)
return failure();
// TBD - support dtype casting.
rewriter.replaceOp(op, {result.value()});
return success();
}
};
// This reduction op legalization template handles op variants that have
// explicit reduce_dims dimensions (provided as a list) and keep_dims
// parameters.
template <typename AtenOpT, ReductionConvFunc ConversionFuncT>
class ConvertAtenMultipleDimsReductionOp
: public ConvertAtenReductionOp<AtenOpT, ConversionFuncT> {
using ConvertAtenReductionOp<AtenOpT,
ConversionFuncT>::ConvertAtenReductionOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult readReduceDimsAndKeepDims(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
ElementsAttr &reduceDimsAttr,
bool &keepDims) const override {
SmallVector<int64_t, 4> reduceDims;
if (!matchPattern(op.getDim(), m_TorchListOfConstantInts(reduceDims)))
return rewriter.notifyMatchFailure(op,
"non-const dim parameter unsupported");
int64_t N = reduceDims.size();
int64_t inputRank =
cast<RankedTensorType>(adaptor.getSelf().getType()).getRank();
for (unsigned i = 0; i < N; i++) {
reduceDims[i] = toPositiveDim(reduceDims[i], inputRank);
if (!isValidDim(reduceDims[i], inputRank))
return rewriter.notifyMatchFailure(op,
"reduce dim is statically invalid");
}
auto reduceDimsType = RankedTensorType::get({N}, rewriter.getI64Type());
reduceDimsAttr =
DenseIntElementsAttr::get(reduceDimsType, llvm::ArrayRef(reduceDims));
keepDims = false;
if (!matchPattern(op.getKeepdim(), m_TorchConstantBool(&keepDims)))
return rewriter.notifyMatchFailure(
op, "non-const keepdim parameter unsupported");
return success();
}
};
// This reduction op legalization template handles op variants that reduce in
// only one explicit dim which is provided as a number (rather than a list), and
// a keep_dims parameter.
template <typename AtenOpT, ReductionConvFunc ConversionFuncT>
class ConvertAtenOneDimReductionOp
: public ConvertAtenReductionOp<AtenOpT, ConversionFuncT> {
using ConvertAtenReductionOp<AtenOpT,
ConversionFuncT>::ConvertAtenReductionOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult readReduceDimsAndKeepDims(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
ElementsAttr &reduceDimsAttr,
bool &keepDims) const override {
int64_t reduceDim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&reduceDim)))
return rewriter.notifyMatchFailure(op,
"non-const dim parameter unsupported");
int64_t inputRank =
cast<RankedTensorType>(adaptor.getSelf().getType()).getRank();
reduceDim = toPositiveDim(reduceDim, inputRank);
if (!isValidDim(reduceDim, inputRank))
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
auto reduceDimsType = RankedTensorType::get({1}, rewriter.getI64Type());
reduceDimsAttr =
DenseIntElementsAttr::get(reduceDimsType, llvm::ArrayRef({reduceDim}));
keepDims = false;
if (!matchPattern(op.getKeepdim(), m_TorchConstantBool(&keepDims)))
return rewriter.notifyMatchFailure(
op, "non-const keepdim parameter unsupported");
return success();
}
};
// This reduction op legalization template handles op variants that reduce all
// dims does not keep dims.
template <typename AtenOpT, ReductionConvFunc ConversionFuncT>
class ConvertAtenAllDimsReductionOp
: public ConvertAtenReductionOp<AtenOpT, ConversionFuncT> {
public:
using ConvertAtenReductionOp<AtenOpT,
ConversionFuncT>::ConvertAtenReductionOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult readReduceDimsAndKeepDims(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
ElementsAttr &reduceDimsAttr,
bool &keepDims) const override {
auto self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
// Select all dims to reduce
SmallVector<int64_t, 4> reduceDims;
for (int64_t i = 0; i < selfTy.getRank(); i++)
reduceDims.push_back(i);
int64_t N = selfTy.getRank();
auto reduceDimsType = RankedTensorType::get({N}, rewriter.getI64Type());
reduceDimsAttr =
DenseIntElementsAttr::get(reduceDimsType, llvm::ArrayRef(reduceDims));
keepDims = false;
return success();
}
};
template <>
LogicalResult ConvertAtenOp<AtenArgmaxOp>::matchAndRewrite(
AtenArgmaxOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA argmax");
int64_t reduceDim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&reduceDim))) {
// NoneType indicates reduce on all dims
reduceDim = -1;
} else {
int64_t inputRank = selfTy.getRank();
reduceDim = toPositiveDim(reduceDim, inputRank);
if (!isValidDim(reduceDim, inputRank))
return rewriter.notifyMatchFailure(op,
"reduce dim is statically invalid");
}
bool keepDim = false;
if (!matchPattern(op.getKeepdim(), m_TorchConstantBool(&keepDim)))
return rewriter.notifyMatchFailure(
op, "non-const keepdim parameter unsupported");
auto resultTy = getTypeConverter()
->convertType(op.getResult().getType())
.cast<RankedTensorType>();
auto outputETy = resultTy.getElementType();
// Create a single instance of tosa.argmax.
// Multiple dims require chained construct.
auto buildArgmax = [&](int64_t reduceDim, Value input) -> Value {
auto inputTy = cast<RankedTensorType>(input.getType());
auto inputShape = makeShapeTorchCompatible(inputTy.getShape());
SmallVector<int64_t> outputShapeArr = {};
int32_t i = 0;
for (auto &dim : inputShape) {
if (i++ != reduceDim) {
outputShapeArr.push_back(dim);
} else {
if (keepDim)
outputShapeArr.push_back(1);
}
}
// Tosa argmax output is i32, while Torch backend mandates i64.
auto outputReduceTy = RankedTensorType::get(
makeShapeLLVMCompatible(ArrayRef<int64_t>(outputShapeArr)),
rewriter.getI32Type());
auto reduceDimAttr =
rewriter.getIntegerAttr(rewriter.getI64Type(), reduceDim);
return rewriter
.create<tosa::ArgMaxOp>(op->getLoc(),
getTypeConverter()->convertType(outputReduceTy),
input, reduceDimAttr)
.getResult();
};
// Convert the final index to i64 for backend finalization, However, i64
// is not a defined type for tosa.cast, so using arith.extsi instead.
auto castToInt64 = [&](Value result) -> LogicalResult {
auto resTy = cast<ShapedType>(result.getType());
if (!resTy)
return rewriter.notifyMatchFailure(op,
"Argmax: Result is not a shaped type");
auto resShape = makeShapeTorchCompatible(resTy.getShape());
auto outTy =
RankedTensorType::get(makeShapeLLVMCompatible(resShape), outputETy);
rewriter.replaceOpWithNewOp<arith::ExtSIOp>(
op, getTypeConverter()->convertType(outTy), result);
return success();
};
if (reduceDim == -1) { // reducing on all dims
Value input = self;
for (int dim = 0; dim < selfTy.getRank(); dim++) {
// progressively reduce each 0-th dim
input = buildArgmax(0, input);
}
return castToInt64(input);
} else {
return castToInt64(buildArgmax(reduceDim, self));
}
return success();
}
template <typename AtenOpT>
class ConvertAtenSqueezeOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
// Each variant must implement corresponding parameter parsing options
virtual LogicalResult
generateSqueezedShape(AtenOpT op, RankedTensorType selfTy,
ConversionPatternRewriter &rewriter,
SmallVector<int64_t> &squeezedShape) const {
return rewriter.notifyMatchFailure(
op, "Unimplemented dim/dim-list parsing function");
}
// Common rewriter for all squeeze ops, calls the specific implementation of
// generateSqueezedShape() needed for the op variant.
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA argmax");
SmallVector<int64_t> newOutputShape;
if (failed(generateSqueezedShape(op, selfTy, rewriter, newOutputShape)))
return rewriter.notifyMatchFailure(op,
"Squeeze could not compute new shape");
auto resultTy = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getResult().getType())
.template cast<RankedTensorType>();
auto resultElemTy = resultTy.getElementType();
auto newOutputTy = RankedTensorType::get(
makeShapeLLVMCompatible(newOutputShape), resultElemTy);
auto reshapeOp = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
newOutputTy),
self, rewriter.getDenseI64ArrayAttr(newOutputShape));
rewriter.replaceOpWithNewOp<tensor::CastOp>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
newOutputTy),
reshapeOp);
return success();
}
};
template <typename AtenOpT>
class ConvertAtenSqueezeOneDimOp : public ConvertAtenSqueezeOp<AtenOpT> {
using ConvertAtenSqueezeOp<AtenOpT>::ConvertAtenSqueezeOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
generateSqueezedShape(AtenOpT op, RankedTensorType selfTy,
ConversionPatternRewriter &rewriter,
SmallVector<int64_t> &squeezedShape) const override {
int64_t squeezeDim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&squeezeDim)))
return rewriter.notifyMatchFailure(op,
"non-const dim parameter unsupported");
// Handle negative dim
if (squeezeDim < 0)
squeezeDim = squeezeDim + selfTy.getRank();
auto selfShape = makeShapeTorchCompatible(selfTy.getShape());
// Only dims statically known to have size=1 are reduced.
// Dynamic dims are treated as unknowns and will not be squeezed
// even if dim parameter says it should be.
uint32_t dimNum = 0;
for (auto &dim : selfShape) {
if (dim != 1 || squeezeDim != dimNum)
squeezedShape.push_back(dim);
dimNum++;
}
return success();
}
};
template <typename AtenOpT>
class ConvertAtenSqueezeAllDimsOp : public ConvertAtenSqueezeOp<AtenOpT> {
using ConvertAtenSqueezeOp<AtenOpT>::ConvertAtenSqueezeOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
generateSqueezedShape(AtenOpT op, RankedTensorType selfTy,
ConversionPatternRewriter &rewriter,
SmallVector<int64_t> &squeezedShape) const override {
auto selfShape = makeShapeTorchCompatible(selfTy.getShape());
// Dims that may dynamically resolve to 1 are not reduced here. Only
// compile-time resolvable dims are handled here.
for (auto &dim : selfShape) {
if (dim != 1)
squeezedShape.push_back(dim);
}
return success();
}
};
template <>
LogicalResult ConvertAtenOp<AtenPowTensorScalarOp>::matchAndRewrite(
AtenPowTensorScalarOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA Pow");
if (!selfTy.getElementType().isa<mlir::FloatType>())
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization supported");
auto outType =
cast<TensorType>(getTypeConverter()->convertType(op.getType()));
Value expTensor;
Value expScalar = op.getExponent();
if (failed(torchScalarToTosaTensor(rewriter, op, expScalar, expTensor,
outType.getElementType(), {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA Pow operation");
auto powOp = tosa::createBinaryOpAndCast<tosa::PowOp>(rewriter, op, outType,
self, expTensor);
rewriter.replaceOp(op, powOp.getResult());
return success();
}
// Perform the basic n-dim matmul operation encompassing the handling of
// broadcasting and dynamic shape propagation.
// All PyTorch ops that leverage matrix multiplication will derive this and
// implement their specialized input processing (e.g transpose), and output
// processing, e.g. GEMM or fully connected bias handling.
template <typename AtenOpT>
class ConvertAtenMatmulBaseOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
// Each variant must implement corresponding parameter parsing options.
// Maintain separate input read functions for each variant because it is not
// necessarily true with all variants that the first two operands are the lhs
// and rhs.
virtual LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
Value &lhs, Value &rhs) const {
return rewriter.notifyMatchFailure(
op,
"Unimplemented matrix multiplication variant input parsing function");
}
LogicalResult performMatmul(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter, Value &lhs,
Value &rhs, Value &output) const {
auto lhsTy = cast<RankedTensorType>(lhs.getType());
auto rhsTy = cast<RankedTensorType>(rhs.getType());
auto lhsRank = lhsTy.getRank();
auto rhsRank = rhsTy.getRank();
auto lhsShape = makeShapeTorchCompatible(lhsTy.getShape());
auto rhsShape = makeShapeTorchCompatible(rhsTy.getShape());
auto lhsElemTy = lhsTy.getElementType();
auto rhsElemTy = rhsTy.getElementType();
if (lhsElemTy != rhsElemTy)
return rewriter.notifyMatchFailure(op,
"Matmul: input datatypes mismatched");
// Legalization constructs may offer input shapes but expect output shapes
// to be inferred, e.g.
// func @forward(%arg0: !torch.vtensor<[14,19],f32>,
// %arg1: !torch.vtensor<[19,28],f32>) ->
// !torch.vtensor<[?,?],f32>
// This is tricky with matmul, since TOSA matmul is on 3D inputs.
// This means the need to reshape potentially both inputs and outputs,
// and reshape to unknown shape is undefined.
auto maxInputRank = lhsRank > rhsRank ? lhsRank : rhsRank;
// If performing dot product on vectors, the RHS is synthetically transposed
if (maxInputRank == 1)
maxInputRank++;
// Obtaining the rank broadcasted shapes of tensors makes it easier to
// construct the input and output reshaping logic.
auto getRankBroadcastedShape = [&](Value tensor,
bool isRHS) -> SmallVector<int64_t> {
auto tensorTy = cast<TensorType>(tensor.getType());
auto tensorShape = makeShapeTorchCompatible(tensorTy.getShape());
auto tensorRank = tensorTy.getRank();
SmallVector<int64_t> bcastedShape;
auto bcastDims = maxInputRank - tensorRank;
if (isRHS && (tensorRank == 1) && bcastDims) {
// RHS with rank1 is special. It be synthetically transposed to dim[:-2]
for (int32_t i = 0; i < bcastDims - 1; i++)
bcastedShape.push_back(1);
bcastedShape.push_back(tensorShape[0]);
bcastedShape.push_back(1);
} else {
if (bcastDims > 0) { // rank broadcast
for (uint32_t i = 0; i < bcastDims; i++)
bcastedShape.push_back(1);
}
for (auto &dim : tensorShape)
bcastedShape.push_back(dim);
}
return bcastedShape;
};
// Step: Rank broadcast the two inputs.
auto lhsBroadcastedShape = getRankBroadcastedShape(lhs, false);
auto lhsBroadcastedTy = RankedTensorType::get(
makeShapeLLVMCompatible(lhsBroadcastedShape), lhsElemTy);
auto rhsBroadcastedShape = getRankBroadcastedShape(rhs, true);
auto rhsBroadcastedTy = RankedTensorType::get(
makeShapeLLVMCompatible(rhsBroadcastedShape), rhsElemTy);
auto rankBroadcastedLhs =
lhsRank == maxInputRank
? lhs
: rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
lhsBroadcastedTy),
lhs, rewriter.getDenseI64ArrayAttr(lhsBroadcastedShape));
auto rankBroadcastedRhs =
rhsRank == maxInputRank
? rhs
: rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
rhsBroadcastedTy),
rhs, rewriter.getDenseI64ArrayAttr(rhsBroadcastedShape));
// TOSA matmul is performed on two 3D inputs and generates a 3D output.
// Lower ranked tensors are dim-1 reshaped up to 3D
auto reshapeUpTo3DTensor = [&](Value tensor) -> Value {
auto tensorTy = cast<TensorType>(tensor.getType());
auto rank = tensorTy.getRank();
assert(rank <= 3 && "reshapeUpTo3D tensor must receive rank <= 3");
if (rank == 3)
return tensor;
auto shape = makeShapeTorchCompatible(tensorTy.getShape());
SmallVector<int64_t> newShape({1, 1, 1});
if (rank == 2) { // batchsize = 1
newShape[1] = shape[0];
newShape[2] = shape[1];
} else { // rank 1
newShape[2] = shape[0];
}
auto newType = RankedTensorType::get(makeShapeLLVMCompatible(newShape),
tensorTy.getElementType());
return rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
newType),
tensor, rewriter.getDenseI64ArrayAttr(newShape));
};
// Where broadcasting is required in one or more batch dims, the following
// is done.
// Where all batch dims are involved in broadcasting:
// Given A: 3x1x5x6 and B: 1x4x6x7
// 1. Reshape A to 1x15x6 (squeeze all batchdims into dim1)
// 2. Transpose B to 6x1x4x7, Reshape to 1x6x28
// 3. tosa.Matmul 1x15x6 1x6x28 = 1x15x28
// 4. Reshape out to 3x5x4x7, Transpose to 3x4x5x7
// Where there are batch dimensions that are broadcast and not, the
// treatment is to have dim0 correspond to product of all non-broadcast
// dimsizes:
// Given A: 4x8x16x32 B: 8x32x17
// 1. Reshape A to 8x64x32 (squeeze all unbroadcasted dims into dim0,
// broadcasted dims into dim1)
// 2. No transpose or reshape of B as its batchdims are not broadcast to.
// 3. tosa.Matmul 8x64x32 8x32x17 = 8x64x17
// 4. Reshape to 8x4x16x17, Transpose to 4x8x16x17
// Check if we need to perform the broadcast on batch dim
// Not needed if max rank < 3, or if maxrank == 3 and dim[0] matches
auto needsBatchDimBroadcast = [&]() -> bool {
if (maxInputRank < 3) {
return false;
} else {
if (maxInputRank == 3 &&
lhsBroadcastedShape[0] == rhsBroadcastedShape[0]) {
return false;
}
return true;
}
};
auto performBatchDimBroadcast = needsBatchDimBroadcast();
// Inputs to the tosa.matmul
Value matmulLhs, matmulRhs;
using TensorShape_t = struct {
int64_t dim;
int64_t shape;
};
// Transpose needs to done if transposeDims are not non-monotonically
// increasing. E.g. [0, 1, 2, 3]: No transpose [1, 0, 2, 3]: Transpose dim0
// and dim1 The order need not be sequential, since one or more dims may
// have been removed due to broadcasting.
auto isTransposeRequired = [](SmallVector<int32_t> transposedDims) -> bool {
int32_t lastDim = -1;
for (auto &dim : transposedDims) {
if (lastDim > dim)
return true;
lastDim = dim;
}
return false;
};
SmallVector<TensorShape_t> batchElems, lhsSqueezedElems, rhsSqueezedElems;
if (!performBatchDimBroadcast) {
// Simple with no broadcasting artifacts. Just reshape up to 3D
matmulLhs = reshapeUpTo3DTensor(rankBroadcastedLhs);
matmulRhs = reshapeUpTo3DTensor(rankBroadcastedRhs);
} else {
// In this case, either or both input matrices involve broadcasting on
// their batch dimensions. For example:
// 4x5x6, 1x6x7 -> 4x5x7
// 4x1x5x6, 1x3x6x7 -> 4x3x5x7
// Though maxInputRank is necessarily >=3 here, individual matrices may be
// lower rank.
// E.g. 3x4x5x6, 6 -> 3x4x5
// These are the accumulated products of the shape of each dim:
// 1. common dimensions: upper dimensions (dims other than two rightmost)
// whose shapes are the same for both LHS and RHS.
// 2. LHS squeezed dimensions: all dimensions of LHS that involve
// broadcasting in either direction, plus the LHS[-2] shape
// 3. RHS squeezed dimensions: all dimensions of RHS that involve
// broadcasting in either direction, plus the RHS[-1] shape
int64_t commonValue = 1, lhsSqueezedValue = 1, rhsSqueezedValue = 1;
// For both LHS and RHS, the dimensions are separated into the common,
// squeezed and remaining dim. E.g. given
// LHS = 3x4x5x6
// RHS = 1x4x6x7
// common = {{dim=1, shape=4}}
// lhs squeezed = {{dim=0, shape=3},
// {dim=2, shape=5}}
// rhs squeezed = {{dim=0, shape=1},
// {dim=2, shape=7}}
// The matmul dim is LHS[-1] and RHS[-2], i.e. 6.
// Once this is obtained, LHS and RHS are expressed as:
// LHS = {common, lhs_squeezed, matmul_dim}
// RHS = {common, matmul_dim, rhs_squeezed}
// The matmul is then performed to obtain output:
// matmul_out = {common, lhs_squeezed, rhs_squeezed}
// Finally, we reshape to 'unsqueeze' the LHS and RHS parts and transpose
// them back to their correct positions.
SmallVector<int64_t> transposedLhsShape;
SmallVector<int32_t> transposedLhsDims;
// Step: generate the common dim/shape information
bool hasDynamicDims = false;
for (uint32_t dim = 0; dim < maxInputRank - 2; dim++) {
bool isDynamicDim = ShapedType::isDynamic(lhsBroadcastedShape[dim]);
hasDynamicDims |= isDynamicDim;
if (isDynamicDim ||
lhsBroadcastedShape[dim] == rhsBroadcastedShape[dim]) {
commonValue *= lhsBroadcastedShape[dim];
batchElems.push_back({dim, lhsBroadcastedShape[dim]});
}
}
commonValue = commonValue < 0 ? kUnknownSize : commonValue;
// TODO: Handle the case when there are dynamic batch dimensions.
if (hasDynamicDims)
commonValue = kUnknownSize;
// Step: generate the LHS squeezed dim/shape information.
for (uint32_t dim = 0; dim < maxInputRank - 2; dim++) {
bool isDynamicDim = ShapedType::isDynamic(lhsBroadcastedShape[dim]);
if (!isDynamicDim &&
lhsBroadcastedShape[dim] != rhsBroadcastedShape[dim]) {
lhsSqueezedValue *= lhsBroadcastedShape[dim];
lhsSqueezedElems.push_back({dim, lhsBroadcastedShape[dim]});
}
}
// including LHS[-2]
lhsSqueezedElems.push_back(
{maxInputRank - 2, lhsBroadcastedShape[maxInputRank - 2]});
lhsSqueezedValue *= lhsBroadcastedShape[maxInputRank - 2];
lhsSqueezedValue = lhsSqueezedValue < 0 ? kUnknownSize : lhsSqueezedValue;
// Step: Create the tosa.transpose array. If this array has a
// non-monotonic series of dims, perform transpose.
// First the common_elems
for (uint32_t i = 0; i < batchElems.size(); i++) {
transposedLhsShape.push_back(batchElems[i].shape);
transposedLhsDims.push_back(batchElems[i].dim);
}
// then the lhs_squeezed elems
for (uint32_t i = 0; i < lhsSqueezedElems.size(); i++) {
transposedLhsShape.push_back(lhsSqueezedElems[i].shape);
transposedLhsDims.push_back(lhsSqueezedElems[i].dim);
}
// then the final dim
transposedLhsDims.push_back(maxInputRank - 1);
transposedLhsShape.push_back(lhsBroadcastedShape[maxInputRank - 1]);
bool lhsNeedsTranspose = isTransposeRequired(transposedLhsDims);
auto lhsReshapeInput = rankBroadcastedLhs;
if (lhsNeedsTranspose) {
auto transposedLhsType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedLhsShape), rhsElemTy);
std::optional<Value> transposedLhsDimsConst =
tosa::getConstTensor<int32_t>(
rewriter, op,
/*vec=*/transposedLhsDims,
/*shape=*/{static_cast<int32_t>(transposedLhsDims.size())});
lhsReshapeInput =
rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(transposedLhsType),
rankBroadcastedLhs, transposedLhsDimsConst.value())
.getResult();
}
// LHS = {common, lhs_squeezed, matmul_dim}
SmallVector<int64_t> newLhsShape(
{1, 1, lhsBroadcastedShape[maxInputRank - 1]});
newLhsShape[0] = commonValue;
newLhsShape[1] = hasDynamicDims ? kUnknownSize : lhsSqueezedValue;
auto newLhsType = RankedTensorType::get(
makeShapeLLVMCompatible(newLhsShape), lhsElemTy);
matmulLhs = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
newLhsType),
lhsReshapeInput, rewriter.getDenseI64ArrayAttr(newLhsShape));
SmallVector<int64_t> transposedRhsShape;
SmallVector<int32_t> transposedRhsDims;
// Step: Create the RHS transpose sequence
// RHS = {common, matmul_dim, rhs_squeezed}
// first the common_dims
for (uint32_t i = 0; i < batchElems.size(); i++) {
transposedRhsShape.push_back(batchElems[i].shape);
transposedRhsDims.push_back(batchElems[i].dim);
}
// The matmul_dim of RHS
transposedRhsDims.push_back(maxInputRank - 2);
transposedRhsShape.push_back(rhsBroadcastedShape[maxInputRank - 2]);
// finally all the rhs_squeeze dims
hasDynamicDims = false;
for (uint32_t dim = 0; dim < maxInputRank - 2; dim++) {
bool isDynamicDim = ShapedType::isDynamic(rhsBroadcastedShape[dim]);
hasDynamicDims |= isDynamicDim;
if (!isDynamicDim &&
rhsBroadcastedShape[dim] != lhsBroadcastedShape[dim]) {
rhsSqueezedElems.push_back({dim, rhsBroadcastedShape[dim]});
rhsSqueezedValue *= rhsBroadcastedShape[dim];
}
}
rhsSqueezedElems.push_back(
{maxInputRank - 1, rhsBroadcastedShape[maxInputRank - 1]});
rhsSqueezedValue *= rhsBroadcastedShape[maxInputRank - 1];
for (uint32_t i = 0; i < rhsSqueezedElems.size(); i++) {
transposedRhsShape.push_back(rhsSqueezedElems[i].shape);
transposedRhsDims.push_back(rhsSqueezedElems[i].dim);
}
auto transposedRhsType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedRhsShape), rhsElemTy);
if (hasDynamicDims)
rhsSqueezedValue = kUnknownSize;
SmallVector<int64_t> newRhsShape(
{commonValue < 0 ? kUnknownSize : commonValue,
rhsBroadcastedShape[maxInputRank - 2], rhsSqueezedValue});
auto newRhsType = RankedTensorType::get(
makeShapeLLVMCompatible(newRhsShape), rhsElemTy);
bool rhsNeedsTranspose = isTransposeRequired(transposedRhsDims);
auto transposedRhsValue = rankBroadcastedRhs;
if (rhsNeedsTranspose) {
std::optional<Value> transposedRhsDimsConst =
tosa::getConstTensor<int32_t>(
rewriter, op,
/*vec=*/transposedRhsDims,
/*shape=*/{static_cast<int32_t>(transposedRhsDims.size())});
transposedRhsValue =
rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(transposedRhsType),
rankBroadcastedRhs, transposedRhsDimsConst.value())
.getResult();
}
// reshape
matmulRhs = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
newRhsType),
transposedRhsValue, rewriter.getDenseI64ArrayAttr(newRhsShape));
}
auto matmulLhsShape = makeShapeTorchCompatible(
cast<RankedTensorType>(matmulLhs.getType()).getShape());
auto matmulRhsShape = makeShapeTorchCompatible(
cast<RankedTensorType>(matmulRhs.getType()).getShape());
// The reshape/transpose should ensure the tosa.matmul always has same
// batch size for either matrix. If if shapes are dynamic, they'll be
// appropriately handled.
assert(matmulLhsShape[0] == matmulRhsShape[0] &&
"tosa.matmul needs same batchsize on LHS and RHS");
SmallVector<int64_t> matmulOutputShape(
{matmulLhsShape[0], matmulLhsShape[1], matmulRhsShape[2]});
Type outputElemTy;
if (isa<mlir::FloatType>(lhsElemTy)) {
outputElemTy = lhsElemTy;
} else { // qint8 emits i32 matmul output
outputElemTy = rewriter.getIntegerType(32);
}
auto mmOutputTy = RankedTensorType::get(
makeShapeLLVMCompatible(matmulOutputShape), outputElemTy);
auto mmOpResult =
rewriter
.create<tosa::MatMulOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
mmOutputTy),
matmulLhs, matmulRhs)
.getResult();
// Perform the reshape to output shape. This is always required unless max
// input rank=3 and there was no broadcasting, in which case the tosa.matmul
// output itself is correctly shaped.
bool performOpReshape = !(maxInputRank == 3 && !performBatchDimBroadcast);
if (performOpReshape) {
// Since the output shape may be unknown, we construct it
// independently and reshape. Otherwise reshape may be expressed for
// an unknown to-be-inferred output shape. The final tensor.cast
// reshapes the known shape to the desired output shape.
auto computeOpShape = [&](SmallVector<int64_t> &reshapedOpShape,
SmallVector<int32_t> &transposedOpDims,
SmallVector<int64_t> &transposedOpShapes) {
if (maxInputRank == 1)
return;
if (maxInputRank == 2) {
if (lhsRank == 2)
reshapedOpShape.push_back(lhsShape[0]);
if (rhsRank == 2)
reshapedOpShape.push_back(rhsShape[1]);
return;
}
// Step: Construct the output transpose/reshape information
// First the common_dims
for (uint32_t i = 0; i < batchElems.size(); i++) {
reshapedOpShape.push_back(batchElems[i].shape);
transposedOpDims.push_back(batchElems[i].dim);
}
// Then the LHS squeezed dims
for (uint32_t i = 0; i < lhsSqueezedElems.size() - 1; i++) {
// Only dims that don't broadcast - broadcasting ones come from the
// other input.
if (lhsSqueezedElems[i].shape != 1) {
reshapedOpShape.push_back(lhsSqueezedElems[i].shape);
transposedOpDims.push_back(lhsSqueezedElems[i].dim);
}
}
// The last squeezed dim is lhs[-2] which needs to be
// checked separately for broadcasting
if (lhsRank > 1) {
reshapedOpShape.push_back(lhsBroadcastedShape[maxInputRank - 2]);
transposedOpDims.push_back(maxInputRank - 2);
}
// then the RHS squeezed dims except rhs[-1] which is handled like
// lhs[-2]
for (uint32_t i = 0; i < rhsSqueezedElems.size() - 1; i++) {
if (rhsSqueezedElems[i].shape != 1) {
reshapedOpShape.push_back(rhsSqueezedElems[i].shape);
transposedOpDims.push_back(rhsSqueezedElems[i].dim);
}
}
// rhs[-1]
if (rhsRank > 1) {
reshapedOpShape.push_back(rhsBroadcastedShape[maxInputRank - 1]);
transposedOpDims.push_back(maxInputRank - 1);
}
// The transposition order is the inverse of what we actually want,
// inversing should fix this:
llvm::SmallVector<int> inverseTransposeDims(transposedOpDims.size());
for (int i = 0, s = transposedOpDims.size(); i < s; ++i)
inverseTransposeDims[transposedOpDims[i]] = i;
transposedOpDims = inverseTransposeDims;
// Final transposed output shape construction
for (uint32_t i = 0; i < maxInputRank - 2; i++) {
if (lhsBroadcastedTy.isDynamicDim(i)) {
transposedOpShapes.push_back(kUnknownSize);
} else {
if (lhsBroadcastedShape[i] == rhsBroadcastedShape[i]) {
transposedOpShapes.push_back(lhsBroadcastedShape[i]);
} else {
transposedOpShapes.push_back(lhsBroadcastedShape[i] == 1
? rhsBroadcastedShape[i]
: lhsBroadcastedShape[i]);
}
}
}
if (lhsRank > 1)
transposedOpShapes.push_back(lhsBroadcastedShape[maxInputRank - 2]);
if (rhsRank > 1)
transposedOpShapes.push_back(rhsBroadcastedShape[maxInputRank - 1]);
return;
};
SmallVector<int64_t> reshapedOpShape, transposedOpShape;
SmallVector<int32_t> transposedOpDims;
computeOpShape(reshapedOpShape, transposedOpDims, transposedOpShape);
bool opNeedsTranspose = isTransposeRequired(transposedOpDims);
// Perform reshape
auto reshapedOpType = RankedTensorType::get(
makeShapeLLVMCompatible(reshapedOpShape), outputElemTy);
auto reshapedOp = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
reshapedOpType),
mmOpResult, rewriter.getDenseI64ArrayAttr(reshapedOpShape));
if (opNeedsTranspose) {
std::optional<Value> transposedOpShapeConst =
tosa::getConstTensor<int32_t>(
rewriter, op,
/*vec=*/transposedOpDims,
/*shape=*/{static_cast<int32_t>(transposedOpDims.size())});
auto transposedOpType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedOpShape), outputElemTy);
output = rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(transposedOpType),
reshapedOp.getResult(), transposedOpShapeConst.value())
.getResult();
} else {
output = reshapedOp.getResult();
}
} else {
output = mmOpResult;
}
return success();
}
// The default version just reads two inputs, computes output and returns it.
// Other versions may add a bias, apply GEMM-style alpha/beta scaling etc.
virtual LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value lhs, rhs;
if (failed(readMatMulInputs(op, adaptor, rewriter, lhs, rhs)))
return rewriter.notifyMatchFailure(op, "Failed to read matmul inputs");
Value output;
if (failed(performMatmul(op, adaptor, rewriter, lhs, rhs, output)))
return rewriter.notifyMatchFailure(op,
"Failed to perform matmul operation");
rewriter.replaceOpWithNewOp<tensor::CastOp>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<RankedTensorType>(),
output);
return success();
}
};
// Legalizes the torch.matmul op for general n-dim matmul.
template <typename AtenOpT>
class ConvertAtenMatMulOp : public ConvertAtenMatmulBaseOp<AtenOpT> {
public:
using ConvertAtenMatmulBaseOp<AtenOpT>::ConvertAtenMatmulBaseOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
Value &lhs, Value &rhs) const override {
lhs = adaptor.getSelf();
auto lhsTy = cast<RankedTensorType>(lhs.getType());
rhs = adaptor.getOther();
auto rhsTy = cast<RankedTensorType>(rhs.getType());
if (!lhsTy || !rhsTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA matmul");
return success();
}
};
// Implements handling of aten.mm and aten.bmm ops.
template <typename AtenOpT>
class ConvertAtenMmOp : public ConvertAtenMatmulBaseOp<AtenOpT> {
public:
using ConvertAtenMatmulBaseOp<AtenOpT>::ConvertAtenMatmulBaseOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
Value &lhs, Value &rhs) const override {
lhs = adaptor.getSelf();
auto lhsTy = cast<RankedTensorType>(lhs.getType());
rhs = adaptor.getMat2();
auto rhsTy = cast<RankedTensorType>(rhs.getType());
if (!lhsTy || !rhsTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA matmul");
auto lhsRank = lhsTy.getRank();
auto rhsRank = rhsTy.getRank();
if (isa<AtenMmOp>(op)) {
// Mm takes two 2D tensors.
if (lhsRank != 2 || rhsRank != 2)
return op.emitError("aten.mm called but matrix rank != 2");
} else if (isa<AtenBmmOp>(op)) {
// Bmm takes two 3D tensors.
if (lhsRank != 3 || rhsRank != 3)
return op.emitError("aten.bmm called but matrix rank != 3");
}
return success();
}
};
// Implements handling of aten.linear op.
template <typename AtenOpT>
class ConvertAtenLinearOp : public ConvertAtenMatmulBaseOp<AtenOpT> {
public:
using ConvertAtenMatmulBaseOp<AtenOpT>::ConvertAtenMatmulBaseOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult readMatMulInputs(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
Value &lhs, Value &rhs) const override {
lhs = adaptor.getInput();
auto lhsTy = cast<RankedTensorType>(lhs.getType());
rhs = adaptor.getWeight();
auto rhsTy = cast<RankedTensorType>(rhs.getType());
if (!lhsTy || !rhsTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA matmul");
auto lhsRank = lhsTy.getRank();
auto rhsRank = rhsTy.getRank();
if (lhsRank != 2 && lhsRank != 3)
return op.emitError("aten.Linear called but input rank not 2 or 3");
if (rhsRank != 2 && rhsRank != 3)
return op.emitError("aten.Linear called but weight rank not 2 or 3");
// Protection against crash due to unguarded code in TOSA->LinAlg.
// TODO: This should be handled in TOSA->LinAlg instead.
if (!lhsTy.hasStaticShape() || !rhsTy.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "aten.Linear needs statically shaped input");
return success();
}
// Override the default rewriter to perform RHS transpose and bias addition as
// well.
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value lhs, rhs;
if (failed(readMatMulInputs(op, adaptor, rewriter, lhs, rhs)))
return rewriter.notifyMatchFailure(op, "Failed to read matmul inputs");
// The aten.Linear op has a bias tensor that is added to the matmul output.
auto bias = adaptor.getBias();
auto biasTy = bias.getType();
// TOSA does not mandate that elementwise op tensors need to be ranked.
if (!isa<Torch::NoneType>(biasTy) && !isa<TensorType>(biasTy))
return rewriter.notifyMatchFailure(
op, "Only tensor types supported in GEMM to TOSA for bias tensor");
// RHS must have its last two dims transposed prior to matrix
// multiplication.
auto rhsTy = cast<RankedTensorType>(rhs.getType());
auto rhsRank = rhsTy.getRank();
auto rhsShape = makeShapeTorchCompatible(rhsTy.getShape());
auto rhsElemTy = rhsTy.getElementType();
// Create a non-const shape array to transpose dims.
SmallVector<int64_t> transposedRhsShape;
for (auto &shape : rhsShape)
transposedRhsShape.push_back(shape);
SmallVector<int32_t> transposedRhsDims;
for (int32_t i = 0; i < rhsRank; i++)
transposedRhsDims.push_back(i);
// Swap the last two dims.
std::swap(transposedRhsShape[rhsRank - 1], transposedRhsShape[rhsRank - 2]);
std::swap(transposedRhsDims[rhsRank - 1], transposedRhsDims[rhsRank - 2]);
std::optional<Value> transposedRhsShapeConst =
tosa::getConstTensor<int32_t>(
rewriter, op,
/*vec=*/transposedRhsDims,
/*shape=*/{static_cast<int32_t>(transposedRhsDims.size())});
auto transposedRhsType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedRhsShape), rhsElemTy);
rhs = rewriter.create<tosa::TransposeOp>(
op->getLoc(),
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
transposedRhsType),
rhs, transposedRhsShapeConst.value());
Value matmulOutput;
if (failed(
this->performMatmul(op, adaptor, rewriter, lhs, rhs, matmulOutput)))
return rewriter.notifyMatchFailure(op,
"Failed to perform matmul operation");
Value matmulPlusBias = matmulOutput;
if (!isa<Torch::NoneType>(biasTy)) {
// Bias addition broadcasts to the matmul output shape.
matmulPlusBias =
rewriter
.create<tosa::AddOp>(op->getLoc(), matmulOutput.getType(),
matmulOutput, bias)
.getResult();
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template cast<RankedTensorType>(),
matmulPlusBias);
return success();
}
};
template <>
LogicalResult ConvertAtenOp<AtenRsubScalarOp>::matchAndRewrite(
AtenRsubScalarOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto self = adaptor.getSelf();
auto otherScalar = op.getOther();
auto alphaScalar = op.getAlpha();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA Rsub");
if (!selfTy.getElementType().isa<mlir::FloatType>())
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization supported");
Value otherTensor, alphaTensor;
if (failed(torchScalarToTosaTensor(rewriter, op, otherScalar, otherTensor,
selfTy.getElementType(), {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA Rsub operation");
if (failed(torchAlphaToTosaTensor(rewriter, op.getOperation(), alphaScalar,
alphaTensor, selfTy.getElementType(),
/*checkForUnity=*/true)))
return failure();
auto multTensor = rewriter.create<tosa::MulOp>(
op->getLoc(), getTypeConverter()->convertType(op.getType()), self,
alphaTensor, /*shift=*/0);
rewriter.replaceOpWithNewOp<tosa::SubOp>(
op, getTypeConverter()->convertType(op.getType()), otherTensor,
multTensor);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenConvolutionOp>::matchAndRewrite(
AtenConvolutionOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
bool transposed;
if (!matchPattern(op.getTransposed(), m_TorchConstantBool(&transposed)))
return rewriter.notifyMatchFailure(
op, "Unimplemented: non-constant value for transposed not supported");
if (transposed)
return rewriter.notifyMatchFailure(
op, "Unimplemented: transposed convolution not supported");
auto input = adaptor.getInput();
auto weight = adaptor.getWeight();
auto inputTy = cast<RankedTensorType>(input.getType());
auto weightTy = cast<RankedTensorType>(weight.getType());
auto outputTy = getTypeConverter()
->convertType(op.getType())
.template cast<RankedTensorType>();
if (!inputTy || !weightTy || !outputTy)
return rewriter.notifyMatchFailure(
op, "Input, weight and output to Convolution must be ranked tensors");
auto inputElemTy = inputTy.getElementType();
auto weightElemTy = weightTy.getElementType();
auto inputShape = makeShapeTorchCompatible(inputTy.getShape());
auto weightShape = makeShapeTorchCompatible(weightTy.getShape());
if (inputTy.getRank() != 4)
return rewriter.notifyMatchFailure(
op, "Unimplemented: only 2D convolutions supported");
if (!weightTy.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Unimplemented: TOSA only supports static weight");
// Bias is optional. TOSA mandates a zero tensor here, so construct one if
// required.
auto bias = adaptor.getBias();
if (isa<Torch::NoneType>(adaptor.getBias().getType())) {
// TBD: This is only valid for quantized 8-bit. For 16-bit, the bias (and
// accumulator) are 48-bit and not 32-bit, and requires the use of APInt to
// define a 48-bit int.
if (isa<quant::QuantizedType>(inputElemTy)) {
SmallVector<int32_t> zeroVec(weightShape[0], 0);
bias = tosa::getConstTensor<int32_t>(
rewriter, op, zeroVec, {static_cast<int32_t>(weightShape[0])})
.value();
} else {
SmallVector<float> zeroVec(weightShape[0], 0);
bias = tosa::getConstTensor<float>(rewriter, op, zeroVec,
{static_cast<int32_t>(weightShape[0])})
.value();
}
} else {
if (!cast<RankedTensorType>(bias.getType()))
return rewriter.notifyMatchFailure(
op, "Bias provided but not a ranked tensor");
}
auto biasElemTy =
isa<mlir::FloatType>(inputElemTy) ? inputElemTy : rewriter.getI32Type();
int64_t groups;
if (!matchPattern(op.getGroups(), m_TorchConstantInt(&groups))) {
return rewriter.notifyMatchFailure(op, "non-const group size unsupported");
} else if (groups != 1 && weightShape[1] != 1) {
return rewriter.notifyMatchFailure(
op, "group size must be 1 (convolution) or weight.dim(1) must be 1 "
"(depthwise convolution)");
}
SmallVector<int64_t, 2> stride;
if (!matchPattern(adaptor.getStride(), m_TorchListOfConstantInts(stride)))
return rewriter.notifyMatchFailure(op, "non-const stride list unsupported");
SmallVector<int64_t, 2> padding_2d;
if (!matchPattern(adaptor.getPadding(),
m_TorchListOfConstantInts(padding_2d)))
return rewriter.notifyMatchFailure(op,
"non-const padding list unsupported");
// TOSA uses 4D padding {t, b, l, r} while Torch defines 2D padding {t, l}.
// The Torch OFM computation uses 2*pad in each spatial direction, implying
// the same t=b and l=r values for TOSA.
SmallVector<int64_t> padding(
{padding_2d[0], padding_2d[0], padding_2d[1], padding_2d[1]});
SmallVector<int64_t, 2> dilation;
if (!matchPattern(adaptor.getDilation(), m_TorchListOfConstantInts(dilation)))
return rewriter.notifyMatchFailure(op,
"non-const dilation list unsupported");
// TOSA works in NHWC and takes OHWI (conv) / HWIM (depthwise conv) weights.
// Perform the necessary transformations.
std::optional<Value> nchwToNhwcTransposeConst =
tosa::getConstTensor<int32_t>(rewriter, op,
/*vec=*/{0, 2, 3, 1},
/*shape=*/{static_cast<int32_t>(4)});
SmallVector<int64_t> transposedInputShape(
{inputShape[0], inputShape[2], inputShape[3], inputShape[1]});
auto transposedInputType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedInputShape), inputElemTy);
auto transposedInput =
rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
getTypeConverter()->convertType(transposedInputType), input,
nchwToNhwcTransposeConst.value())
.getResult();
SmallVector<int64_t> transformedWeightShape;
RankedTensorType transformedWeightType;
Value transformedWeight;
int64_t outputCDim;
if (groups == 1) {
// full convolution: O(I/G)HW-> OHWI
transformedWeightShape = {weightShape[0], weightShape[2], weightShape[3],
weightShape[1]};
transformedWeightType = RankedTensorType::get(
makeShapeLLVMCompatible(transformedWeightShape), weightElemTy);
transformedWeight =
rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
getTypeConverter()->convertType(transformedWeightType), weight,
nchwToNhwcTransposeConst.value())
.getResult();
outputCDim = transformedWeightShape[0];
} else if (weightShape[1] == 1) {
// depthwise convolution: O(I/G)HW-> HWIM)
// transpose: O(I/G)HW -> HWO(I/G)
std::optional<Value> transposeConst =
tosa::getConstTensor<int32_t>(rewriter, op,
/*vec=*/{2, 3, 0, 1},
/*shape=*/{static_cast<int32_t>(4)});
SmallVector<int64_t> transposedWeightShape = {
weightShape[2], weightShape[3], weightShape[0], weightShape[1]};
auto transposedWeightType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedWeightShape), weightElemTy);
auto transposedWeight =
rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
getTypeConverter()->convertType(transposedWeightType), weight,
transposeConst.value())
.getResult();
// reshape: HWO(I/G) -> HWIM
outputCDim = makeShapeTorchCompatible(outputTy.getShape())[1];
if (outputCDim == kUnknownSize) {
return rewriter.notifyMatchFailure(
op, "number of output channels must be statically known for "
"depthwise convolutions");
}
transformedWeightShape = {
transposedWeightShape[0],
transposedWeightShape[1],
groups,
outputCDim / groups,
};
transformedWeightType = RankedTensorType::get(
makeShapeLLVMCompatible(transformedWeightShape), weightElemTy);
transformedWeight =
rewriter
.create<tosa::ReshapeOp>(
op->getLoc(),
getTypeConverter()->convertType(transformedWeightType),
transposedWeight,
rewriter.getDenseI64ArrayAttr(transformedWeightShape))
.getResult();
} else {
llvm_unreachable("Unhandled convolution type");
}
int64_t outputHDim, outputWDim;
if (inputTy.hasStaticShape()) {
int64_t inputHDim = inputShape[2];
int64_t inputWDim = inputShape[3];
int64_t weightHDim = weightShape[2];
int64_t weightWDim = weightShape[3];
outputHDim = (inputHDim + padding[0] + padding[1] -
dilation[0] * (weightHDim - 1) - 1) /
stride[0] +
1;
outputWDim = (inputWDim + padding[2] + padding[3] -
dilation[1] * (weightWDim - 1) - 1) /
stride[1] +
1;
} else {
outputHDim = kUnknownSize;
outputWDim = kUnknownSize;
}
// Output shape is NHWC, to be transposed back to NCHW. Output elemTy for
// quantized input is i32, which gets rescaled down to quantized output range.
SmallVector<int64_t> outputShape = {transposedInputShape[0], outputHDim,
outputWDim, outputCDim};
auto convOpTy =
RankedTensorType::get(makeShapeLLVMCompatible(outputShape), biasElemTy);
Value convOpResult;
if (groups == 1) {
// full convolution
convOpResult =
rewriter
.create<tosa::Conv2DOp>(op->getLoc(),
getTypeConverter()->convertType(convOpTy),
transposedInput, transformedWeight, bias,
rewriter.getDenseI64ArrayAttr(padding),
rewriter.getDenseI64ArrayAttr(stride),
rewriter.getDenseI64ArrayAttr(dilation))
.getResult();
} else if (weightShape[1] == 1) {
// depthwise convolution
convOpResult =
rewriter
.create<tosa::DepthwiseConv2DOp>(
op->getLoc(), getTypeConverter()->convertType(convOpTy),
transposedInput, transformedWeight, bias,
rewriter.getDenseI64ArrayAttr(padding),
rewriter.getDenseI64ArrayAttr(stride),
rewriter.getDenseI64ArrayAttr(dilation))
.getResult();
} else {
llvm_unreachable("Unhandled convolution type");
}
std::optional<Value> nhwcToNchwTransposeConst =
tosa::getConstTensor<int32_t>(rewriter, op,
/*vec=*/{0, 3, 1, 2},
/*shape=*/{static_cast<int32_t>(4)});
SmallVector<int64_t> transposedOutputShape(
{outputShape[0], outputShape[3], outputShape[1], outputShape[2]});
auto transposedOutputType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedOutputShape), biasElemTy);
auto transposedOutput =
rewriter
.create<tosa::TransposeOp>(
op->getLoc(),
getTypeConverter()->convertType(transposedOutputType),
convOpResult, nhwcToNchwTransposeConst.value())
.getResult();
Value rescaledResult = transposedOutput;
if (isa<quant::QuantizedType>(inputElemTy)) {
rescaledResult = tosa::buildRescaleOpConvOutput(
rewriter, op, transposedOutput, inputTy, weightTy, outputTy);
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(
op, getTypeConverter()->convertType(op.getType()), rescaledResult);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenReshapeOp>::matchAndRewrite(
AtenReshapeOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA Reshape");
// Check that at most one dimension is -1
SmallVector<int64_t> newShape;
if (!matchPattern(op.getShape(), m_TorchListOfConstantInts(newShape)))
return rewriter.notifyMatchFailure(
op, "Only constant shape supported in TOSA Reshape");
int auto_sz = 0;
for (auto s : newShape)
auto_sz += (s == -1 ? 1 : 0);
if (auto_sz > 1)
return rewriter.notifyMatchFailure(
op, "At most one dimension may be specified as -1 to "
"automatically calculate its size");
auto newType = RankedTensorType::get(makeShapeLLVMCompatible(newShape),
selfTy.getElementType());
rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
op, getTypeConverter()->convertType(newType), self,
rewriter.getDenseI64ArrayAttr(newShape));
return success();
}
Value computeBatchNorm(Operation *op, ConversionPatternRewriter &rewriter,
Type outType, Value input, Value variance, Value eps,
Value mean, Value weight, Value bias) {
// For PyTorch:
// scale = gamma = weight
// offset = beta = bias
// Lowering:
// fused batchnorm = (input-mean) * scale * rsqrt(var+epsilon)) + offset
//
// shape_0 = ones(input.rank)
// shape_0[input.rank-1] = input.shape[input.rank-1]
// shape_1 = ones(1)
//
// bmean = reshape(mean, shape_0)
// bscale = reshape(scale, shape_0)
// boffset= reshape(offset, shape_0)
// beps = reshape(epsilon, shape_1)
//
// op1 = sub(input, bmean)
// op2 = add(var, beps)
// op3 = rsqrt(op2)
// bvar = reshape(op3, shape_0)
// op4 = mul(op1, bvar)
// op5 = mul(op4, bscale)
// op6 = add(op5, boffset)
auto op1SubInputMean =
rewriter.create<tosa::SubOp>(op->getLoc(), outType, input, mean);
auto op2AddVarEpsilon = rewriter.create<tosa::AddOp>(
op->getLoc(), variance.getType(), variance, eps);
auto op3RsqrtOp2 = rewriter.create<tosa::RsqrtOp>(
op->getLoc(), variance.getType(), op2AddVarEpsilon.getResult());
auto op4MulOp1Op3 = rewriter.create<tosa::MulOp>(op->getLoc(), outType,
op1SubInputMean.getResult(),
op3RsqrtOp2.getResult(), 0);
auto op5MulOp4Scale = rewriter.create<tosa::MulOp>(
op->getLoc(), outType, op4MulOp1Op3.getResult(), weight, 0);
return rewriter
.create<tosa::AddOp>(op->getLoc(), outType, op5MulOp4Scale.getResult(),
bias)
.getResult();
}
// This lowering is based on the TensorFlow to TOSA lowering.
template <>
LogicalResult ConvertAtenOp<AtenBatchNormOp>::matchAndRewrite(
AtenBatchNormOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a ranked tensor output
if (!dyn_cast<RankedTensorType>(adaptor.getInput().getType()))
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types are supported");
auto outType = getTypeConverter()->convertType(op.getType());
// Note: cudnn_enabled is not handled.
// FIXME: Handle training and momentum.
if (op.getMomentum().getType().isa<Torch::NoneType>())
return rewriter.notifyMatchFailure(op, "Unsupported None for momentum");
auto meanType = dyn_cast<TensorType>(adaptor.getRunningMean().getType());
auto varianceType = dyn_cast<TensorType>(adaptor.getRunningVar().getType());
if (!varianceType || !meanType)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types are supported");
// Normalization ops perform elementwise ops of a single mean/stdev value
// against the feature map and because input is NCHW, the rank-1 value must be
// reshaped so it sits on the same dim as 'C'.
auto reshapeToNormInputDim = [&](Operation *op,
ConversionPatternRewriter &rewriter,
const TypeConverter *converter, Type outType,
const Value toBcast, Value &result) {
RankedTensorType toBcastType =
dyn_cast<RankedTensorType>(toBcast.getType());
if (toBcastType.getRank() > 1)
return rewriter.notifyMatchFailure(op, "Rank cannot be more than 1");
RankedTensorType outTensorType = cast<RankedTensorType>(outType);
SmallVector<int64_t> newShape = {
makeShapeTorchCompatible(toBcastType.getShape())[0]};
for (auto i = 2; i < outTensorType.getRank(); ++i)
newShape.push_back(1);
auto newType = RankedTensorType::get(makeShapeLLVMCompatible(newShape),
outTensorType.getElementType());
result = rewriter.create<tosa::ReshapeOp>(
op->getLoc(), newType, toBcast,
rewriter.getDenseI64ArrayAttr(newShape));
return success();
};
Value meanVal, varianceVal, weightVal, biasVal;
assert(meanType.getNumElements() != 0 && varianceType.getNumElements() != 0);
if (failed(reshapeToNormInputDim(op.getOperation(), rewriter,
getTypeConverter(), outType,
adaptor.getRunningMean(), meanVal)))
return rewriter.notifyMatchFailure(op, "Failed to reshape running mean");
if (failed(reshapeToNormInputDim(op.getOperation(), rewriter,
getTypeConverter(), outType,
adaptor.getRunningVar(), varianceVal)))
return rewriter.notifyMatchFailure(op,
"Failed to reshape running variance");
if (failed(reshapeToNormInputDim(op.getOperation(), rewriter,
getTypeConverter(), outType,
adaptor.getWeight(), weightVal)))
return rewriter.notifyMatchFailure(op, "Failed to reshape weight");
if (failed(reshapeToNormInputDim(op.getOperation(), rewriter,
getTypeConverter(), outType,
adaptor.getBias(), biasVal)))
return rewriter.notifyMatchFailure(op, "Failed to reshape bias");
double eps;
if (!matchPattern(op.getEps(), m_TorchConstantFloat(&eps)))
return rewriter.notifyMatchFailure(op, "eps must be a scalar constant");
auto epsilonConst = tosa::getConstTensor<float>(rewriter, op.getOperation(),
{static_cast<float>(eps)}, {},
meanType.getElementType())
.value();
auto batchNorm =
computeBatchNorm(op, rewriter, outType, adaptor.getInput(), varianceVal,
epsilonConst, meanVal, weightVal, biasVal);
rewriter.replaceOp(op, {batchNorm});
return success();
}
// This lowering is loosely based on Torch to LinAlg lowering.
template <>
LogicalResult ConvertAtenOp<AtenNativeLayerNormOp>::matchAndRewrite(
AtenNativeLayerNormOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// The key difference from BatchNorm is that a specified set of dims
// (normalized_shape) are chosen to compute the mean and variance from input.
// Where as in BatchNorm the mean and variance are operands. tosa::ReduceSumOp
// is used to sum up the these dims for mean and for variance. The results
// eventually being reshaped for broadcasting.
// Not a ranked tensor output
if (!dyn_cast<RankedTensorType>(adaptor.getInput().getType()))
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types are supported");
auto inputType = cast<RankedTensorType>(adaptor.getInput().getType());
if (inputType.getRank() > 4)
return rewriter.notifyMatchFailure(op,
"Only up to 4D tensors are supported");
auto outType = getTypeConverter()->convertType(op.getType(0));
// Note: cudnn_enabled is not handled.
// FIXME: Handle the None cases for the optional parameters.
if (adaptor.getWeight().getType().isa<Torch::NoneType>())
return rewriter.notifyMatchFailure(op, "Unsupported None for weight");
if (adaptor.getBias().getType().isa<Torch::NoneType>())
return rewriter.notifyMatchFailure(op, "Unsupported None for bias");
auto weightType = cast<RankedTensorType>(adaptor.getWeight().getType());
auto biasType = cast<RankedTensorType>(adaptor.getBias().getType());
int64_t inputRank = inputType.getRank();
Type elemTy = inputType.getElementType();
SmallVector<int64_t> inputTypeShape(
makeShapeTorchCompatible(inputType.getShape()));
// Check if all the arguments meet the requirements.
SmallVector<int64_t> normalizedShapeSizesInt;
if (!matchPattern(op.getNormalizedShape(),
m_TorchListOfConstantInts(normalizedShapeSizesInt))) {
return rewriter.notifyMatchFailure(op, "Unimplemented normalized_shape not"
"constructed from ListConstruct");
}
int64_t normalizedShapeRank = normalizedShapeSizesInt.size();
if (weightType.getRank() != normalizedShapeRank ||
biasType.getRank() != normalizedShapeRank ||
inputRank < normalizedShapeRank || normalizedShapeRank < 1)
return rewriter.notifyMatchFailure(op, "Input or weight or bias shape or"
"normalized shape not compatible");
// Check all the dimensions match the normalized_shape, only static shapes as
// of now
int64_t meanAndVarShapeRank = inputRank - normalizedShapeSizesInt.size();
for (auto en : llvm::enumerate((normalizedShapeSizesInt))) {
int64_t index = en.index();
int64_t value = en.value();
if (inputTypeShape[index + meanAndVarShapeRank] != value ||
makeShapeTorchCompatible(weightType.getShape())[index] != value ||
makeShapeTorchCompatible(biasType.getShape())[index] != value)
return rewriter.notifyMatchFailure(op,
"mismatching contracting dimension");
}
// Helper for computing mean and variance.
auto computeSumAndReshape = [&](Value toReduce, RankedTensorType toReduceType,
Type outType, SmallVector<int64_t> outShape) {
Value sumDiv = toReduce;
SmallVector<int64_t> toReduceShape(
makeShapeTorchCompatible(toReduceType.getShape()));
for (int64_t i = toReduceShape.size() - 1; i >= meanAndVarShapeRank; i--) {
toReduceShape[i] = 1;
sumDiv = rewriter.create<tosa::ReduceSumOp>(
op.getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(toReduceShape),
inputType.getElementType()),
sumDiv, rewriter.getI32IntegerAttr(i));
}
return rewriter.create<tosa::ReshapeOp>(
op.getLoc(), outType, sumDiv, rewriter.getDenseI64ArrayAttr(outShape));
};
// TOSA has integer Div so, compute reciprocal of element count to be used in
// mul.
int64_t elemCnt = 1;
for (auto i : normalizedShapeSizesInt)
elemCnt *= i;
auto elemCntConst =
tosa::getConstTensor<float>(rewriter, op.getOperation(),
{static_cast<float>(elemCnt)}, {1}, elemTy)
.value();
Value elemCntRcp = rewriter.create<tosa::ReciprocalOp>(
op.getLoc(), elemCntConst.getType(), elemCntConst);
// Broadcast type and shape for various intermediate values.
SmallVector<int64_t> bcastOutShape;
for (auto en : llvm::enumerate(inputTypeShape)) {
bcastOutShape.push_back(
static_cast<int64_t>(en.index()) >= meanAndVarShapeRank ? 1
: en.value());
}
auto bcastOutType =
RankedTensorType::get(makeShapeLLVMCompatible(bcastOutShape), elemTy);
// Compute mean.
Value sum = computeSumAndReshape(adaptor.getInput(), inputType, bcastOutType,
bcastOutShape);
Value meanVal = rewriter.create<tosa::MulOp>(op.getLoc(), bcastOutType, sum,
elemCntRcp, /*shift=*/0);
// Compute variance.
Value squareSumSub = rewriter.create<tosa::SubOp>(
op.getLoc(), inputType, adaptor.getInput(), meanVal);
Value squareSum = rewriter.create<tosa::MulOp>(op.getLoc(), inputType,
squareSumSub, squareSumSub, 0);
Value squareSumReduced =
computeSumAndReshape(squareSum, inputType, bcastOutType, bcastOutShape);
Value varianceVal = rewriter.create<tosa::MulOp>(
op.getLoc(), bcastOutType, squareSumReduced, elemCntRcp, /*shift=*/0);
// Reshape weight and bias.
SmallVector<int64_t> weightAndBiasBcastShape;
for (auto en :
llvm::enumerate(makeShapeTorchCompatible(inputType.getShape()))) {
weightAndBiasBcastShape.push_back(
static_cast<int64_t>(en.index()) < meanAndVarShapeRank ? 1
: en.value());
}
auto weightAndMeanBcastType = RankedTensorType::get(
makeShapeLLVMCompatible(weightAndBiasBcastShape), elemTy);
Value weightVal = rewriter.create<tosa::ReshapeOp>(
op.getLoc(), weightAndMeanBcastType, adaptor.getWeight(),
rewriter.getDenseI64ArrayAttr(weightAndBiasBcastShape));
Value biasVal = rewriter.create<tosa::ReshapeOp>(
op.getLoc(), weightAndMeanBcastType, adaptor.getBias(),
rewriter.getDenseI64ArrayAttr(weightAndBiasBcastShape));
double eps;
if (!matchPattern(op.getEps(), m_TorchConstantFloat(&eps)))
return rewriter.notifyMatchFailure(op, "eps must be a scalar constant");
auto epsilonConst =
tosa::getConstTensor<float>(rewriter, op.getOperation(),
{static_cast<float>(eps)}, {}, elemTy)
.value();
// Compute layer norm.
auto layerNorm =
computeBatchNorm(op, rewriter, outType, adaptor.getInput(), varianceVal,
epsilonConst, meanVal, weightVal, biasVal);
rewriter.replaceOp(op, {layerNorm, meanVal, varianceVal});
return success();
}
// Torch constants are converted to tosa.const .
template <>
LogicalResult ConvertAtenOp<ValueTensorLiteralOp>::matchAndRewrite(
ValueTensorLiteralOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto outputTy = getTypeConverter()
->convertType(op.getType())
.template 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 = dyn_cast<DenseIntElementsAttr>(op.getValueAttr())) {
if (elements.getElementType().isSignedInteger()) {
Type builtinTensorElemTy = outputTy.getElementType();
unsigned bitWidth = builtinTensorElemTy.getIntOrFloatBitWidth();
DenseElementsAttr valueAttr =
elements.mapValues(builtinTensorElemTy, [&](const APInt &v) {
return APInt(bitWidth, v.getSExtValue());
});
rewriter.replaceOpWithNewOp<tosa::ConstOp>(op, outputTy, valueAttr);
return success();
}
}
rewriter.replaceOpWithNewOp<tosa::ConstOp>(op, outputTy, adaptor.getValue());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenFlattenUsingIntsOp>::matchAndRewrite(
AtenFlattenUsingIntsOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a ranked tensor type
auto selfType = dyn_cast<RankedTensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(op,
"Only ranked tensor types supported");
int64_t selfRank = selfType.getRank();
int64_t start_dim, end_dim;
if (!matchPattern(op.getStartDim(), m_TorchConstantInt(&start_dim)))
return rewriter.notifyMatchFailure(op,
"start_dim must be a Scalar constant");
start_dim = toPositiveDim(start_dim, selfRank);
if (!matchPattern(op.getEndDim(), m_TorchConstantInt(&end_dim)))
return rewriter.notifyMatchFailure(op, "end_dim must be a Scalar constant");
end_dim = toPositiveDim(end_dim, selfRank);
if (selfRank > 0 && !isValidDim(start_dim, selfRank))
return rewriter.notifyMatchFailure(op, "start_dim is statically invalid");
if (selfRank > 0 && !isValidDim(end_dim, selfRank))
return rewriter.notifyMatchFailure(op, "end_dim is statically invalid");
if (end_dim < start_dim)
return rewriter.notifyMatchFailure(op,
"end_dim must be larger than start_dim");
SmallVector<int64_t> newShape;
for (auto s :
llvm::enumerate(makeShapeTorchCompatible(selfType.getShape()))) {
int64_t idx = s.index();
if (idx < start_dim || idx > end_dim) {
newShape.push_back(s.value());
} else {
if (idx == start_dim)
newShape.push_back(s.value());
// Only updating when the shapes are static
else if (s.value() != kUnknownSize && newShape.back() != kUnknownSize)
newShape.back() *= s.value();
else
newShape.back() = kUnknownSize;
}
}
// Handle the Scalar case
if (newShape.size() == 0)
newShape.push_back(1);
auto newType = RankedTensorType::get(makeShapeLLVMCompatible(newShape),
selfType.getElementType());
auto reshapeOp =
rewriter.create<tosa::ReshapeOp>(op.getLoc(), newType, adaptor.getSelf(),
rewriter.getDenseI64ArrayAttr(newShape));
rewriter.replaceOpWithNewOp<tensor::CastOp>(
op, getTypeConverter()->convertType(op.getType()), reshapeOp);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenUnflattenIntOp>::matchAndRewrite(
AtenUnflattenIntOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a ranked tensor type
auto selfType = dyn_cast<RankedTensorType>(adaptor.getSelf().getType());
if (!selfType || !selfType.hasStaticShape())
return rewriter.notifyMatchFailure(
op,
"Only ranked tensor types with static shapes are currently supported");
int64_t selfRank = selfType.getRank();
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(op, "dim must be a Scalar constant");
SmallVector<int64_t> sizes;
if (!matchPattern(op.getSizes(), m_TorchListOfConstantInts(sizes)))
return rewriter.notifyMatchFailure(
op, "Only constant sizes are currently supported");
if (selfRank > 0 && !isValidDim(dim, selfRank))
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
SmallVector<int64_t> newShape;
for (auto s :
llvm::enumerate(makeShapeTorchCompatible(selfType.getShape()))) {
int64_t idx = s.index();
if (idx < dim || idx > dim) {
newShape.push_back(s.value());
} else {
auto sum = 1;
for (auto newDims : sizes) {
newShape.push_back(newDims);
sum *= newDims;
}
if (sum != s.value())
return rewriter.notifyMatchFailure(op,
"sizes mismatch with original dim");
}
}
auto newType = RankedTensorType::get(makeShapeLLVMCompatible(newShape),
selfType.getElementType());
rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
op, getTypeConverter()->convertType(newType), adaptor.getSelf(),
rewriter.getDenseI64ArrayAttr(newShape));
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenPermuteOp>::matchAndRewrite(
AtenPermuteOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a ranked tensor type
auto selfType = dyn_cast<RankedTensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op,
"Only ranked tensor types with static shapes are currently supported");
SmallVector<int64_t> dimListInt;
if (!matchPattern(adaptor.getDims(), m_TorchListOfConstantInts(dimListInt)))
return rewriter.notifyMatchFailure(
op, "Only constant dimensions are currently supported");
int64_t selfRank = selfType.getRank();
// TODO: If this is already verified on the op then we can drop checking here.
for (auto &d : dimListInt) {
d = toPositiveDim(d, selfRank);
if (!isValidDim(d, selfRank))
return rewriter.notifyMatchFailure(op, "Not all dims are valid");
}
auto transposeDimsConst = mlir::tosa::getConstTensor<int64_t>(
rewriter, op.getOperation(), dimListInt, {selfRank});
rewriter.replaceOpWithNewOp<tosa::TransposeOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getSelf(),
transposeDimsConst.value());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenLog2Op>::matchAndRewrite(
AtenLog2Op op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
// Constant value of ln2.
SmallVector<int64_t> ln2Shape(selfType.getRank(), 1);
auto ln2Op = tosa::getConstTensor<float>(rewriter, op, {0.69314718056f},
ln2Shape, selfType.getElementType())
.value();
auto rcpOp =
rewriter.create<tosa::ReciprocalOp>(op.getLoc(), ln2Op.getType(), ln2Op);
auto outType = getTypeConverter()->convertType(op.getType());
auto logOp =
rewriter.create<tosa::LogOp>(op.getLoc(), outType, adaptor.getSelf());
rewriter.replaceOpWithNewOp<tosa::MulOp>(op, outType, logOp, rcpOp,
/*shift=*/0);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenThresholdOp>::matchAndRewrite(
AtenThresholdOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
auto selfElemTy = selfType.getElementType();
if (!selfElemTy.isIntOrFloat())
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
// Integer types with width > 32 are not supported
auto selfIntType = dyn_cast<IntegerType>(selfElemTy);
if (selfIntType && selfIntType.getWidth() > 32) {
return rewriter.notifyMatchFailure(
op, "Integer types with width greater than 32 are not supported");
}
SmallVector<int64_t> constTypeShape(selfType.getRank(), 1);
Value threshold, value;
if (failed(torchScalarToTosaTensor(rewriter, op, op.getThreshold(), threshold,
selfElemTy, constTypeShape)))
return rewriter.notifyMatchFailure(
op, "Only scalar constant is supported for threshold");
if (failed(torchScalarToTosaTensor(rewriter, op, op.getValue(), value,
selfElemTy, constTypeShape)))
return rewriter.notifyMatchFailure(
op, "Only scalar constant is supported for value");
// Threshold only clamps the upper values. tosa::ClampOp has the same
// value for both threshold and clamped value so cannot be used.
auto outType = getTypeConverter()->convertType(op.getType());
auto cmpOp = rewriter.create<tosa::GreaterOp>(
op.getLoc(),
RankedTensorType::get(selfType.getShape(), rewriter.getIntegerType(1)),
adaptor.getSelf(), threshold);
rewriter.replaceOpWithNewOp<tosa::SelectOp>(op, outType, cmpOp,
adaptor.getSelf(), value);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenUnsqueezeOp>::matchAndRewrite(
AtenUnsqueezeOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType) {
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
}
auto selfRank = selfType.getRank();
auto selfElemTy = selfType.getElementType();
if (!selfElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(op, "dim must be a Scalar constant");
// toPositiveDim converts negative dims to the range [0, inputRank). So, -1
// will be converted to inputRank-1. For `torch.unsqueeze` op, -1 has to be
// converted to inputRank, and the valid dim range is [0, inputRank + 1).
dim = toPositiveDim(dim, selfRank + 1);
if (!isValidDim(dim, selfRank + 1))
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
SmallVector<int64_t> outShape;
for (auto en :
llvm::enumerate(makeShapeTorchCompatible(selfType.getShape()))) {
if (static_cast<int64_t>(en.index()) == dim)
outShape.push_back(1);
outShape.push_back(en.value());
}
if (dim == selfRank)
outShape.push_back(1);
rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getSelf(),
rewriter.getDenseI64ArrayAttr(outShape));
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenContiguousOp>::matchAndRewrite(
AtenContiguousOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
// FIXME: memory_format is not handled.
rewriter.replaceOp(op, adaptor.getSelf());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenDropoutOp>::matchAndRewrite(
AtenDropoutOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getInput().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
// FIXME: train and p are not handled.
bool train;
if (!matchPattern(op.getTrain(), m_TorchConstantBool(&train)))
return rewriter.notifyMatchFailure(op, "train must be a Scalar constant");
if (train)
return rewriter.notifyMatchFailure(op, "train must be false");
rewriter.replaceOpWithNewOp<tosa::CastOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getInput());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenViewOp>::matchAndRewrite(
AtenViewOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
auto selfElemTy = selfType.getElementType();
if (!selfElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
SmallVector<int64_t> outShape;
if (!matchPattern(op.getSize(), m_TorchListOfConstantInts(outShape)))
return rewriter.notifyMatchFailure(op,
"size must consist of Scalar constants");
// the shape -1 is inferred from other dimensions
size_t countNegativeShape{0};
// Check at most one -1 shape
for (size_t i = 0; i < outShape.size(); i++) {
if (outShape[i] < 0) {
countNegativeShape++;
if (countNegativeShape > 1)
return rewriter.notifyMatchFailure(op, "At most one -1 shape");
}
}
auto inputShape = selfType.getShape();
size_t totalSize = 1;
for (size_t i = 0; i < inputShape.size(); i++) {
totalSize *= inputShape[i];
}
size_t otherSize = 1;
for (size_t i = 0; i < outShape.size(); i++) {
if (outShape[i] > 0) {
otherSize *= outShape[i];
}
}
for (size_t i = 0; i < outShape.size(); i++) {
if (outShape[i] < 0) {
outShape[i] = totalSize / otherSize;
break;
}
}
rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getSelf(),
rewriter.getDenseI64ArrayAttr(outShape));
return success();
}
static Value approximateErfOp(ConversionPatternRewriter &rewriter,
Operation *op, Value x, Type dtype) {
// Using:
// https://en.wikipedia.org/wiki/Error_function#Numerical_approximations with
// maximum error as 5 x 10^-4 where a1 = 0.278393, a2 = 0.230389, a3 =
// 0.000972, a4 = 0.078108.
//
// Erf = 1 - 1 / (1 + a1X + a2X + a3X + a4X)^4
auto outType = cast<TensorType>(x.getType());
auto loc = op->getLoc();
auto absX = rewriter.create<tosa::AbsOp>(loc, outType, x);
auto zero = tosa::getConstTensor<float>(rewriter, op, 0, {}, dtype).value();
auto one = tosa::getConstTensor<float>(rewriter, op, 1, {}, dtype).value();
auto a1 =
tosa::getConstTensor<float>(rewriter, op, 0.278393f, {}, dtype).value();
auto a1X = rewriter.create<tosa::MulOp>(loc, outType, a1, absX, /*shift=*/0);
auto sum = rewriter.create<tosa::AddOp>(loc, outType, a1X, one);
auto a2 =
tosa::getConstTensor<float>(rewriter, op, 0.230389f, {}, dtype).value();
auto x2 = rewriter.create<tosa::MulOp>(loc, outType, absX, absX, /*shift=*/0);
auto a2X = rewriter.create<tosa::MulOp>(loc, outType, a2, x2, /*shift=*/0);
sum = rewriter.create<tosa::AddOp>(loc, outType, sum, a2X);
auto a3 =
tosa::getConstTensor<float>(rewriter, op, 0.000972f, {}, dtype).value();
auto x3 = rewriter.create<tosa::MulOp>(loc, outType, x2, absX, /*shift=*/0);
auto a3X = rewriter.create<tosa::MulOp>(loc, outType, a3, x3, /*shift=*/0);
sum = rewriter.create<tosa::AddOp>(loc, outType, sum, a3X);
auto a4 =
tosa::getConstTensor<float>(rewriter, op, 0.078108f, {}, dtype).value();
auto x4 = rewriter.create<tosa::MulOp>(loc, outType, x3, absX, /*shift=*/0);
auto a4X = rewriter.create<tosa::MulOp>(loc, outType, a4, x4, /*shift=*/0);
sum = rewriter.create<tosa::AddOp>(loc, outType, sum, a4X);
auto rcprl = rewriter.create<tosa::ReciprocalOp>(loc, outType, sum);
auto rcprl2 =
rewriter.create<tosa::MulOp>(loc, outType, rcprl, rcprl, /*shift=*/0);
auto rcprl4 =
rewriter.create<tosa::MulOp>(loc, outType, rcprl2, rcprl2, /*shift=*/0);
auto erf = rewriter.create<tosa::SubOp>(loc, outType, one, rcprl4);
// Deal with negative x.
auto cond = rewriter.create<tosa::GreaterEqualOp>(
loc,
RankedTensorType::get(outType.getShape(), rewriter.getIntegerType(1)), x,
zero);
auto negateErf = rewriter.create<tosa::NegateOp>(loc, outType, erf);
return rewriter.create<tosa::SelectOp>(loc, outType, cond, erf, negateErf);
}
static Value buildUnitNormalCdf(ConversionPatternRewriter &rewriter,
Operation *op, Value x, Type dtype) {
auto zero = tosa::getConstTensor<float>(rewriter, op, 0, {}, dtype).value();
auto one = tosa::getConstTensor<float>(rewriter, op, 1, {}, dtype).value();
auto loc = op->getLoc();
// buildNormalCdf, mean = zero, sigma = one
auto outType = x.getType();
auto mean = zero;
Value xMinusMean = rewriter.create<tosa::SubOp>(loc, outType, x, mean);
// rsqrt of 2
Value rsqrt2 =
tosa::getConstTensor<float>(rewriter, op, 0.70710678f, {}, dtype).value();
Value erfArg = rewriter.create<tosa::MulOp>(loc, outType, xMinusMean, rsqrt2,
/*shift=*/0);
Value erf = approximateErfOp(rewriter, op, erfArg, dtype);
Value erfPlus1 = rewriter.create<tosa::AddOp>(loc, outType, one, erf);
Value oneHalf =
tosa::getConstTensor<float>(rewriter, op, 0.5, {}, dtype).value();
Value normalCdf = rewriter.create<tosa::MulOp>(loc, outType, oneHalf,
erfPlus1, /*shift=*/0);
return normalCdf;
}
// This lowering is based on Torch to LinAlg lowering.
template <>
LogicalResult ConvertAtenOp<AtenGeluOp>::matchAndRewrite(
AtenGeluOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
auto selfElemTy = selfType.getElementType();
if (!isa<mlir::FloatType>(selfElemTy)) {
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization supported");
}
// TODO: Handle approximate.
std::string approximate;
if (!matchPattern(op.getApproximate(), m_TorchConstantStr(approximate)) ||
approximate != "none") {
return rewriter.notifyMatchFailure(op, "Unsupported value of approximate");
}
Value cdf = buildUnitNormalCdf(rewriter, op, adaptor.getSelf(), selfElemTy);
cdf = rewriter.createOrFold<tosa::CastOp>(
op->getLoc(),
cast<RankedTensorType>(cdf.getType()).cloneWith({}, selfElemTy), cdf);
rewriter.replaceOpWithNewOp<tosa::MulOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getSelf(), cdf,
/*shift=*/0);
return success();
}
// This lowering is based on Torch to LinAlg lowering.
template <>
LogicalResult ConvertAtenOp<AtenGeluBackwardOp>::matchAndRewrite(
AtenGeluBackwardOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
auto selfElemTy = selfType.getElementType();
if (!isa<mlir::FloatType>(selfElemTy)) {
return rewriter.notifyMatchFailure(
op, "Only floating-point datatype legalization supported");
}
// TODO: Handle approximate.
std::string approximate;
if (!matchPattern(op.getApproximate(), m_TorchConstantStr(approximate)) ||
approximate != "none") {
return rewriter.notifyMatchFailure(op, "Unsupported value of approximate");
}
auto loc = op->getLoc();
const float cstAlpha0 = 1.12837916709551257390f;
const float cstAlpha1 = 0.70710678118654752440f;
const float oneHalf = 0.5f;
const float kAlpha = cstAlpha0 * cstAlpha1;
Value kAlphaHalf = tosa::getConstTensor<float>(rewriter, op, kAlpha * oneHalf,
{}, selfElemTy)
.value();
Value negOneHalf =
tosa::getConstTensor<float>(rewriter, op, -0.5f, {}, selfElemTy).value();
Value inputSquared = rewriter.create<tosa::MulOp>(
loc, selfType, adaptor.getSelf(), adaptor.getSelf(), /*shift=*/0);
Value negHalfInputSquared = rewriter.create<tosa::MulOp>(
loc, selfType, inputSquared, negOneHalf, /*shift=*/0);
Value dinput =
rewriter.create<tosa::ExpOp>(loc, selfType, negHalfInputSquared);
Value cdf = buildUnitNormalCdf(rewriter, op, adaptor.getSelf(), selfElemTy);
Value dinputInput = rewriter.create<tosa::MulOp>(
loc, selfType, dinput, adaptor.getSelf(), /*shift=*/0);
Value dinputInputAlpha = rewriter.create<tosa::MulOp>(
loc, selfType, dinputInput, kAlphaHalf, /*shift=*/0);
Value cdfExt =
rewriter.create<tosa::AddOp>(loc, selfType, dinputInputAlpha, cdf);
rewriter.replaceOpWithNewOp<tosa::MulOp>(
op, getTypeConverter()->convertType(op.getType()),
adaptor.getGradOutput(), cdfExt,
/*shift=*/0);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenHardtanhBackwardOp>::matchAndRewrite(
AtenHardtanhBackwardOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType) {
return rewriter.notifyMatchFailure(
op, "Only tensor types are currently supported");
}
auto selfElemTy = selfType.getElementType();
if (!selfElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
// Integer types with width > 32 are not supported
auto selfIntType = dyn_cast<IntegerType>(selfElemTy);
if (selfIntType && selfIntType.getWidth() > 32) {
return rewriter.notifyMatchFailure(
op, "Integer types with width greater than 32 are not supported");
}
Value gradOutput = adaptor.getGradOutput();
auto gradOutputType = dyn_cast<TensorType>(adaptor.getSelf().getType());
Type gradOutputElemType = gradOutputType.getElementType();
if (selfElemTy != gradOutputElemType) {
return rewriter.notifyMatchFailure(
op,
"Input element type should be same as the grad_output element type.");
}
SmallVector<int64_t> constTypeShape(selfType.getRank(), 1);
Value maxVal, minVal;
if (failed(torchScalarToTosaTensor(rewriter, op, op.getMinVal(), minVal,
selfElemTy, constTypeShape))) {
return rewriter.notifyMatchFailure(op, "Only scalar constant is supported");
}
if (failed(torchScalarToTosaTensor(rewriter, op, op.getMaxVal(), maxVal,
selfElemTy, constTypeShape))) {
return rewriter.notifyMatchFailure(op, "Only scalar constant is supported");
}
Value replace =
tosa::getConstTensor<float>(rewriter, op, 0, {}, selfElemTy).value();
Type outType = getTypeConverter()->convertType(op.getType());
Value lesser = rewriter.create<tosa::GreaterOp>(
op.getLoc(),
RankedTensorType::get(selfType.getShape(), rewriter.getIntegerType(1)),
minVal, adaptor.getSelf());
Value greater = rewriter.create<tosa::GreaterOp>(
op.getLoc(),
RankedTensorType::get(selfType.getShape(), rewriter.getIntegerType(1)),
adaptor.getSelf(), maxVal);
Value cmp = rewriter.create<tosa::LogicalOrOp>(
op.getLoc(),
RankedTensorType::get(selfType.getShape(), rewriter.getIntegerType(1)),
lesser, greater);
rewriter.replaceOpWithNewOp<tosa::SelectOp>(op, outType, cmp, replace,
gradOutput);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenEmbeddingOp>::matchAndRewrite(
AtenEmbeddingOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value weight = adaptor.getWeight();
Value indices = adaptor.getIndices();
RankedTensorType outType =
cast<RankedTensorType>(typeConverter->convertType(op.getType()));
auto indicesType = dyn_cast<RankedTensorType>(indices.getType());
if (!indicesType || !indicesType.getElementType().isa<IntegerType>())
return rewriter.notifyMatchFailure(
op, "Indices must be of integer tensor type");
auto weightType = cast<RankedTensorType>(weight.getType());
if (weightType.getRank() != 2)
return op.emitError("weight must be of rank 2");
// FIXME: padding_idx, scale_grad_by_freq and sparse are not handled yet.
int64_t paddingIdx;
if (!matchPattern(op.getPaddingIdx(), m_TorchConstantInt(&paddingIdx)))
return rewriter.notifyMatchFailure(
op, "only supports constant int padding_idx for embedding op");
bool scaleGradByFreq;
if (!matchPattern(op.getScaleGradByFreq(),
m_TorchConstantBool(&scaleGradByFreq)))
return rewriter.notifyMatchFailure(
op, "only supports constant bool scale_grad_by_freq for embedding op");
if (scaleGradByFreq)
return rewriter.notifyMatchFailure(
op,
"only supports scale_grad_by_freq equals to False for embedding op");
bool isSparse;
if (!matchPattern(op.getSparse(), m_TorchConstantBool(&isSparse)))
return rewriter.notifyMatchFailure(
op, "only supports constant bool sparse for embedding op");
if (isSparse)
return rewriter.notifyMatchFailure(
op, "only support sparse equals to False for embedding op");
// For inference:
// Weights [num_embeddings, embedding_dim], Indices [X, Y]
// Output [X, Y, embedding_dim] = Weights[Indices[x, y]] forall x in X, y
// in Y
//
// Condition: num_embeddings > Indices [x, y] forall x in X, y in Y
// Reshape the weight, since tosa.gather expects a 3D tensor
auto indicesShape = makeShapeTorchCompatible(indicesType.getShape());
auto weightShape = makeShapeTorchCompatible(weightType.getShape());
SmallVector<int64_t> newWeightShape = {1};
for (auto s : weightShape)
newWeightShape.push_back(s);
auto reshapedWeight = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(newWeightShape),
weightType.getElementType()),
weight, rewriter.getDenseI64ArrayAttr(newWeightShape));
int64_t numIndices = 1;
if (indicesType.hasStaticShape()) {
for (auto s : indicesShape)
numIndices *= s;
} else {
numIndices = kUnknownSize;
}
SmallVector<int64_t> newIndicesShape = {1, numIndices};
auto reshapedIndices = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(newIndicesShape),
indicesType.getElementType()),
indices, rewriter.getDenseI64ArrayAttr(newIndicesShape));
auto castIndices = rewriter.create<tosa::CastOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(newIndicesShape),
rewriter.getIntegerType(32)),
reshapedIndices);
SmallVector<int64_t> intermediateOutShape = {1, numIndices, weightShape[1]};
auto gatherOp = rewriter.create<tosa::GatherOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(intermediateOutShape),
weightType.getElementType()),
reshapedWeight, castIndices);
rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
op, outType, gatherOp,
rewriter.getDenseI64ArrayAttr(
makeShapeTorchCompatible(outType.getShape())));
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenTransposeIntOp>::matchAndRewrite(
AtenTransposeIntOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(op, "Only tensor types are supported");
// Only statically resolvable values are currently supported
int64_t dim0, dim1;
if (!matchPattern(op.getDim0(), m_TorchConstantInt(&dim0)))
return rewriter.notifyMatchFailure(op, "dim0 must be a Scalar constant");
if (!matchPattern(op.getDim1(), m_TorchConstantInt(&dim1)))
return rewriter.notifyMatchFailure(op, "dim1 must be a Scalar constant");
dim0 = toPositiveDim(dim0, selfType.getRank());
dim1 = toPositiveDim(dim1, selfType.getRank());
auto selfRank = selfType.getRank();
if (!isValidDim(dim0, selfRank) || !isValidDim(dim1, selfRank))
return rewriter.notifyMatchFailure(
op, "dim0 and dim1 must be less than tensor rank");
SmallVector<int32_t> transposeDims;
for (auto i = 0; i < selfType.getRank(); ++i)
transposeDims.push_back(i);
transposeDims[dim0] = dim1;
transposeDims[dim1] = dim0;
auto transposeDimsConst = mlir::tosa::getConstTensor<int32_t>(
rewriter, op.getOperation(), transposeDims, {selfType.getRank()});
rewriter.replaceOpWithNewOp<tosa::TransposeOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getSelf(),
transposeDimsConst.value());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenMaxDimOp>::matchAndRewrite(
AtenMaxDimOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(op, "Only tensor types are supported");
auto indicesType =
dyn_cast<TensorType>(getTypeConverter()->convertType(op.getType(1)));
if (!indicesType)
return rewriter.notifyMatchFailure(op, "Only tensor types are supported");
auto selfElemType = selfType.getElementType();
auto indicesElemType = indicesType.getElementType();
// Only statically deducible values are currently supported
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(op, "dim must be a Scalar constant");
dim = toPositiveDim(dim, selfType.getRank());
if (!isValidDim(dim, selfType.getRank()))
return rewriter.notifyMatchFailure(op, "dim must be less than tensor rank");
bool keepDim;
if (!matchPattern(op.getKeepdim(), m_TorchConstantBool(&keepDim)))
return rewriter.notifyMatchFailure(op, "keepdim must be a Scalar constant");
SmallVector<int64_t> reducedShape, prunedShape;
for (auto en :
llvm::enumerate(makeShapeTorchCompatible(selfType.getShape()))) {
if (static_cast<int64_t>(en.index()) == dim) {
reducedShape.push_back(1);
continue;
}
reducedShape.push_back(en.value());
prunedShape.push_back(en.value());
}
auto dimAttr = rewriter.getIntegerAttr(rewriter.getI32Type(), dim);
auto prunedShapeAttr = rewriter.getDenseI64ArrayAttr(prunedShape);
Value reduceMax = rewriter.create<tosa::ReduceMaxOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(reducedShape),
selfElemType),
adaptor.getSelf(), dimAttr);
Value argMax = rewriter.create<tosa::ArgMaxOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(prunedShape),
indicesElemType),
adaptor.getSelf(), dimAttr);
if (argMax.getType() != indicesType) {
argMax = rewriter.create<tosa::ReshapeOp>(
op->getLoc(), indicesType, argMax,
rewriter.getDenseI64ArrayAttr(reducedShape));
}
if (!keepDim) {
reduceMax = rewriter.create<tosa::ReshapeOp>(
op->getLoc(),
RankedTensorType::get(makeShapeLLVMCompatible(prunedShape),
selfElemType),
reduceMax, prunedShapeAttr);
}
rewriter.replaceOp(op, {reduceMax, argMax});
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenSliceTensorOp>::matchAndRewrite(
AtenSliceTensorOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType || !selfType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only tensor types with static shape are supported");
// Only statically deducible values are currently supported
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(op, "dim must be a Scalar constant");
dim = toPositiveDim(dim, selfType.getRank());
if (!isValidDim(dim, selfType.getRank()))
return rewriter.notifyMatchFailure(op, "dim must less than tensor rank");
int64_t start;
if (!matchPattern(op.getStart(), m_TorchConstantInt(&start)))
return rewriter.notifyMatchFailure(op, "start must be a Scalar constant");
if (start < 0) {
start = toPositiveDim(start, selfType.getShape()[dim]);
if (!isValidDim(start, selfType.getShape()[dim]))
return rewriter.notifyMatchFailure(op, "start is not a valid index");
}
start = std::min(selfType.getShape()[dim], start);
int64_t end;
if (!matchPattern(op.getEnd(), m_TorchConstantInt(&end))) {
if (isa<ConstantNoneOp>(op.getEnd().getDefiningOp()))
end = selfType.getShape()[dim];
else
return rewriter.notifyMatchFailure(op, "end must be a Scalar constant");
}
// support for end < 0
end = toPositiveDim(end, selfType.getShape()[dim]);
// support for end out of upper bound
end = (end > selfType.getShape()[dim] ? selfType.getShape()[dim] : end);
// FIXME: add support for start < 0 and end < start
if (end < start)
return rewriter.notifyMatchFailure(op,
"Currently unsupported: end < start");
int64_t step;
if (!matchPattern(op.getStep(), m_TorchConstantInt(&step)))
return rewriter.notifyMatchFailure(op, "step must be a Scalar constant");
if (step != 1)
return rewriter.notifyMatchFailure(
op, "step value other than 1 is currently unsupported");
SmallVector<int64_t> startSlice(selfType.getRank(), 0);
SmallVector<int64_t> sizeSlice =
llvm::to_vector(makeShapeTorchCompatible(selfType.getShape()));
startSlice[dim] = start;
sizeSlice[dim] = end - start;
rewriter.replaceOpWithNewOp<tosa::SliceOp>(
op, getTypeConverter()->convertType(op.getType()), adaptor.getSelf(),
rewriter.getDenseI64ArrayAttr(startSlice),
rewriter.getDenseI64ArrayAttr(sizeSlice));
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenBroadcastToOp>::matchAndRewrite(
AtenBroadcastToOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType || !selfType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only tensor types with static shape are supported");
auto selfElemTy = selfType.getElementType();
if (!selfElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
SmallVector<int64_t> resultShape;
if (!matchPattern(op.getSize(), m_TorchListOfConstantInts(resultShape)))
return rewriter.notifyMatchFailure(op,
"size must consist of Scalar constants");
// Get the result type
auto resultType = getTypeConverter()->convertType(op.getType());
SmallVector<int64_t> inputShape(
makeShapeTorchCompatible(selfType.getShape()));
// Result dimension -1 means not changing the size of that dimension.
// Adjust it by assigning its inputShape.
for (auto shape : llvm::enumerate(makeShapeTorchCompatible(inputShape))) {
auto index = shape.index();
if (resultShape[index] == -1)
resultShape[index] = shape.value();
}
// Check for identity case i.e, for ex: [a, b, c] -> [a, b, c]. If this is
// true then we can replace the op result with the input operand directly.
if (llvm::equal(inputShape, resultShape)) {
// If we reach here, then it means that the broadcasting is not required
// since the input and result are of same shape.
op.replaceAllUsesWith(op.getSelf());
rewriter.eraseOp(op);
return success();
} else if (selfType.hasRank() &&
(selfType.getRank() == (int64_t)resultShape.size() ||
selfType.getRank() == 0)) {
// Right now to support limited cases where input and result shape are not
// equal, we can put a constraint that either the input should be of rank
// 0 or the rank of input tensor and result should be equal. And then we
// can check for broadcasting compatibility for the latter case. For
// broadcasting compatibility, either the shape of input and result should
// be equal at each dimenion or one of them should be 1.
if (selfType.getRank() != 0) {
for (unsigned i = 0; i < inputShape.size(); i++) {
if (inputShape[i] != resultShape[i] && inputShape[i] != 1 &&
resultShape[i] != 1) {
return rewriter.notifyMatchFailure(
op, "unimplemented: either the shape of input and result should "
"be equal at each dimenion or one of them should be 1.");
}
}
}
// If the above condition hold true then we can directly create a const
// zero tensor of shape same as the result shape.
SmallVector<int64_t> zeroTensorShape{resultShape};
// create the 0 constant tensor
int64_t totalNumElements = 1;
for (auto dimSize : zeroTensorShape) {
totalNumElements = dimSize * totalNumElements;
}
// There is some danger here. For edge cases in floating point, x + 0 != x.
// The cases are denormalized values, which may get flushed, and -0 + 0 =
// +0. (sign bit flips). These are probably acceptable in the short term,
// but we should put a comment acknowledging the danger, as there isn't an
// op that avoids the denorm flushing.
Value zeroTensor =
tosa::getZerosLikeTensor(rewriter, op, resultType).value();
// Use add broadcast
rewriter.replaceOpWithNewOp<tosa::AddOp>(op, resultType, adaptor.getSelf(),
zeroTensor);
return success();
}
return rewriter.notifyMatchFailure(
op,
"unimplemented: broadcasts other than same rank or zero ranked tensor.");
}
template <>
LogicalResult ConvertAtenOp<AtenGatherOp>::matchAndRewrite(
AtenGatherOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// For easy understanding of this algorithm, I will comment the code with an
// exact example: torch.aten.gather (!torch.vtensor<[1,4,3],f32>,
// !torch.int-1, !torch.vtensor<[1,4,2],si64>)
// -> !torch.vtensor<[1,4,2],f32>
// https://gist.github.com/AmosLewis/2f18434397025211da4491735bcc6db6
// Not a tensor type.
auto input = adaptor.getSelf();
auto inputType = dyn_cast<RankedTensorType>(adaptor.getSelf().getType());
if (!inputType)
return rewriter.notifyMatchFailure(
op, "Only RankedTensorType input are currently supported");
auto index = adaptor.getIndex();
auto indexType = dyn_cast<RankedTensorType>(adaptor.getIndex().getType());
auto inputShape = inputType.getShape();
int paramsRank = inputShape.size();
if (!indexType)
return rewriter.notifyMatchFailure(
op, "Only RankedTensorType index are currently supported");
// Check `index` and `input` param should have the same rank
if (indexType.getRank() != inputType.getRank())
return rewriter.notifyMatchFailure(
op, "`index` and `input` param should have the same rank");
// Dynamic shape check
if (!inputType.hasStaticShape() || !indexType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "AtenGatherOp: support for dynamic input "
"shape not implemented");
// index i64 to i32 for tosa compatitable
if (indexType.getElementType() != rewriter.getIntegerType(32)) {
index = rewriter.create<tosa::CastOp>(
op->getLoc(),
RankedTensorType::get(indexType.getShape(),
rewriter.getIntegerType(32)),
index);
}
// Get positive dim
int64_t dim{0};
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(
op, "unimplemented: value `dim` should be a torch constant int");
dim = toPositiveDim(dim, paramsRank);
if (!isValidDim(dim, paramsRank))
return rewriter.notifyMatchFailure(op, "Not dim are invalid");
// check sparseGrad is bool type
bool sparseGrad = false;
if (!matchPattern(op.getSparseGrad(), m_TorchConstantBool(&sparseGrad)))
return rewriter.notifyMatchFailure(
op, "only constant boolean `sparse_grad` param supported");
if (sparseGrad)
return rewriter.notifyMatchFailure(
op, "only constant boolean `sparse_grad` == false supported");
// Get the output type
auto outType = getTypeConverter()->convertType(op.getType());
// convert torch style index and dim into tf style indices
// tensor<[1,4,2],si64> -> tensor<[1,4,2,3],si64>
auto indicesTf =
tosa::convertTorchIndexToTfIndices(rewriter, op, input, index, dim);
if (!indicesTf) {
return rewriter.notifyMatchFailure(op,
"Convert TorchIndex To TfIndices fail.");
}
// do the tf gathernp algorithm with tf style indices as input.
auto result =
tosa::convertGatherNdOp(rewriter, op, outType, input, indicesTf.value());
if (!result) {
return rewriter.notifyMatchFailure(op, "Convert GatherNdOp fail.");
}
rewriter.replaceOp(op, {result.value()});
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenIndexPutHackedTwinOp>::matchAndRewrite(
AtenIndexPutHackedTwinOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// a = torch.tensor([[0, 1, 2, 3]])
// a[..., 1:] = torch.tensor([4, 5, 6])
// = a[..., 1:4] = torch.tensor([4, 5, 6])
// = a[[0, 0, 0], [1, 2, 3]] = torch.tensor([4, 5, 6]) # tensor([[0, 4, 5,
// 6]]) = torch.ops.aten.index_put(torch.tensor([[0, 1, 2, 3]]), # input
// (torch.tensor([0, 0, 0]), torch.tensor([1, 2,
// 3])), # indicies torch.tensor([4, 5, 6])) #
// value
// = torch.ops.aten.index_put(torch.tensor([[0, 1, 2, 3]]), # input
// (None, torch.tensor([1, 2, 3]),),# indicies
// torch.tensor([4, 5, 6])) # value
// Not a tensor type.
auto input = adaptor.getSelf();
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
auto fillValues = adaptor.getValues();
auto valuesType = dyn_cast<TensorType>(adaptor.getValues().getType());
if (!valuesType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
// Deal with torch.prim.ListConstruct of non const value to get the index
auto tensorList = op.getIndices();
SmallVector<Value> tensorsTorchType;
if (!getListConstructElements(tensorList, tensorsTorchType))
return op.emitError(
"unimplemented: the tensor list is not from list construct");
auto indexTensors = getTypeConvertedValues(
rewriter, op->getLoc(), getTypeConverter(), tensorsTorchType);
auto outType = getTypeConverter()->convertType(op.getType());
// convert list of indices with none into indices tensor without none
// indexTensors (none,[1,2,3]) -> ([0,0,0],[1,2,3])
// ([[0],[0],[0]],[[1],[2],[3]])-> [[0,1],[0,2], [0,3]]
if (indexTensors.size() <= 1) {
return rewriter.notifyMatchFailure(
op, "Only support indexput with multiple index.");
}
SmallVector<Value> indicesTfConcatTensors;
SmallVector<int64_t> indexesRank;
SmallVector<SmallVector<int64_t>> indexesShape;
// concat index tensor into to indices tensor for concat
for (size_t i = 0; i < indexTensors.size(); i++) {
auto index = indexTensors[i];
auto indexTorch = tensorsTorchType[i];
// TODO add support for none index other than i==0, like (index0, None)
// (None, index1)
if (i == 0 && indexTorch.getType().isa<Torch::NoneType>()) {
// convert None to [0,0,0]
auto indexNext = indexTensors[i + 1];
auto indexNextTorch = tensorsTorchType[i + 1];
if (indexNextTorch.getType().isa<Torch::NoneType>()) {
return rewriter.notifyMatchFailure(
op, "Multiple None index is not support for now.");
}
auto indexNextType = dyn_cast<RankedTensorType>(indexNext.getType());
auto indexNextShape = indexNextType.getShape();
int64_t size = 1;
for (auto s : indexNextShape)
size *= s;
SmallVector<int32_t> values(size, i);
index =
tosa::getConstTensor<int32_t>(rewriter, op, values, indexNextShape)
.value();
}
auto indexType = dyn_cast<RankedTensorType>(index.getType());
auto indexShape = indexType.getShape();
indexesShape.push_back(makeShapeTorchCompatible(indexShape));
indexesRank.push_back(indexType.getRank());
// index i64 to i32 for tosa compatible
if (indexType.getElementType() != rewriter.getIntegerType(32)) {
index = rewriter.create<tosa::CastOp>(
op->getLoc(),
RankedTensorType::get(indexShape, rewriter.getIntegerType(32)),
index);
}
// Expand last dim of index to tf indices [3] -> [3,1]
// convert [0,0,0] to [[0],[0],[0]]
SmallVector<int64_t> indiceShapeOneDim;
for (auto shape : indexShape) {
indiceShapeOneDim.push_back(shape);
}
indiceShapeOneDim.push_back(1);
auto indicesTfOneDim = tosa::CreateOpAndInfer<tosa::ReshapeOp>(
rewriter, op->getLoc(),
RankedTensorType::get(indiceShapeOneDim, rewriter.getIntegerType(32)),
index, rewriter.getDenseI64ArrayAttr(indiceShapeOneDim));
// create concat tensor for indicesTf
// ([[0],[0],[0]], [[1],[2],[3]])
indicesTfConcatTensors.push_back(indicesTfOneDim.getResult());
}
// Right now only support multiple indexes with same shape
// TODO for different shape multiple indexes, add broadcast_to for small
// shape
for (auto indexShapeOneDim : indexesShape) {
if (!llvm::equal(indexesShape[0], indexShapeOneDim)) {
return rewriter.notifyMatchFailure(
op, "unimplemented: Only support multi indexes with same shape");
}
}
// concat each indices into indicesTf: shape ([3,1],[3,1]) -> [3,2]
// ([0,0,0],[1,2,3]) -> [[0,1],[0,2], [0,3]]
auto indicesShapeConcat = indexesShape[0];
uint64_t lastDim = indexesRank[0];
indicesShapeConcat.push_back(indicesTfConcatTensors.size());
auto indicesTf = tosa::CreateOpAndInfer<tosa::ConcatOp>(
rewriter, op->getLoc(),
GetTypeFromTensorShape(indicesShapeConcat, rewriter.getIntegerType(32)),
indicesTfConcatTensors, lastDim);
if (!indicesTf) {
return rewriter.notifyMatchFailure(op,
"Convert TorchIndex To TfIndices fail.");
}
// do the tf scatterNd algorithm with tf style indices as input, algorithm
// mostly take from convertGatherNdOp.
auto result = tosa::convertScatterNdOp(rewriter, op, outType, input,
indicesTf.getResult(), fillValues);
if (!result) {
return rewriter.notifyMatchFailure(
op, "Convert ScatterNdOp fail for index tensor.");
}
rewriter.replaceOp(op, {result.value()});
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenIndexTensorHackedTwinOp>::matchAndRewrite(
AtenIndexTensorHackedTwinOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// t = tf.constant([[1, 2, 3, 4, 5],[6,7,8,9,10],
// [11,12,13,14,15],[16,17,18,19,20]]) # 4*5
// i = tf.constant([[1,2,3], [3,2,1]]) # 2*3
// i_expand = tf.expand_dims(i,axis=2) # 2*3*1
// IndexTensorOutput = tf.gather_nd(t,tf.i_expand)
// = torch.ops.aten.index(t, (i, )) = t[i] # 2*3*5
// [[[ 6, 7, 8, 9, 10], [11, 12, 13, 14, 15], [16, 17, 18, 19, 20]],
// [[16, 17, 18, 19, 20], [11, 12, 13, 14, 15], [ 6, 7, 8, 9, 10]]]
auto input = adaptor.getSelf();
auto inputTensorType =
dyn_cast<RankedTensorType>(adaptor.getSelf().getType());
// Check input is a tensor type.
if (!inputTensorType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
// Deal with torch.prim.ListConstruct of non const value to get the index
auto tensorList = op.getIndices();
SmallVector<Value> tensorsTorchType;
if (!getListConstructElements(tensorList, tensorsTorchType))
return op.emitError(
"unimplemented: the tensor list is not from list construct");
auto indexTensors = getTypeConvertedValues(
rewriter, op->getLoc(), getTypeConverter(), tensorsTorchType);
auto outType = getTypeConverter()->convertType(op.getType());
// Support for multiple indexes
if (indexTensors.size() > 1) {
// t[i, i]
// = torch.ops.aten.index(t,(i,i))
// = tensor([[ t[1,1], t[2,2], t[3,3]],
// [ t[3,3], t[2,2], t[1,1]]])
// = tensor([[ 7, 13, 19], [19, 13, 7]])
// = tf.gather_nd(t,tf.ii_expand)
// ii_expand
// = tf.concat((i_expand,i_expand), dim=2)
// = tf.constant([[[1,1],[2,2],[3,3]],
// [[3,3],[2,2],[1,1]]]) # 2*3*2
SmallVector<Value> indicesTfConcatTensors;
SmallVector<int64_t> indexesRank;
SmallVector<SmallVector<int64_t>> indexesShape;
// concat index tensor into to indices tensor for concat
for (size_t i = 0; i < indexTensors.size(); i++) {
auto index = indexTensors[i];
auto indexType = dyn_cast<RankedTensorType>(index.getType());
auto indexShape = indexType.getShape();
indexesShape.push_back(makeShapeTorchCompatible(indexShape));
indexesRank.push_back(indexType.getRank());
// Make type of index tosa compatible, i64 to i32.
if (indexType.getElementType() != rewriter.getIntegerType(32)) {
index = rewriter.create<tosa::CastOp>(
op->getLoc(),
RankedTensorType::get(indexShape, rewriter.getIntegerType(32)),
index);
}
// Expand last dim of index to tf indices [2,3] -> [2,3,1]
SmallVector<int64_t> indiceShapeOneDim;
for (auto shape : indexShape) {
indiceShapeOneDim.push_back(shape);
}
indiceShapeOneDim.push_back(1);
auto indicesTfOneDim = tosa::CreateOpAndInfer<tosa::ReshapeOp>(
rewriter, op->getLoc(),
RankedTensorType::get(indiceShapeOneDim, rewriter.getIntegerType(32)),
index, rewriter.getDenseI64ArrayAttr(indiceShapeOneDim));
// create concat tensor for indicesTf
indicesTfConcatTensors.push_back(indicesTfOneDim.getResult());
}
// Right now only support multiple indexes with same shape
// TODO for different shape multiple indexes, add broadcast_to for small
// shape
for (auto indexShapeOneDim : indexesShape) {
if (!llvm::equal(indexesShape[0], indexShapeOneDim)) {
return rewriter.notifyMatchFailure(
op, "unimplemented: Only support multi indexes with same shape");
}
}
// concat each indices into indicesTf: shape [2,3,1],[2,3,1] -> [2,3,2]
auto indicesShapeConcat = indexesShape[0];
uint64_t lastDim = indexesRank[0];
indicesShapeConcat.push_back(indicesTfConcatTensors.size());
auto indicesTf = tosa::CreateOpAndInfer<tosa::ConcatOp>(
rewriter, op->getLoc(),
GetTypeFromTensorShape(indicesShapeConcat, rewriter.getIntegerType(32)),
indicesTfConcatTensors, lastDim);
if (!indicesTf) {
return rewriter.notifyMatchFailure(
op, "Convert TorchIndex To TfIndices fail.");
}
// do the tf gathernp algorithm with tf style indices as input.
auto result = tosa::convertGatherNdOp(rewriter, op, outType, input,
indicesTf.getResult());
if (!result) {
return rewriter.notifyMatchFailure(
op, "Convert GatherNdOp fail for index tensor.");
}
rewriter.replaceOp(op, {result.value()});
return success();
}
// Support for multiple index
auto index = indexTensors[0];
auto indexType = dyn_cast<RankedTensorType>(index.getType());
auto indexShape = indexType.getShape();
// index i64 to i32 for tosa compatible
if (indexType.getElementType() != rewriter.getIntegerType(32)) {
index = rewriter.create<tosa::CastOp>(
op->getLoc(),
RankedTensorType::get(indexShape, rewriter.getIntegerType(32)), index);
}
// Expand last dim of index to tf indices [2,3] -> [2,3,1]
SmallVector<int64_t> indicesShape;
for (auto shape : indexShape) {
indicesShape.push_back(shape);
}
indicesShape.push_back(1);
auto indicesTf = tosa::CreateOpAndInfer<tosa::ReshapeOp>(
rewriter, op->getLoc(),
RankedTensorType::get(indicesShape, rewriter.getIntegerType(32)), index,
rewriter.getDenseI64ArrayAttr(indicesShape));
if (!indicesTf) {
return rewriter.notifyMatchFailure(op,
"Convert TorchIndex To TfIndices fail.");
}
// do the tf gathernp algorithm with tf style indices as input.
auto result = tosa::convertGatherNdOp(rewriter, op, outType, input,
indicesTf.getResult());
if (!result) {
return rewriter.notifyMatchFailure(
op, "Convert GatherNdOp fail for index tensor.");
}
rewriter.replaceOp(op, {result.value()});
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenAbsOp>::matchAndRewrite(
AtenAbsOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
auto outType = getTypeConverter()->convertType(op.getType());
rewriter.replaceOpWithNewOp<tosa::AbsOp>(op, outType, adaptor.getSelf());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenWhereSelfOp>::matchAndRewrite(
AtenWhereSelfOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
auto condType = dyn_cast<TensorType>(adaptor.getCondition().getType());
if (!condType)
return rewriter.notifyMatchFailure(
op, "Only tensor types condition are currently supported");
auto outType = getTypeConverter()->convertType(op.getType());
rewriter.replaceOpWithNewOp<tosa::SelectOp>(
op, outType, adaptor.getCondition(), adaptor.getSelf(),
adaptor.getOther());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenLeTensorOp>::matchAndRewrite(
AtenLeTensorOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
auto otherType = dyn_cast<TensorType>(adaptor.getOther().getType());
if (!otherType)
return rewriter.notifyMatchFailure(
op, "Only tensor types condition are currently supported");
auto outType = getTypeConverter()->convertType(op.getType());
auto greaterOp = rewriter.create<tosa::GreaterOp>(
op.getLoc(), outType, adaptor.getSelf(), adaptor.getOther());
rewriter.replaceOpWithNewOp<tosa::LogicalNotOp>(op, outType,
greaterOp.getOutput());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenIscloseOp>::matchAndRewrite(
AtenIscloseOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// check args
double rtol, atol;
bool equalNan;
if (!matchPattern(op.getRtol(), m_TorchConstantFloat(&rtol)))
return rewriter.notifyMatchFailure(op, "rtol must be a scalar constant");
if (!matchPattern(op.getAtol(), m_TorchConstantFloat(&atol)))
return rewriter.notifyMatchFailure(op, "atol must be a scalar constant");
if (!matchPattern(op.getEqualNan(), m_TorchConstantBool(&equalNan)))
return rewriter.notifyMatchFailure(
op, "unimplemented: equal_nan is expected to be false");
// check tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
auto otherType = dyn_cast<TensorType>(adaptor.getOther().getType());
if (!selfType || !otherType)
return rewriter.notifyMatchFailure(
op, "Only tensor types input are currently supported");
if (!selfType.hasStaticShape() || !otherType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only tensor types with static shape are supported");
if (!selfType.getElementType().isa<mlir::FloatType>() ||
!otherType.getElementType().isa<mlir::FloatType>()) {
return rewriter.notifyMatchFailure(
op, "unimplemented: only FP element type is supported");
}
auto rhsSubOp = rewriter.create<tosa::SubOp>(
op->getLoc(), selfType, adaptor.getSelf(), adaptor.getOther());
auto rhsAbsOp =
rewriter.create<tosa::AbsOp>(op->getLoc(), selfType, rhsSubOp);
auto lhsAbsOp =
rewriter.create<tosa::AbsOp>(op->getLoc(), otherType, adaptor.getOther());
auto rtolConstOp =
tosa::getTosaConstTensorSingleF32(rewriter, op, static_cast<float>(rtol));
auto mulOp = rewriter.create<tosa::MulOp>(op->getLoc(), otherType,
rtolConstOp, lhsAbsOp, /*shift=*/0);
auto atolConstOp =
tosa::getTosaConstTensorSingleF32(rewriter, op, static_cast<float>(atol));
auto addOp =
rewriter.create<tosa::AddOp>(op->getLoc(), otherType, atolConstOp, mulOp);
auto outType = getTypeConverter()->convertType(op.getType());
rewriter.replaceOpWithNewOp<tosa::GreaterEqualOp>(op, outType, addOp,
rhsAbsOp);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenClampOp>::matchAndRewrite(
AtenClampOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType)
return rewriter.notifyMatchFailure(
op, "only tensor types input are currently supported");
IntegerAttr min_int =
rewriter.getI64IntegerAttr(std::numeric_limits<int64_t>::min());
IntegerAttr max_int =
rewriter.getI64IntegerAttr(std::numeric_limits<int64_t>::max());
FloatAttr min_fp =
rewriter.getF32FloatAttr(std::numeric_limits<float>::lowest());
FloatAttr max_fp =
rewriter.getF32FloatAttr(std::numeric_limits<float>::max());
auto getValAttr = [&](Value operand, IntegerAttr &intAttr,
FloatAttr &fpAttr) -> LogicalResult {
double valFloat;
int64_t valInt;
if (matchPattern(operand, m_TorchConstantFloat(&valFloat))) {
intAttr = rewriter.getI64IntegerAttr(static_cast<int64_t>(valFloat));
fpAttr = rewriter.getF32FloatAttr(static_cast<float>(valFloat));
} else if (matchPattern(operand, m_TorchConstantInt(&valInt))) {
intAttr = rewriter.getI64IntegerAttr(valInt);
fpAttr = rewriter.getF32FloatAttr(static_cast<float>(valInt));
} else {
return failure();
}
return success();
};
LogicalResult minAttrResult = getValAttr(op.getMin(), min_int, min_fp);
LogicalResult maxAttrResult = getValAttr(op.getMax(), max_int, max_fp);
if (failed(minAttrResult) && failed(maxAttrResult)) {
return rewriter.notifyMatchFailure(
op, "either `min` or `max` should be a torch constant");
}
if (failed(minAttrResult) &&
succeeded(checkNotNone(rewriter, op, op.getMin()))) {
return rewriter.notifyMatchFailure(op,
"min attr should be a torch constant");
}
if (failed(maxAttrResult) &&
succeeded(checkNotNone(rewriter, op, op.getMax()))) {
return rewriter.notifyMatchFailure(op,
"max attr should be a torch constant");
}
auto outType = getTypeConverter()->convertType(op.getType());
rewriter.replaceOpWithNewOp<tosa::ClampOp>(op, outType, adaptor.getSelf(),
min_int, max_int, min_fp, max_fp);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenArangeStartStepOp>::matchAndRewrite(
AtenArangeStartStepOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
const TypeConverter *typeConverter = this->getTypeConverter();
RankedTensorType resultType =
typeConverter->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
// At this point all tensors should have value semantics, and hence the
// `layout` check can be ignored.
// TODO: Add support for pin_memory features.
// The pin_memory should be either `False` or `none`.
bool pinMemory;
if (!op.getPinMemory().getType().isa<Torch::NoneType>() &&
(!matchPattern(op.getPinMemory(), m_TorchConstantBool(&pinMemory)) ||
pinMemory)) {
return rewriter.notifyMatchFailure(
op, "unimplemented: pin_memory must be either None or false");
}
// Stores a range value (a start, end, or step value) and whether or not it
// was initiated with a constant integer, an constant float or neither.
class ConstRangeValue {
public:
explicit ConstRangeValue(double v)
: vDouble(v), fromDouble(true), vInt(static_cast<int64_t>(v)),
fromInt(false) {}
explicit ConstRangeValue(int64_t v)
: vDouble(static_cast<double>(v)), fromDouble(false), vInt(v),
fromInt(true) {}
// Constructor for the case where there is no constant value to use.
ConstRangeValue()
: vDouble(0), fromDouble(false), vInt(0), fromInt(false) {}
static ConstRangeValue fromValue(Value v) {
int64_t intVal{0};
double floatVal{0.0};
if (matchPattern(v, m_TorchConstantFloat(&floatVal))) {
return ConstRangeValue(floatVal);
} else if (matchPattern(v, m_TorchConstantInt(&intVal))) {
return ConstRangeValue(intVal);
}
return ConstRangeValue();
}
bool hasConstInt() const { return fromInt; }
bool hasConstDouble() const { return fromDouble; }
bool hasConst() const { return fromInt || fromDouble; }
double getDouble() const { return vDouble; }
int64_t getInt() const { return vInt; }
private:
double vDouble;
bool fromDouble;
int64_t vInt;
bool fromInt;
};
auto start = ConstRangeValue::fromValue(op.getStart());
if (!start.hasConst()) {
return rewriter.notifyMatchFailure(
op, "unimplemented: case where `start` is not a constant int or float");
}
auto end = ConstRangeValue::fromValue(op.getEnd());
if (!end.hasConst()) {
return rewriter.notifyMatchFailure(
op,
"unimplemented: case where value `end` is not a constant int or float");
}
auto step = ConstRangeValue::fromValue(op.getStep());
if (!step.hasConst()) {
return rewriter.notifyMatchFailure(op,
"unimplemented: case where value `step` "
"is not a constant int or float");
}
auto getRange = [](auto start, auto end, auto step) {
// Initialize a small vector of the same type as start:
using T = decltype(start);
SmallVector<T> values;
uint64_t counter{0};
if (start == end) {
return values;
}
assert(step != T(0));
values.reserve(
1 + static_cast<size_t>(std::abs((end - start) / std::abs(step))));
if (step > 0) {
while (start + T(counter) * step < end) {
values.push_back(start + counter * step);
counter++;
}
} else {
while (start + T(counter) * step > end) {
values.push_back(start + counter * step);
counter++;
}
}
return values;
};
const auto isIntType =
resultType.getElementType().dyn_cast_or_null<mlir::IntegerType>();
const auto isDoubleType =
resultType.getElementType().dyn_cast_or_null<mlir::FloatType>();
auto maybeResult = [&]() -> std::optional<Value> {
// Integer output type, and start / end / range are all integers.
if (isIntType && start.hasConstInt() && end.hasConstInt() &&
step.hasConstInt()) {
auto values = getRange(start.getInt(), end.getInt(), step.getInt());
return tosa::getConstTensor<int64_t>(rewriter, op, values, values.size());
}
// Get a double range.
auto values =
getRange(start.getDouble(), end.getDouble(), step.getDouble());
if (isIntType) {
SmallVector<int64_t> values_i64;
values_i64.reserve(values.size());
for (auto v : values) {
values_i64.push_back(static_cast<int64_t>(v));
}
return tosa::getConstTensor<int64_t>(rewriter, op, values_i64,
values.size());
}
if (!isDoubleType) {
return {};
}
SmallVector<float> values_f32;
values_f32.reserve(values.size());
for (auto v : values) {
values_f32.push_back(static_cast<float>(v));
}
auto vs = tosa::getConstTensor<float>(rewriter, op, values_f32,
values_f32.size());
return vs;
}();
if (!maybeResult.has_value()) {
return rewriter.notifyMatchFailure(
op, "failed to generate constant tensor for arange");
}
auto result = maybeResult.value();
rewriter.replaceOpWithNewOp<tosa::CastOp>(op, resultType, result);
return success();
}
template <>
LogicalResult ConvertAtenOp<PrimNumToTensorScalarOp>::matchAndRewrite(
PrimNumToTensorScalarOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
const TypeConverter *typeConverter = this->getTypeConverter();
RankedTensorType resultType =
typeConverter->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
// Only supports integer operand type, because for the floating point operand
// type result tensor has to be of type `f64` which is not supported in the
// tosa.
double doubleValue;
auto isDouble = matchPattern(op.getA(), m_TorchConstantFloat(&doubleValue));
int64_t intValue;
auto isInt = matchPattern(op.getA(), m_TorchConstantInt(&intValue));
if (!isDouble && !isInt)
return rewriter.notifyMatchFailure(op,
"Unable to extract the scalar constant");
auto outElemTy = resultType.getElementType();
if (isa<mlir::IntegerType>(outElemTy)) {
rewriter.replaceOpWithNewOp<tosa::ConstOp>(
op, resultType, DenseElementsAttr::get(resultType, {intValue}));
} else if (outElemTy.isF64()) {
rewriter.replaceOpWithNewOp<tosa::ConstOp>(
op, resultType, DenseElementsAttr::get(resultType, {doubleValue}));
}
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenCopyOp>::matchAndRewrite(
AtenCopyOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
auto srcType = dyn_cast<TensorType>(adaptor.getSrc().getType());
if (!selfType || !selfType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only tensor types with static shape are supported");
if (!srcType || !srcType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only tensor types with static shape are supported");
// The non_blocking should be a constant `False`.
bool nonBlocking;
if (!matchPattern(op.getNonBlocking(), m_TorchConstantBool(&nonBlocking))) {
return rewriter.notifyMatchFailure(
op, "unimplemented: non_blocking must be a constant");
} else if (nonBlocking) {
return rewriter.notifyMatchFailure(
op, "unimplemented: non_blocking is expected to be false");
}
SmallVector<int64_t> selfShape(makeShapeTorchCompatible(selfType.getShape()));
SmallVector<int64_t> srcShape(makeShapeTorchCompatible(srcType.getShape()));
if (llvm::equal(selfShape, srcShape) || selfShape.size() == 0) {
// If we reach here, then it means the given case is handled by implicit
// broadcasting done by tosa.
Value result;
if (failed(tosa::tosaCastTensorToType(
rewriter, op, adaptor.getSrc(),
getTypeConverter()->convertType(op.getType()), result)))
return rewriter.notifyMatchFailure(
op, "unimplemented: cast to result type not supported");
rewriter.replaceOp(op, result);
return success();
}
return rewriter.notifyMatchFailure(
op, "unimplemented: valsem.aten.copy op not supported for this case.");
}
// Legalizes the torch.aten.to.dtype op
template <>
LogicalResult ConvertAtenOp<AtenToDtypeOp>::matchAndRewrite(
AtenToDtypeOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType || !selfType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only tensor types with static shape are supported");
// The non_blocking arg should be a constant `False`.
bool nonBlocking;
if (!matchPattern(op.getNonBlocking(), m_TorchConstantBool(&nonBlocking))) {
return rewriter.notifyMatchFailure(
op, "unimplemented: non_blocking arg must be a constant");
} else if (nonBlocking) {
return rewriter.notifyMatchFailure(
op, "unimplemented: non_blocking arg is expected to be false");
}
// The copy arg should be a constant `False`.
bool copy;
if (!matchPattern(op.getCopy(), m_TorchConstantBool(&copy))) {
return rewriter.notifyMatchFailure(
op, "unimplemented: copy arg must be a constant");
} else if (copy) {
return rewriter.notifyMatchFailure(
op, "unimplemented: copy arg is expected to be false");
}
// Only `none`, `contiguous` and `preserve` memory_format is supported.
if (!op.getMemoryFormat().getType().isa<Torch::NoneType>()) {
int64_t memoryFormat;
if (!matchPattern(op.getMemoryFormat(), m_TorchConstantInt(&memoryFormat)))
return rewriter.notifyMatchFailure(
op, "unimplemented: the memory format should be specified in "
"an integer constant");
if (memoryFormat != torch_upstream::MemoryFormat::Contiguous &&
memoryFormat != torch_upstream::MemoryFormat::Preserve)
return rewriter.notifyMatchFailure(
op, "unimplemented: only none, contiguous and preserve "
"memory_format is supported");
}
auto resultTy = getTypeConverter()
->convertType(op.getResult().getType())
.cast<RankedTensorType>();
Value result;
if (failed(tosa::tosaCastTensorToType(rewriter, op, adaptor.getSelf(),
resultTy, result)))
return rewriter.notifyMatchFailure(op, "conversion to result type failed");
rewriter.replaceOp(op, result);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenRemainderScalarOp>::matchAndRewrite(
AtenRemainderScalarOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Value self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(
op, "Only ranked tensor types supported in TOSA Remainder");
auto outType =
cast<TensorType>(getTypeConverter()->convertType(op.getType()));
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat())
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
Value otherTensor;
Value other = op.getOther();
if (failed(torchScalarToTosaTensor(rewriter, op, other, otherTensor,
outElemTy, {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA Remainder operation");
if (selfTy.getElementType() != outElemTy)
self = rewriter.create<tosa::CastOp>(op.getLoc(), outType, self);
auto divTensor = self;
if (isa<mlir::FloatType>(outElemTy)) {
auto otherTensorReciprocal = rewriter.create<tosa::ReciprocalOp>(
op.getLoc(), otherTensor.getType(), otherTensor);
divTensor = rewriter.create<tosa::MulOp>(
op.getLoc(), outType, self, otherTensorReciprocal, /*shift=*/0);
divTensor = rewriter.create<tosa::FloorOp>(op.getLoc(), outType, divTensor);
} else {
divTensor = rewriter.create<tosa::IntDivOp>(op.getLoc(), outType, self,
otherTensor);
}
auto mulTensor =
rewriter.create<tosa::MulOp>(op.getLoc(), outType, otherTensor, divTensor,
/*shift=*/0);
rewriter.replaceOpWithNewOp<tosa::SubOp>(op, outType, self, mulTensor);
return success();
}
template <typename AtenOpT, typename TosaOpT>
class ConvertAtenPoolingBaseOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
// Different pooling variants need to process inputs differently, e.g.
// adaptive pooling generates the kernel size rather than receive it. This
// function also transposes inputs.
virtual LogicalResult processInputs(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter,
Value &input, DenseI64ArrayAttr &kernel,
DenseI64ArrayAttr &stride,
DenseI64ArrayAttr &pad,
Type &outputTy) const {
return rewriter.notifyMatchFailure(
op, "Unimplemented pooling input parsing function");
}
static int64_t getOutputDim(int64_t inputDim, int64_t kernelDim,
int64_t stride, int64_t padBefore,
int64_t padAfter, int64_t dilation,
bool ceilMode = false) {
if (inputDim == kUnknownSize) {
return kUnknownSize;
} else {
int64_t dimSize =
inputDim + padBefore + padAfter - dilation * (kernelDim - 1) - 1;
if (ceilMode && (dimSize % stride != 0))
return dimSize / stride + 2;
return dimSize / stride + 1;
}
}
// Apply the transposeDims vector on input to generate a transposed form.
Value transposeTensor(AtenOpT op, ConversionPatternRewriter &rewriter,
Value input, ArrayRef<int32_t> transposeDims) const {
auto inputTy = cast<RankedTensorType>(input.getType());
auto inputElemTy = inputTy.getElementType();
auto inputShape = makeShapeTorchCompatible(inputTy.getShape());
auto inputRank = inputTy.getRank();
std::optional<Value> transposeDimsConst = tosa::getConstTensor<int32_t>(
rewriter, op,
/*vec=*/transposeDims,
/*shape=*/{static_cast<int32_t>(inputRank)});
SmallVector<int64_t> transposedInputShape;
for (auto &dim : transposeDims)
transposedInputShape.push_back(inputShape[dim]);
auto transposedInputType = RankedTensorType::get(
makeShapeLLVMCompatible(transposedInputShape), inputElemTy);
return rewriter
.create<tosa::TransposeOp>(op->getLoc(), transposedInputType, input,
transposeDimsConst.value())
.getResult();
}
Value transposePoolingInputToHwc(AtenOpT op,
ConversionPatternRewriter &rewriter,
Value input) const {
auto inputRank = cast<RankedTensorType>(input.getType()).getRank();
SmallVector<int32_t> nchwToNhwc4DTransposeDims({0, 2, 3, 1});
SmallVector<int32_t> chwToHwc3DTransposeDims({1, 2, 0});
return transposeTensor(op, rewriter, input,
inputRank == 3 ? chwToHwc3DTransposeDims
: nchwToNhwc4DTransposeDims);
}
Value transposePoolingOutputToChw(AtenOpT op,
ConversionPatternRewriter &rewriter,
Value input) const {
auto inputTy = cast<RankedTensorType>(input.getType());
auto inputRank = inputTy.getRank();
SmallVector<int32_t> nhwcToNchw4DTransposeDims({0, 3, 1, 2});
SmallVector<int32_t> hwcToChw3DTransposeDims({2, 0, 1});
return transposeTensor(op, rewriter, input,
inputRank == 3 ? hwcToChw3DTransposeDims
: nhwcToNchw4DTransposeDims);
}
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value input;
DenseI64ArrayAttr kernel, stride, pad;
Type outputTy;
// Attempts to read input and kernel parameters, or synthesize them in the
// case of adaptive pooling. Also performs input CHW->HWC transpose.
if (failed(processInputs(op, adaptor, rewriter, input, kernel, stride, pad,
outputTy)))
return rewriter.notifyMatchFailure(
op, "Failed to process inputs for pooling");
Value pooledOutput;
static_assert(std::is_same<TosaOpT, tosa::MaxPool2dOp>::value ||
std::is_same<TosaOpT, tosa::AvgPool2dOp>::value,
"Expected either tosa::MaxPool2dOp or tosa::AvgPool2dOp");
if constexpr (std::is_same<TosaOpT, tosa::MaxPool2dOp>::value) {
pooledOutput = rewriter
.create<TosaOpT>(op->getLoc(), outputTy, input, kernel,
stride, pad)
.getResult();
} else if constexpr (std::is_same<TosaOpT, tosa::AvgPool2dOp>::value) {
TypeAttr accType;
if (failed(tosa::getAvgPool2dAccType(rewriter, input, accType)))
return rewriter.notifyMatchFailure(
op, "Failed to get accumulator type for pooling");
pooledOutput = rewriter
.create<TosaOpT>(op->getLoc(), outputTy, input, kernel,
stride, pad, accType)
.getResult();
}
auto transposedOutput =
ConvertAtenPoolingBaseOp<AtenOpT, TosaOpT>::transposePoolingOutputToChw(
op, rewriter, pooledOutput);
rewriter.replaceOpWithNewOp<tensor::CastOp>(
op,
OpConversionPattern<AtenOpT>::getTypeConverter()->convertType(
op.getType()),
transposedOutput);
return success();
}
};
template <typename AtenOpT, typename TosaOpT>
class ConvertAtenAdaptivePoolingOp
: public ConvertAtenPoolingBaseOp<AtenOpT, TosaOpT> {
public:
using ConvertAtenPoolingBaseOp<AtenOpT, TosaOpT>::ConvertAtenPoolingBaseOp;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult processInputs(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter, Value &input,
DenseI64ArrayAttr &kernel,
DenseI64ArrayAttr &stride, DenseI64ArrayAttr &pad,
Type &outputTy) const override {
auto inputXchw = adaptor.getSelf();
auto inputTy = cast<RankedTensorType>(inputXchw.getType());
if (!inputTy)
return rewriter.notifyMatchFailure(
op, "Adaptive avgpool requires ranked tensor input");
auto inputShape = makeShapeTorchCompatible(inputTy.getShape());
auto inputRank = inputTy.getRank();
auto inputElemTy = inputTy.getElementType();
// Rank sanity check.
if (inputTy.getRank() != 4 && inputRank != 3)
return rewriter.notifyMatchFailure(
op, "NCHW->NHWC transpose requires 3D or 4D tensor");
int64_t inputHDim = inputShape[inputRank - 2];
int64_t inputWDim = inputShape[inputRank - 1];
SmallVector<int64_t> outputSize;
if (!matchPattern(op.getOutputSize(),
m_TorchListOfConstantInts(outputSize)))
return rewriter.notifyMatchFailure(
op, "Non-const output_size for adaptive pooling unsupported.");
SmallVector<int64_t> kernelDims;
int64_t outputHDim, outputWDim;
if (outputSize.size() == 1) {
outputHDim = outputWDim = outputSize[0];
} else {
if (outputSize.size() != 2)
return rewriter.notifyMatchFailure(
op, "Adaptive avgpool output_size not 1 or 2 elements.");
// Assumes 'None' (e.g. output_size=(None, 5) ) is expressed as <=0.
outputHDim =
(outputSize[0] <= 0) ? inputShape[inputRank - 2] : outputSize[0];
outputWDim =
(outputSize[1] <= 0) ? inputShape[inputRank - 1] : outputSize[1];
}
// In adaptive pooling,
// stride = inputDim // outputDim
// kernel = inputDim - (outputDim-1)* stride
// pad = 0, dilation = 1
int64_t strideH = inputShape[inputRank - 2] / outputHDim;
int64_t strideW = inputShape[inputRank - 1] / outputWDim;
kernelDims.push_back(inputHDim - (outputHDim - 1) * strideH);
kernelDims.push_back(inputWDim - (outputWDim - 1) * strideW);
SmallVector<int64_t> outputShape;
if (inputRank > 3)
outputShape.push_back(inputShape[0]);
outputShape.push_back(outputHDim);
outputShape.push_back(outputWDim);
outputShape.push_back(inputShape[inputRank - 3]);
// Transpose to xHWC
input =
ConvertAtenPoolingBaseOp<AtenOpT, TosaOpT>::transposePoolingInputToHwc(
op, rewriter, inputXchw);
kernel = rewriter.getDenseI64ArrayAttr(kernelDims);
stride = rewriter.getDenseI64ArrayAttr({strideH, strideW});
// Adaptive pooling does unit dilation and zero pad.
pad = rewriter.getDenseI64ArrayAttr({0, 0, 0, 0});
outputTy = RankedTensorType::get(makeShapeLLVMCompatible(outputShape),
inputElemTy);
return success();
}
};
template <typename AtenOpT, typename tosaOp>
static Type getOutputTypeForNonAdaptivePoolingOp(
RankedTensorType inputTy, SmallVectorImpl<int64_t> &kernelSize,
SmallVectorImpl<int64_t> &strideArray, SmallVectorImpl<int64_t> &padArray,
SmallVectorImpl<int64_t> &dilationArray, bool ceilMode = false) {
auto inputShape = makeShapeTorchCompatible(inputTy.getShape());
auto inputRank = inputTy.getRank();
auto inputElemTy = inputTy.getElementType();
int64_t outputHDim = ConvertAtenPoolingBaseOp<AtenOpT, tosaOp>::getOutputDim(
inputShape[inputRank - 2], kernelSize[0], strideArray[0], padArray[0],
padArray[0], dilationArray[0], ceilMode);
int64_t outputWDim = ConvertAtenPoolingBaseOp<AtenOpT, tosaOp>::getOutputDim(
inputShape[inputRank - 1], kernelSize[1], strideArray[1], padArray[1],
padArray[1], dilationArray[1], ceilMode);
padArray[0] = (outputHDim - 1) * strideArray[0] +
dilationArray[0] * kernelSize[0] - dilationArray[0] + 1 -
padArray[0] * 2 - inputShape[inputRank - 2];
padArray[1] = (outputWDim - 1) * strideArray[1] +
dilationArray[0] * kernelSize[1] - dilationArray[0] + 1 -
padArray[1] * 2 - inputShape[inputRank - 1];
SmallVector<int64_t> outputShape;
if (inputRank > 3)
outputShape.push_back(inputShape[0]);
outputShape.push_back(outputHDim);
outputShape.push_back(outputWDim);
outputShape.push_back(inputShape[inputRank - 3]);
return RankedTensorType::get(makeShapeLLVMCompatible(outputShape),
inputElemTy);
}
// Checks the validity of pooling parameters and stores them in the respective
// vector. Also, gets the output type for the pooling op.
template <typename AtenOpT, typename tosaOp>
static LogicalResult getOutputTypeAndPoolingParameters(
AtenOpT op, ConversionPatternRewriter &rewriter, Value inputXchw,
SmallVectorImpl<int64_t> &dilationArray, Type &outputTy,
DenseI64ArrayAttr &kernel, DenseI64ArrayAttr &stride,
DenseI64ArrayAttr &pad) {
RankedTensorType inputTy = cast<RankedTensorType>(inputXchw.getType());
if (!inputTy)
return rewriter.notifyMatchFailure(
op, "Pooling op requires ranked tensor input");
auto inputRank = inputTy.getRank();
// Rank sanity check.
if (inputTy.getRank() != 4 && inputRank != 3)
return rewriter.notifyMatchFailure(
op, "NCHW->NHWC transpose requires 3D or 4D tensor");
SmallVector<int64_t, 2> kernelSizeInts, strideInts, paddingInts;
if (!matchPattern(op.getKernelSize(),
m_TorchListOfConstantInts(kernelSizeInts)))
return rewriter.notifyMatchFailure(
op, "Non-const kernel_size for pooling op unsupported");
if (!matchPattern(op.getStride(), m_TorchListOfConstantInts(strideInts)))
return rewriter.notifyMatchFailure(
op, "Non-const stride for pooling op unsupported");
// If `stride` is not specified by the user, it is assigned the value of empty
// list during import. For such a case, the stride value is the kernel size.
// See:
// https://pytorch.org/docs/stable/generated/torch.nn.MaxPool2d.html#torch.nn.MaxPool2d
if (strideInts.empty())
strideInts.assign(kernelSizeInts);
if (!matchPattern(op.getPadding(), m_TorchListOfConstantInts(paddingInts)))
return rewriter.notifyMatchFailure(
op, "Non-const padding factor for pooling op unsupported");
SmallVector<int64_t, 4> padArr = {paddingInts[0], paddingInts[0],
paddingInts[1], paddingInts[1]};
kernel = rewriter.getDenseI64ArrayAttr(kernelSizeInts);
stride = rewriter.getDenseI64ArrayAttr(strideInts);
bool ceilMode;
if (!matchPattern(op.getCeilMode(), m_TorchConstantBool(&ceilMode)))
return rewriter.notifyMatchFailure(
op, "only support constant bool ceil_mode for pooling op");
outputTy = getOutputTypeForNonAdaptivePoolingOp<AtenOpT, tosaOp>(
inputTy, kernelSizeInts, strideInts, paddingInts, dilationArray,
ceilMode);
padArr[1] = padArr[1] + paddingInts[0];
padArr[3] = padArr[3] + paddingInts[1];
pad = rewriter.getDenseI64ArrayAttr(
{padArr[0], padArr[1], padArr[2], padArr[3]});
return success();
}
class ConvertAtenMaxPool2dOp
: public ConvertAtenPoolingBaseOp<AtenMaxPool2dOp, tosa::MaxPool2dOp> {
public:
using ConvertAtenPoolingBaseOp<AtenMaxPool2dOp,
tosa::MaxPool2dOp>::ConvertAtenPoolingBaseOp;
LogicalResult processInputs(AtenMaxPool2dOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter, Value &input,
DenseI64ArrayAttr &kernel,
DenseI64ArrayAttr &stride, DenseI64ArrayAttr &pad,
Type &outputTy) const override {
SmallVector<int64_t, 2> dilationArray;
if (!matchPattern(op.getDilation(),
m_TorchListOfConstantInts(dilationArray)))
return rewriter.notifyMatchFailure(
op, "Non-const dilation for pooling op unsupported.");
// TOSA pooling only supports unit dilation.
if (dilationArray[0] > 1 || dilationArray[1] > 1)
return rewriter.notifyMatchFailure(
op, "Cannot process non-unit pooling dilation.");
if (failed(getOutputTypeAndPoolingParameters<AtenMaxPool2dOp,
tosa::MaxPool2dOp>(
op, rewriter, adaptor.getSelf(), dilationArray, outputTy, kernel,
stride, pad)))
return rewriter.notifyMatchFailure(
op, "invalid pooling parameters or input type");
// Transpose to xHWC
input = ConvertAtenPoolingBaseOp<AtenMaxPool2dOp, tosa::MaxPool2dOp>::
transposePoolingInputToHwc(op, rewriter, adaptor.getSelf());
return success();
}
};
class ConvertAtenAvgPool2dOp
: public ConvertAtenPoolingBaseOp<AtenAvgPool2dOp, tosa::AvgPool2dOp> {
public:
using ConvertAtenPoolingBaseOp<AtenAvgPool2dOp,
tosa::AvgPool2dOp>::ConvertAtenPoolingBaseOp;
LogicalResult processInputs(AtenAvgPool2dOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter, Value &input,
DenseI64ArrayAttr &kernel,
DenseI64ArrayAttr &stride, DenseI64ArrayAttr &pad,
Type &outputTy) const override {
SmallVector<int64_t, 2> dilationArray{1, 1};
if (failed(getOutputTypeAndPoolingParameters<AtenAvgPool2dOp,
tosa::AvgPool2dOp>(
op, rewriter, adaptor.getSelf(), dilationArray, outputTy, kernel,
stride, pad)))
return rewriter.notifyMatchFailure(
op, "invalid pooling parameters or input type");
// Transpose to xHWC
input = ConvertAtenPoolingBaseOp<AtenAvgPool2dOp, tosa::AvgPool2dOp>::
transposePoolingInputToHwc(op, rewriter, adaptor.getSelf());
return success();
}
};
// Ref: Error checking based on the Torch to LinAlg lowering
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 rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat())
return rewriter.notifyMatchFailure(
op, "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?
// The layout arg should be either `none` or `0` i.e. strided.
if (!isa<Torch::NoneType>(op.getLayout().getType())) {
int64_t tensorLayout;
if (!matchPattern(op.getLayout(), m_TorchConstantInt(&tensorLayout)))
return rewriter.notifyMatchFailure(
op, "The layout arg should be either `none` or `0` i.e. strided.");
else if (tensorLayout != torch_upstream::Layout::Strided)
return rewriter.notifyMatchFailure(
op, "The layout arg should be either `none` or `0` i.e. strided.");
}
bool pinMemory;
if (!isa<Torch::NoneType>(op.getPinMemory().getType()) &&
(!matchPattern(op.getPinMemory(), m_TorchConstantBool(&pinMemory)) ||
pinMemory)) {
return rewriter.notifyMatchFailure(
op, "Unsupported pin_memory, should be either None or false");
}
SmallVector<int64_t> shape;
if (!matchPattern(op.getSize(), m_TorchListOfConstantInts(shape))) {
return rewriter.notifyMatchFailure(
op, "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 =
tosa::getConstTensor<int32_t>(rewriter, op, values, shape).value();
rewriter.replaceOpWithNewOp<tosa::CastOp>(op, outType, constOp);
return success();
}
};
template <typename AtenOpT>
class ConvertAtenFillScalarOp : 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 || !outType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only Tensor types with static shapes are currently supported");
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
Value constOp;
if (failed(torchScalarToTosaTensor(
rewriter, op, op.getValue(), constOp, outElemTy,
makeShapeTorchCompatible(outType.getShape()))))
return rewriter.notifyMatchFailure(
op, "Supplied value must be a Scalar constant");
rewriter.replaceOpWithNewOp<tosa::CastOp>(op, outType, constOp);
return success();
}
};
template <typename AtenOpT>
class ConvertAtenMaskedFillOp : 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 || !outType.hasStaticShape())
return rewriter.notifyMatchFailure(
op, "Only Tensor types with static shapes are currently supported");
Type outElemTy = outType.getElementType();
if (!outElemTy.isIntOrFloat()) {
return rewriter.notifyMatchFailure(
op, "Only floating-point or integer datatype legalization supported");
}
// Not a tensor type.
auto selfType = dyn_cast<TensorType>(adaptor.getSelf().getType());
if (!selfType || !outType.hasStaticShape())
return rewriter.notifyMatchFailure(
op,
"Only tensor types with static shapes input are currently supported");
auto maskType = dyn_cast<TensorType>(adaptor.getMask().getType());
if (!maskType)
return rewriter.notifyMatchFailure(
op, "Only tensor types mask are currently supported");
Value rhs = adaptor.getValue();
auto rhsType = dyn_cast<TensorType>(rhs.getType());
Value rhsAsTensor;
if (!rhsType) { // scalar
if (failed(torchScalarToTosaTensor(rewriter, op, op.getValue(),
rhsAsTensor, rhs.getType(), {})))
return rewriter.notifyMatchFailure(
op, "Currently only scalar constants are supported for "
"conversion in TOSA operation");
} else { // tensor
rhsType = dyn_cast<TensorType>(rhs.getType());
}
auto rhsTensor = rhsType ? rhs : rhsAsTensor;
auto rhsTensorType = dyn_cast<TensorType>(rhsTensor.getType());
if (rhsTensorType.getElementType() != outElemTy)
rhsTensor = rewriter.create<tosa::CastOp>(
op.getLoc(),
RankedTensorType::get(rhsTensorType.getShape(), outElemTy),
rhsTensor);
rewriter.replaceOpWithNewOp<tosa::SelectOp>(op, outType, adaptor.getMask(),
rhsTensor, adaptor.getSelf());
return success();
}
};
// Legalizes the torch.clone op.
template <typename AtenOpT>
class ConvertAtenCloneOp : public OpConversionPattern<AtenOpT> {
public:
using OpConversionPattern<AtenOpT>::OpConversionPattern;
using OpAdaptor = typename AtenOpT::Adaptor;
LogicalResult
matchAndRewrite(AtenOpT op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
int64_t memoryFormat;
if (!isa<Torch::NoneType>(op.getMemoryFormat().getType()) &&
(!matchPattern(op.getMemoryFormat(),
m_TorchConstantInt(&memoryFormat)) ||
(memoryFormat != torch_upstream::MemoryFormat::Contiguous &&
memoryFormat != torch_upstream::MemoryFormat::ChannelsLast))) {
return op.emitError(
"unimplemented: only contiguous and channels last memory "
"format is supported");
}
auto outType = OpConversionPattern<AtenOpT>::getTypeConverter()
->convertType(op.getType())
.template dyn_cast<TensorType>();
rewriter.replaceOpWithNewOp<tosa::CastOp>(op, outType, adaptor.getSelf());
return success();
}
};
template <>
LogicalResult ConvertAtenOp<AtenConstantPadNdOp>::matchAndRewrite(
AtenConstantPadNdOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op.getLoc();
Value self = adaptor.getSelf();
auto selfTy = cast<RankedTensorType>(self.getType());
auto selfElemTy = selfTy.getElementType();
int64_t rank = selfTy.getRank();
// START the code snippet from
// lib/Conversion/TorchToLinalg/TensorConstructors.cpp (see:
// ConvertAtenConstantPadNdOp) Pattern match against the op's original
// operands, because otherwise we will get the lowered version of the operands
// which is harder to pattern match.
SmallVector<int64_t> padInts;
if (!matchPattern(op.getPad(), m_TorchListOfConstantInts(padInts)))
return rewriter.notifyMatchFailure(op,
"only support constant int pad ranges");
uint64_t padRank = padInts.size() / 2;
if (padRank * 2 != padInts.size())
return rewriter.notifyMatchFailure(op, "pad range size is not even");
if (rank < 0 || padRank > (uint64_t)rank)
return rewriter.notifyMatchFailure(op, "padding exceeds tensor rank");
// Initialize low/high paddings with 0 for all the dims.
SmallVector<int64_t> lowPadding(/*Size=*/rank, /*Value=*/0);
SmallVector<int64_t> highPadding(/*Size=*/rank, /*Value=*/0);
// Add the requested padding - note op.pad() is highest dim first ordered
// pairs of low,high.
for (uint64_t i = 0; i < padRank; ++i) {
lowPadding[rank - i - 1] = padInts[i * 2];
highPadding[rank - i - 1] = padInts[i * 2 + 1];
}
// END the code snippet from
// lib/Conversion/TorchToLinalg/TensorConstructors.cpp (see:
// ConvertAtenConstantPadNdOp)
llvm::SmallVector<int64_t> translatePadsList;
for (unsigned int i = 0; i < rank; i++) {
translatePadsList.push_back(lowPadding[i]);
translatePadsList.push_back(highPadding[i]);
}
DenseElementsAttr paddingAttr = DenseIntElementsAttr::get(
RankedTensorType::get({rank, 2}, rewriter.getI64Type()),
translatePadsList);
Value padsList1 = rewriter.create<mlir::tosa::ConstOp>(
loc, paddingAttr.getType(), paddingAttr);
Value padValue = adaptor.getValue();
Operation *padOp = padValue.getDefiningOp();
padValue = padOp->getOperand(0);
Value padTensor;
if (failed(torchScalarToTosaTensor(rewriter, op.getOperation(), padValue,
padTensor, selfElemTy, {})))
return rewriter.notifyMatchFailure(
op, "Pad value needs to be a scalar constant for conversion to "
"TOSA pad operation");
rewriter.replaceOpWithNewOp<mlir::tosa::PadOp>(
op, getTypeConverter()->convertType(op.getType()), self, padsList1,
padTensor);
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenCatOp>::matchAndRewrite(
AtenCatOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
const TypeConverter *typeConverter = this->getTypeConverter();
auto outType =
cast<RankedTensorType>(typeConverter->convertType(op.getType()));
int64_t rank = outType.getRank();
int64_t dim;
if (!outType || !outType.hasStaticShape()) {
return rewriter.notifyMatchFailure(
op, "Only Tensor types with static shapes are currently supported");
}
Location loc = op.getLoc();
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim))) {
return rewriter.notifyMatchFailure(op,
"unimplemented: dim is not constant");
}
dim = toPositiveDim(dim, rank);
if (!isValidDim(dim, rank)) {
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
}
auto tensorList = op.getTensors();
SmallVector<Value> tensorsTorchType;
if (!getListConstructElements(tensorList, tensorsTorchType)) {
return rewriter.notifyMatchFailure(
op, "unimplemented: the tensor list is not from list construct");
}
auto builtinTensors =
getTypeConvertedValues(rewriter, loc, typeConverter, tensorsTorchType);
auto result = tosa::CreateOpAndInfer<tosa::ConcatOp>(
rewriter, loc, outType, builtinTensors, rewriter.getI32IntegerAttr(dim));
rewriter.replaceOp(op, result.getResult());
return success();
}
template <>
LogicalResult ConvertAtenOp<AtenSqrtOp>::matchAndRewrite(
AtenSqrtOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
// Converts AtenSqrtOp into pow(x, 0.5)
auto self = adaptor.getSelf();
auto selfTy = dyn_cast<TensorType>(self.getType());
if (!selfTy)
return rewriter.notifyMatchFailure(op,
"Only Tensor types supported in TOSA");
auto resultType = typeConverter->convertType(op.getType())
.template cast<RankedTensorType>();
auto elementType = resultType.getElementType();
if (isa<mlir::IntegerType>(selfTy.getElementType())) {
self = rewriter.createOrFold<tosa::CastOp>(
op->getLoc(), RankedTensorType::get(resultType.getShape(), elementType),
self);
}
auto oneHalf =
tosa::getConstTensor<float>(rewriter, op, 0.5, {}, elementType).value();
rewriter.replaceOpWithNewOp<tosa::PowOp>(op, resultType, self, oneHalf);
return success();
}
} // namespace
// -----------------------------------------------------------------------------
// TorchToTosa Pass
// -----------------------------------------------------------------------------
namespace {
class ConvertTorchToTosa : public ConvertTorchToTosaBase<ConvertTorchToTosa> {
public:
void getDependentDialects(DialectRegistry &registry) const override {
registry.insert<tosa::TosaDialect>();
registry.insert<tensor::TensorDialect>();
registry.insert<arith::ArithDialect>();
TorchConversion::getBackendTypeConversionDependentDialects(registry);
}
void runOnOperation() override {
MLIRContext *context = &getContext();
ConversionTarget target(*context);
target.addLegalDialect<tosa::TosaDialect, tensor::TensorDialect,
arith::ArithDialect>();
TypeConverter typeConverter;
typeConverter.addConversion([](Type type) { return type; });
TorchConversion::setupBackendTypeConversion(target, typeConverter);
// The following ops are never the primary reason why lowering fails.
// The backend contract only allows functions to return tensors thus there
// is always another op using them.
// When we have a chain of torch.constant.int followed by a unsupported
// torch op, we want the pass to mention the unsupported torch op
// in the error message.
target.addLegalOp<ConstantNoneOp>();
target.addLegalOp<ConstantBoolOp>();
target.addLegalOp<ConstantIntOp>();
target.addLegalOp<ConstantFloatOp>();
target.addLegalOp<ConstantStrOp>();
target.addLegalOp<ConstantDeviceOp>();
target.addLegalOp<PrimListConstructOp>();
target.addLegalOp<PrimTupleConstructOp>();
target.addIllegalDialect<Torch::TorchDialect>();
RewritePatternSet patterns(context);
#define INSERT_UNARY_FPONLY_PATTERN(AtenOp, TosaOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenUnaryFPOnlyOp<AtenOp, TosaOp>>(typeConverter, \
context);
INSERT_UNARY_FPONLY_PATTERN(AtenLogOp, tosa::LogOp)
INSERT_UNARY_FPONLY_PATTERN(AtenExpOp, tosa::ExpOp)
#undef INSERT_UNARY_FPONLY_PATTERN
#define INSERT_UNARY_PATTERN(AtenOp, TosaOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenUnaryOp<AtenOp, TosaOp>>(typeConverter, context);
INSERT_UNARY_PATTERN(AtenNegOp, tosa::NegateOp)
INSERT_UNARY_PATTERN(AtenFloorOp, tosa::FloorOp)
INSERT_UNARY_PATTERN(AtenRsqrtOp, tosa::RsqrtOp)
INSERT_UNARY_PATTERN(AtenBitwiseNotOp, tosa::BitwiseNotOp)
INSERT_UNARY_PATTERN(AtenCeilOp, tosa::CeilOp)
INSERT_UNARY_PATTERN(AtenReciprocalOp, tosa::ReciprocalOp)
#undef INSERT_UNARY_PATTERN
#define INSERT_BINARY_PATTERN(AtenOp, TosaOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenBinaryOp<AtenOp, TosaOp>>(typeConverter, context);
INSERT_BINARY_PATTERN(AtenMaximumOp, tosa::MaximumOp)
INSERT_BINARY_PATTERN(AtenMinimumOp, tosa::MinimumOp)
INSERT_BINARY_PATTERN(AtenLogicalOrOp, tosa::LogicalOrOp)
#undef INSERT_BINARY_PATTERN
#define INSERT_BINARY_ADDSUB_PATTERN(AtenOp, TosaOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenAddSubOp<AtenOp, TosaOp>>(typeConverter, context);
INSERT_BINARY_ADDSUB_PATTERN(AtenAddTensorOp, tosa::AddOp)
INSERT_BINARY_ADDSUB_PATTERN(AtenAddScalarOp, tosa::AddOp)
INSERT_BINARY_ADDSUB_PATTERN(AtenSubTensorOp, tosa::SubOp)
INSERT_BINARY_ADDSUB_PATTERN(AtenSubScalarOp, tosa::SubOp)
#undef INSERT_BINARY_ADDSUB_PATTERN
#define INSERT_BINARY_COMPARE_PATTERN(AtenOp, TosaOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenCompareOp<AtenOp, TosaOp>>(typeConverter, context);
INSERT_BINARY_COMPARE_PATTERN(AtenGtTensorOp, tosa::GreaterOp)
INSERT_BINARY_COMPARE_PATTERN(AtenGeScalarOp, tosa::GreaterEqualOp)
INSERT_BINARY_COMPARE_PATTERN(AtenGtScalarOp, tosa::GreaterOp)
INSERT_BINARY_COMPARE_PATTERN(AtenLtTensorOp, tosa::GreaterOp)
INSERT_BINARY_COMPARE_PATTERN(AtenLtScalarOp, tosa::GreaterOp)
INSERT_BINARY_COMPARE_PATTERN(AtenEqTensorOp, tosa::EqualOp)
INSERT_BINARY_COMPARE_PATTERN(AtenEqScalarOp, tosa::EqualOp)
INSERT_BINARY_COMPARE_PATTERN(AtenNeTensorOp, tosa::EqualOp)
INSERT_BINARY_COMPARE_PATTERN(AtenNeScalarOp, tosa::EqualOp)
INSERT_BINARY_COMPARE_PATTERN(AtenBitwiseAndTensorOp, tosa::BitwiseAndOp)
INSERT_BINARY_COMPARE_PATTERN(AtenBitwiseOrTensorOp, tosa::BitwiseOrOp)
INSERT_BINARY_COMPARE_PATTERN(AtenBitwiseXorTensorOp, tosa::BitwiseXorOp)
#undef INSERT_BINARY_COMPARE_PATTERN
#define INSERT_BINARY_MUL_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenMulOp<AtenOp>>(typeConverter, context);
INSERT_BINARY_MUL_PATTERN(AtenMulTensorOp);
INSERT_BINARY_MUL_PATTERN(AtenMulScalarOp);
#undef INSERT_BINARY_MUL_PATTERN
#define INSERT_BINARY_DIV_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenDivOp<AtenOp>>(typeConverter, context);
INSERT_BINARY_DIV_PATTERN(AtenDivTensorOp);
INSERT_BINARY_DIV_PATTERN(AtenDivScalarOp);
#undef INSERT_BINARY_DIV_PATTERN
#define INSERT_NDIMS_REDUCTION_OP_PATTERN(AtenOp, ConversionFunc) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenMultipleDimsReductionOp<AtenOp, ConversionFunc>>( \
typeConverter, context);
INSERT_NDIMS_REDUCTION_OP_PATTERN(AtenMeanDimOp,
mlir::tosa::convertReduceMeanOp)
INSERT_NDIMS_REDUCTION_OP_PATTERN(AtenSumDimIntListOp,
mlir::tosa::convertReduceSumOp)
INSERT_NDIMS_REDUCTION_OP_PATTERN(AtenLinalgVectorNormOp,
mlir::tosa::convertLinalgVectorNormOp)
#undef INSERT_NDIMS_REDUCTION_OP_PATTERN
#define INSERT_ONEDIM_REDUCTION_OP_PATTERN(AtenOp, ConversionFunc) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenOneDimReductionOp<AtenOp, ConversionFunc>>( \
typeConverter, context);
INSERT_ONEDIM_REDUCTION_OP_PATTERN(AtenAnyDimOp,
mlir::tosa::convertReduceAnyOp)
#undef INSERT_ONEDIM_REDUCTION_OP_PATTERN
#define INSERT_ALLDIMS_REDUCTION_OP_PATTERN(AtenOp, ConversionFunc) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenAllDimsReductionOp<AtenOp, ConversionFunc>>( \
typeConverter, context);
INSERT_ALLDIMS_REDUCTION_OP_PATTERN(AtenAllOp,
mlir::tosa::convertReduceAllOp)
INSERT_ALLDIMS_REDUCTION_OP_PATTERN(AtenAnyOp,
mlir::tosa::convertReduceAnyOp)
INSERT_ALLDIMS_REDUCTION_OP_PATTERN(AtenSumOp,
mlir::tosa::convertReduceSumOp)
#undef INSERT_ALLDIMS_REDUCTION_OP_PATTERN
#define INSERT_SQUEEZE_OP_PATTERN(AtenOp, TemplateForm) \
target.addIllegalOp<AtenOp>(); \
patterns.add<TemplateForm<AtenOp>>(typeConverter, context);
INSERT_SQUEEZE_OP_PATTERN(AtenSqueezeOp, ConvertAtenSqueezeAllDimsOp)
INSERT_SQUEEZE_OP_PATTERN(AtenSqueezeDimOp, ConvertAtenSqueezeOneDimOp)
#undef INSERT_SQUEEZE_OP_PATTERN
#define INSERT_MATMUL_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenMatMulOp<AtenOp>>(typeConverter, context);
INSERT_MATMUL_ATENOP_PATTERN(AtenMatmulOp);
#undef INSERT_MATMUL_ATEMOP_PATTERN
#define INSERT_MM_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenMmOp<AtenOp>>(typeConverter, context);
INSERT_MM_ATENOP_PATTERN(AtenMmOp);
INSERT_MM_ATENOP_PATTERN(AtenBmmOp);
#undef INSERT_MM_ATEMOP_PATTERN
#define INSERT_LINEAR_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenLinearOp<AtenOp>>(typeConverter, context);
INSERT_LINEAR_ATENOP_PATTERN(AtenLinearOp);
#undef INSERT_LINEAR_ATEMOP_PATTERN
#define INSERT_ADAPTIVE_POOLING_ATENOP_PATTERN(AtenOp, TosaOpT) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenAdaptivePoolingOp<AtenOp, TosaOpT>>(typeConverter, \
context);
INSERT_ADAPTIVE_POOLING_ATENOP_PATTERN(AtenAdaptiveAvgPool2dOp,
tosa::AvgPool2dOp);
#undef INSERT_ADAPTIVE_POOLING_ATEMOP_PATTERN
target.addIllegalOp<AtenMaxPool2dOp>();
patterns.add<ConvertAtenMaxPool2dOp>(typeConverter, context);
target.addIllegalOp<AtenAvgPool2dOp>();
patterns.add<ConvertAtenAvgPool2dOp>(typeConverter, context);
#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_FILL_SCALAR_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenFillScalarOp<AtenOp>>(typeConverter, context);
INSERT_FILL_SCALAR_PATTERN(AtenFill_ScalarOp);
#undef INSERT_FILL_SCALAR_PATTERN
#define INSERT_MASKED_FILL_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenMaskedFillOp<AtenOp>>(typeConverter, context);
INSERT_MASKED_FILL_PATTERN(AtenMaskedFillScalarOp);
INSERT_MASKED_FILL_PATTERN(AtenMaskedFillTensorOp);
#undef INSERT_MASKED_FILL_PATTERN
#define INSERT_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenOp<AtenOp>>(typeConverter, context);
INSERT_ATENOP_PATTERN(AtenTanhOp);
INSERT_ATENOP_PATTERN(AtenHardtanhBackwardOp);
INSERT_ATENOP_PATTERN(AtenSigmoidOp);
INSERT_ATENOP_PATTERN(AtenReluOp);
INSERT_ATENOP_PATTERN(AtenLeakyReluOp);
INSERT_ATENOP_PATTERN(AtenArgmaxOp);
INSERT_ATENOP_PATTERN(AtenPowTensorScalarOp);
INSERT_ATENOP_PATTERN(AtenRsubScalarOp);
INSERT_ATENOP_PATTERN(AtenConvolutionOp);
INSERT_ATENOP_PATTERN(ValueTensorLiteralOp);
INSERT_ATENOP_PATTERN(AtenReshapeOp);
INSERT_ATENOP_PATTERN(AtenBatchNormOp);
INSERT_ATENOP_PATTERN(AtenNativeLayerNormOp);
INSERT_ATENOP_PATTERN(AtenFlattenUsingIntsOp);
INSERT_ATENOP_PATTERN(AtenUnflattenIntOp);
INSERT_ATENOP_PATTERN(AtenPermuteOp);
INSERT_ATENOP_PATTERN(AtenLog2Op);
INSERT_ATENOP_PATTERN(AtenThresholdOp);
INSERT_ATENOP_PATTERN(AtenUnsqueezeOp);
INSERT_ATENOP_PATTERN(AtenContiguousOp);
INSERT_ATENOP_PATTERN(AtenDropoutOp);
INSERT_ATENOP_PATTERN(AtenViewOp);
INSERT_ATENOP_PATTERN(AtenGeluOp);
INSERT_ATENOP_PATTERN(AtenGeluBackwardOp);
INSERT_ATENOP_PATTERN(AtenEmbeddingOp);
INSERT_ATENOP_PATTERN(AtenTransposeIntOp);
INSERT_ATENOP_PATTERN(AtenMaxDimOp);
INSERT_ATENOP_PATTERN(AtenSliceTensorOp);
INSERT_ATENOP_PATTERN(AtenBroadcastToOp);
INSERT_ATENOP_PATTERN(AtenGatherOp);
INSERT_ATENOP_PATTERN(AtenIndexPutHackedTwinOp);
INSERT_ATENOP_PATTERN(AtenIndexTensorHackedTwinOp);
INSERT_ATENOP_PATTERN(AtenAbsOp);
INSERT_ATENOP_PATTERN(AtenWhereSelfOp);
INSERT_ATENOP_PATTERN(AtenLeTensorOp);
INSERT_ATENOP_PATTERN(AtenClampOp);
INSERT_ATENOP_PATTERN(AtenArangeStartStepOp);
INSERT_ATENOP_PATTERN(PrimNumToTensorScalarOp);
INSERT_ATENOP_PATTERN(AtenCopyOp);
INSERT_ATENOP_PATTERN(AtenToDtypeOp);
INSERT_ATENOP_PATTERN(AtenConstantPadNdOp);
INSERT_ATENOP_PATTERN(AtenRemainderScalarOp);
INSERT_ATENOP_PATTERN(AtenCatOp);
INSERT_ATENOP_PATTERN(AtenSqrtOp);
INSERT_ATENOP_PATTERN(AtenIscloseOp);
#undef INSERT_ATENOP_PATTERN
#define INSERT_CLONE_ATENOP_PATTERN(AtenOp) \
target.addIllegalOp<AtenOp>(); \
patterns.add<ConvertAtenCloneOp<AtenOp>>(typeConverter, context);
INSERT_CLONE_ATENOP_PATTERN(AtenCloneOp);
#undef INSERT_CLONE_ATENOP_PATTERN
if (failed(applyPartialConversion(getOperation(), target,
std::move(patterns))))
return signalPassFailure();
}
};
} // namespace
std::unique_ptr<OperationPass<func::FuncOp>>
mlir::torch::createConvertTorchToTosaPass() {
return std::make_unique<ConvertTorchToTosa>();
}
Reference in New Issue View Git Blame Copy Permalink