torch-mlir/lib/Conversion/TorchOnnxToTorch/DefaultDomainAtoF.cpp

2062 lines
92 KiB
C++

//===------------------------------------------------------------*- C++ -*-===//
//
// This file is licensed 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 "mlir/IR/DialectResourceBlobManager.h"
#include "torch-mlir/Conversion/TorchOnnxToTorch/Patterns.h"
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "llvm/Support/FormatVariadic.h"
using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::onnx_c;
class Endian {
private:
static constexpr uint32_t uint32_ = 0x01020304;
static constexpr uint8_t magic_ = (const uint8_t &)uint32_;
public:
static constexpr bool little = magic_ == 0x04;
static constexpr bool big = magic_ == 0x01;
static_assert(little || big, "Cannot determine endianness!");
private:
Endian() = delete;
};
static int64_t onnxDtypeIntToTorchDtypeInt(int64_t dtypeIntOnnx) {
// TODO: Add complete mapping.
// Where are the ONNX and PyTorch dtype enums defined?
// ONNX:
// https://github.com/shouxieai/tensorRT_Pro/blob/main/onnx/onnx-ml.proto
// PyTorch:
// https://github.com/llvm/torch-mlir/blob/main/include/torch-mlir/Dialect/Torch/Utils/TorchUpstream.h#L88
int64_t dtypeIntTorch = [dtypeIntOnnx]() {
switch (dtypeIntOnnx) {
case 1:
return 6; // float
case 2:
return 0; // uint8
case 3:
return 1; // int8
case 6:
return 3; // int32
case 7:
return 4; // int64
case 9:
return 11; // bool
case 10:
return 5; // half
case 11:
return 7; // double
case 16:
return 15; // bfloat16
default:
return -1; // No dtype
}
}();
return dtypeIntTorch;
}
static LogicalResult createTorchTransposeOp(ConversionPatternRewriter &rewriter,
Location loc, Value input,
int64_t dimA, int64_t dimB,
Value &transposed) {
Type transposedType;
if (failed(getTransposedType(input.getType().cast<Torch::BaseTensorType>(),
dimA, dimB, transposedType)))
return failure();
Value cstDimA = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(dimA));
Value cstDimB = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(dimB));
transposed = rewriter.create<Torch::AtenTransposeIntOp>(
loc, transposedType, input, cstDimA, cstDimB);
return success();
}
// Simple rewrites for the default domain.
// See: https://onnx.ai/onnx/operators/
// For operators that are effectively version invariant, we register with
// sinceVersion==1. We interpret this to include the following spec
// diffs that are irrelevant to this level of lowering:
// * Supported element types.
// * Limited broadcasting to full broadcasting support.
//
// There are a lot of spec revisions that basically generalized elementwise
// to be more normal and a direct translation vs a special case. This
// results in a lot of ONNX test cases that all reduce to the exact same
// thing here, so we simplify.
void mlir::torch::onnx_c::populateDefaultDomainAtoF(
OnnxCustomOpConversionPattern &patterns) {
patterns.onOp("Abs", 1,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAbsOp>(
binder.op, resultType, operand);
return success();
});
// Add became forward compatible with Torch in version 7.
patterns.onOp("Add", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
Value const1 = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), 1));
rewriter.replaceOpWithNewOp<Torch::AtenAddTensorOp>(
binder.op, resultType, lhs, rhs, const1);
return success();
});
// TODO: AffineGrid
patterns.onOp("And", 1,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenLogicalAndOp>(
binder.op, resultType, lhs, rhs);
return success();
});
patterns.onOp(
"ArgMax", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
bool keepDims;
int64_t axis;
bool selectLastIndex;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType) ||
binder.s64BoolAttr(keepDims, "keepdims", true) ||
binder.s64IntegerAttr(axis, "axis", 0) ||
binder.s64BoolAttr(selectLastIndex, "select_last_index", false))
return failure();
if (selectLastIndex) {
// TODO: Figure out how to support this case. Need to add a reverse
// or something.
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: select_last_index=true");
}
// ONNX allows negative axis.
if (axis < 0)
axis +=
cast<Torch::ValueTensorType>(operand.getType()).getSizes().size();
Value constAxis = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), axis));
Value constKeepDims = rewriter.create<Torch::ConstantBoolOp>(
binder.getLoc(), rewriter.getType<Torch::BoolType>(),
rewriter.getBoolAttr(keepDims));
rewriter.replaceOpWithNewOp<Torch::AtenArgmaxOp>(
binder.op, resultType, operand, constAxis, constKeepDims);
return success();
});
patterns.onOp(
"ArgMin", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
bool keepDims;
int64_t axis;
bool selectLastIndex;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType) ||
binder.s64BoolAttr(keepDims, "keepdims", true) ||
binder.s64IntegerAttr(axis, "axis", 0) ||
binder.s64BoolAttr(selectLastIndex, "select_last_index", false))
return failure();
if (selectLastIndex) {
// TODO: Figure out how to support this case. Need to add a reverse
// or something.
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: select_last_index=true");
}
// ONNX allows negative axis.
if (axis < 0)
axis +=
cast<Torch::ValueTensorType>(operand.getType()).getSizes().size();
Value constAxis = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), axis));
Value constKeepDims = rewriter.create<Torch::ConstantBoolOp>(
binder.getLoc(), rewriter.getType<Torch::BoolType>(),
rewriter.getBoolAttr(keepDims));
rewriter.replaceOpWithNewOp<Torch::AtenArgminOp>(
binder.op, resultType, operand, constAxis, constKeepDims);
return success();
});
patterns.onOp("Asin", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAsinOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Asinh", 9,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAsinhOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Atan", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAtanOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Atanh", 9,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAtanhOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Acos", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAcosOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Acosh", 9,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenAcoshOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("BatchNormalization", 15,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value input, weight, bias, runningMean, runningVar;
bool training;
float momentum, eps;
if (binder.s64BoolAttr(training, "training_mode", 0))
return failure();
if (training) {
// TODO: Add support for training = true
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: training = true");
}
if (binder.tensorOperandAtIndex(input, 0) ||
binder.tensorOperandAtIndex(weight, 1) ||
binder.tensorOperandAtIndex(bias, 2) ||
binder.tensorOperandAtIndex(runningMean, 3) ||
binder.tensorOperandAtIndex(runningVar, 4) ||
binder.f32FloatAttr(momentum, "momentum", 0.9f) ||
binder.f32FloatAttr(eps, "epsilon", 1e-05f) ||
binder.tensorResultType(resultType))
return failure();
Value cstFalse = rewriter.create<Torch::ConstantBoolOp>(
binder.getLoc(), false);
Value cstMomentum = rewriter.create<Torch::ConstantFloatOp>(
binder.getLoc(), rewriter.getF64FloatAttr(momentum));
Value cstEps = rewriter.create<Torch::ConstantFloatOp>(
binder.getLoc(), rewriter.getF64FloatAttr(eps));
rewriter.replaceOpWithNewOp<Torch::AtenBatchNormOp>(
binder.op, resultType, input, weight, bias, runningMean,
runningVar, /*training=*/cstFalse, cstMomentum, cstEps,
/*cudnn_enabled=*/cstFalse);
return success();
});
patterns.onOp(
"AveragePool", 11,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
std::string autoPad;
SmallVector<int64_t> dilation;
if (binder.customOpNameStringAttr(autoPad, "auto_pad", "NOTSET"))
return failure();
if (autoPad != "NOTSET") {
// TODO: Add support for `auto_pad` != "NOTSET"
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: auto_pad != NOTSET");
}
if (binder.s64IntegerArrayAttr(dilation, "dilations", {})) {
return failure();
}
if (dilation.size() > 0) {
return rewriter.notifyMatchFailure(
binder.op, "dilation is not supported by torch.aten.avgpool op");
}
Torch::ValueTensorType resultType;
Value operand;
bool ceilMode, countIncludePad;
if (binder.tensorOperand(operand) ||
binder.s64BoolAttr(ceilMode, "ceil_mode", false) ||
binder.s64BoolAttr(countIncludePad, "count_include_pad", false) ||
binder.tensorResultType(resultType))
return failure();
// Determine the rank of input tensor.
std::optional<unsigned> maybeRank = Torch::getTensorRank(operand);
if (!maybeRank)
return rewriter.notifyMatchFailure(binder.op,
"Unimplemented: unranked tensor");
unsigned rank = *maybeRank;
SmallVector<int64_t> kernel, padding, strides;
if (binder.s64IntegerArrayAttr(kernel, "kernel_shape", {})) {
return failure();
}
if (kernel.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op, "kernel list size does not match the number of axes");
}
SmallVector<int64_t> defaultPadding(2 * (rank - 2), 0);
if (binder.s64IntegerArrayAttr(padding, "pads", defaultPadding)) {
return failure();
}
if (padding.size() != 2 * (rank - 2)) {
return rewriter.notifyMatchFailure(
binder.op,
"padding list size does not match twice the number of axes");
}
if (binder.s64IntegerArrayAttr(
strides, "strides", llvm::SmallVector<int64_t>(rank - 2, 1))) {
return failure();
}
if (strides.size() != 1 && strides.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op, "strides list size does not match the number of axes");
}
SmallVector<Value> cstKernel, cstPadding, cstStrides;
for (int64_t i : kernel) {
cstKernel.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
for (int64_t i : padding) {
cstPadding.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
for (int64_t i : strides) {
cstStrides.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
Value kernelSizeList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstKernel);
Value paddingList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstPadding);
Value stridesList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstStrides);
Value cstCeilMode =
rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), ceilMode);
Value cstCountIncludePad = rewriter.create<Torch::ConstantBoolOp>(
binder.getLoc(), countIncludePad);
Value cstNone = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
if (rank == 3) {
rewriter.replaceOpWithNewOp<Torch::AtenAvgPool1dOp>(
binder.op, resultType, operand, kernelSizeList, stridesList,
paddingList, cstCeilMode, cstCountIncludePad);
return success();
} else if (rank == 4) {
rewriter.replaceOpWithNewOp<Torch::AtenAvgPool2dOp>(
binder.op, resultType, operand, kernelSizeList, stridesList,
paddingList, cstCeilMode, cstCountIncludePad,
/*divisor_override=*/cstNone);
return success();
} else if (rank == 5) {
rewriter.replaceOpWithNewOp<Torch::AtenAvgPool3dOp>(
binder.op, resultType, operand, kernelSizeList, stridesList,
paddingList, cstCeilMode, cstCountIncludePad,
/*divisor_override=*/cstNone);
return success();
}
return failure();
});
patterns.onOp(
"Bernoulli", 15,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value input;
int64_t dtypeIntOnnx, dtypeIntTorch;
if (binder.tensorOperand(input) ||
binder.s64IntegerAttr(dtypeIntOnnx, "dtype", -1) ||
binder.tensorResultType(resultType))
return failure();
SmallString<64> name("torch.onnx.");
name.append("seed");
auto attr = binder.op->getAttr(name);
if (attr) {
return rewriter.notifyMatchFailure(
binder.op,
"unimplemented: support not present for seed attribute");
}
Value none = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
Value bernoulli = rewriter.create<Torch::AtenBernoulliOp>(
binder.getLoc(), input.getType(), input, /*generator=*/none);
if (dtypeIntOnnx == -1) {
// True, if dtype attribute value is not present.
rewriter.replaceOp(binder.op, bernoulli);
return success();
}
dtypeIntTorch = onnxDtypeIntToTorchDtypeInt(dtypeIntOnnx);
if (dtypeIntTorch == -1) {
return rewriter.notifyMatchFailure(
binder.op,
"unimplemented support for the given dtype conversion");
}
Value constDtype = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64),
dtypeIntTorch));
Value cstFalse =
rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), false);
rewriter.replaceOpWithNewOp<Torch::AtenToDtypeOp>(
binder.op, resultType, bernoulli, constDtype,
/*non_blocking=*/cstFalse, /*copy=*/cstFalse,
/*memory_format=*/none);
return success();
});
patterns.onOp(
"BitShift", 11, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
std::string direction;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType) ||
binder.customOpNameStringAttr(direction, "direction", ""))
return failure();
if (direction == "LEFT") {
rewriter.replaceOpWithNewOp<Torch::AtenBitwiseLeftShiftTensorOp>(
binder.op, resultType, lhs, rhs);
} else {
rewriter.replaceOpWithNewOp<Torch::AtenBitwiseRightShiftTensorOp>(
binder.op, resultType, lhs, rhs);
}
return success();
});
patterns.onOp("BitwiseAnd", 18,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
std::string direction;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenBitwiseAndTensorOp>(
binder.op, resultType, lhs, rhs);
return success();
});
patterns.onOp("BitwiseOr", 18,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
std::string direction;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenBitwiseOrTensorOp>(
binder.op, resultType, lhs, rhs);
return success();
});
patterns.onOp("BitwiseNot", 18,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenBitwiseNotOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("BitwiseXor", 18,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
std::string direction;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenBitwiseXorTensorOp>(
binder.op, resultType, lhs, rhs);
return success();
});
patterns.onOp(
"Cast", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
int64_t dtypeIntOnnx, dtypeIntTorch;
if (binder.tensorOperand(operand) ||
binder.s64IntegerAttr(dtypeIntOnnx, "to") ||
binder.tensorResultType(resultType))
return failure();
dtypeIntTorch = onnxDtypeIntToTorchDtypeInt(dtypeIntOnnx);
if (dtypeIntTorch == -1) {
auto message = llvm::formatv("unimplemented support for the given "
"dtype conversion (onnx 'type' = {0})",
dtypeIntOnnx);
auto y = rewriter.notifyMatchFailure(binder.op, message);
return y;
}
Value constDtype = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64),
dtypeIntTorch));
Value none = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
Value cstFalse =
rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), false);
rewriter.replaceOpWithNewOp<Torch::AtenToDtypeOp>(
binder.op, resultType, operand, constDtype,
/*non_blocking=*/cstFalse, /*copy=*/cstFalse,
/*memory_format=*/none);
return success();
});
patterns.onOp(
"CastLike", 15, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value input, target;
if (binder.tensorOperands(input, target) ||
binder.tensorResultType(resultType))
return failure();
// TODO: Add support to handle the `saturate` attribute.
// Ignoring it right now, since it's only using during the float8
// conversions which are not supported in Torch-MLIR right now.
Torch::ValueTensorType targetTy =
target.getType().cast<Torch::ValueTensorType>();
if (!targetTy.hasDtype()) {
return rewriter.notifyMatchFailure(binder.op,
"target tensor must have a dtype");
}
Type targetDtype = targetTy.getDtype();
Value constDtype = Torch::getDtypeIntValueForType(
rewriter, binder.getLoc(), targetDtype);
Value none = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
Value cstFalse =
rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), false);
rewriter.replaceOpWithNewOp<Torch::AtenToDtypeOp>(
binder.op, resultType, input, constDtype,
/*non_blocking=*/cstFalse, /*copy=*/cstFalse,
/*memory_format=*/none);
return success();
});
patterns.onOp("Ceil", 13,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenCeilOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp(
"Celu", 12, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
float alpha;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType) ||
binder.f32FloatAttr(alpha, "alpha", 1.0f))
return failure();
// exp(x/alpha)
Value constAlpha = rewriter.create<Torch::ConstantFloatOp>(
binder.getLoc(), rewriter.getType<Torch::FloatType>(),
rewriter.getF64FloatAttr(alpha));
Value xDivAlpha = rewriter.create<Torch::AtenDivScalarOp>(
binder.getLoc(), resultType, operand, constAlpha);
Value expXDivAlpha = rewriter.create<Torch::AtenExpOp>(
binder.getLoc(), resultType, xDivAlpha);
// alpha * (exp(x/alpha) - 1)
Value constantOne = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(1));
Value subOne = rewriter.create<Torch::AtenSubScalarOp>(
binder.getLoc(), resultType, expXDivAlpha, constantOne,
constantOne);
Value mulAlpha = rewriter.create<Torch::AtenMulScalarOp>(
binder.getLoc(), resultType, subOne, constAlpha);
Value constantZero = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(0));
Value zeroTensor = createRank0Tensor(rewriter, binder.getLoc(),
resultType, constantZero);
// min(0, alpha * (exp(x/alpha) - 1))
Value minExpression = rewriter.create<Torch::AtenMinimumOp>(
binder.getLoc(), resultType, zeroTensor, mulAlpha);
// max(0, x)
Value maxExpression = rewriter.create<Torch::AtenMaximumOp>(
binder.getLoc(), resultType, zeroTensor, operand);
// max(0,x) + min(0, alpha * (exp(x/alpha) - 1))
rewriter.replaceOpWithNewOp<Torch::AtenAddTensorOp>(
binder.op, resultType, maxExpression, minExpression, constantOne);
return success();
});
patterns.onOp(
"Clip", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
// https://onnx.ai/onnx/operators/onnx__Clip.html
// Inputs and outputs must be tensors.
Value source;
Torch::ValueTensorType resultType;
if (binder.tensorOperandAtIndex(source, 0) ||
binder.tensorResultType(resultType)) {
return failure();
}
// Min and max can be args (version 11+) or attributes (version 6-).
// They default to numeric_limits::lowest() and numeric_limits::max().
Value min;
Value max;
if (binder.op->getNumOperands() >= 2)
min = binder.op->getOperand(1);
if (binder.op->getNumOperands() == 3)
max = binder.op->getOperand(2);
// Note: attribute versions of the op only support float types.
auto resultDtype = resultType.getDtype();
if (!min && binder.op->hasAttr("torch.onnx.min")) {
float minValue;
if (binder.f32FloatAttr(minValue, "min",
std::numeric_limits<float>::lowest()))
return failure();
auto minSplatAttr = SplatElementsAttr::get(
resultType.toBuiltinTensor().clone(resultDtype),
rewriter.getFloatAttr(resultDtype, minValue));
min = rewriter.create<Torch::ValueTensorLiteralOp>(
binder.getLoc(), resultType, minSplatAttr);
}
if (!max && binder.op->hasAttr("torch.onnx.max")) {
float maxValue;
if (binder.f32FloatAttr(maxValue, "max",
std::numeric_limits<float>::max()))
return failure();
auto maxSplatAttr = SplatElementsAttr::get(
resultType.toBuiltinTensor().clone(resultDtype),
rewriter.getFloatAttr(resultDtype, maxValue));
max = rewriter.create<Torch::ValueTensorLiteralOp>(
binder.getLoc(), resultType, maxSplatAttr);
}
if (!min && !max) {
// Cliping with no limits is a no-op.
rewriter.replaceOp(binder.op, source);
return success();
}
if (!max) {
rewriter.replaceOpWithNewOp<Torch::AtenClampMinTensorOp>(
binder.op, resultType, source, min);
return success();
}
rewriter.replaceOpWithNewOp<Torch::AtenClampTensorOp>(
binder.op, resultType, source, min, max);
return success();
});
patterns.onOp(
"Compress", 9, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand, conditionTensor;
int64_t axis;
if (binder.tensorOperands(operand, conditionTensor) ||
binder.s64IntegerAttr(axis, "axis", INT64_MAX) ||
binder.tensorResultType(resultType))
return failure();
auto shapeSizes =
dyn_cast<Torch::ValueTensorType>(operand.getType()).getSizes();
auto resultSizes = resultType.getSizes();
// flatten input tensor if using default axis
if (axis == INT64_MAX) {
SmallVector<int64_t> nonzeroShape = {resultSizes[0]};
auto dtype =
dyn_cast<Torch::ValueTensorType>(conditionTensor.getType())
.getDtype();
auto nonzeroType =
rewriter.getType<Torch::ValueTensorType>(nonzeroShape, dtype);
Value indexVal = rewriter.create<Torch::AtenNonzeroOp>(
binder.getLoc(), nonzeroType, conditionTensor);
Value cstZero = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(0));
Value cstNegOne = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(-1));
int64_t numElements = 1;
for (auto i : shapeSizes) {
numElements *= i;
}
SmallVector<int64_t> flattenShape = {numElements};
auto flattenType = rewriter.getType<Torch::ValueTensorType>(
flattenShape, resultType.getDtype());
Value flattenTensor = rewriter.create<Torch::AtenFlattenUsingIntsOp>(
binder.getLoc(), flattenType, operand, cstZero, cstNegOne);
rewriter.replaceOpWithNewOp<Torch::AtenIndexSelectOp>(
binder.op, resultType, flattenTensor, cstZero, indexVal);
return success();
}
// Negative axis value means counting dimensions from the back
if (axis < 0)
axis += shapeSizes.size();
SmallVector<int64_t> nonzeroShape = {resultSizes[axis]};
auto dtype = dyn_cast<Torch::ValueTensorType>(conditionTensor.getType())
.getDtype();
auto nonzeroType =
rewriter.getType<Torch::ValueTensorType>(nonzeroShape, dtype);
Value indexVal = rewriter.create<Torch::AtenNonzeroOp>(
binder.getLoc(), nonzeroType, conditionTensor);
Value dimVal = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(axis));
rewriter.replaceOpWithNewOp<Torch::AtenIndexSelectOp>(
binder.op, resultType, operand, dimVal, indexVal);
return success();
});
patterns.onOp(
"Concat", 13, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
SmallVector<Value> tensors;
int64_t dim;
if (binder.tensorOperands(tensors, binder.op->getNumOperands()) ||
binder.s64IntegerAttr(dim, "axis", 0) ||
binder.tensorResultType(resultType))
return failure();
Type listElemType =
tensors[0]
.getType()
.cast<Torch::BaseTensorType>()
.getWithSizesAndDtype(/*optionalSizes=*/std::nullopt,
/*optionalDtype=*/nullptr);
Type listType = Torch::ListType::get(listElemType);
Value tensorList = rewriter.create<Torch::PrimListConstructOp>(
binder.op->getLoc(), listType, tensors);
Value cstDim = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(dim));
rewriter.replaceOpWithNewOp<Torch::AtenCatOp>(binder.op, resultType,
tensorList, cstDim);
return success();
});
patterns.onOp(
"Constant", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
if (binder.tensorResultType(resultType))
return failure();
auto dtype = resultType.getDtype();
float floatValue;
if (binder.op->hasAttr("torch.onnx.value_float") &&
!binder.f32FloatAttr(floatValue, "value_float", 0.0)) {
auto splatAttr =
SplatElementsAttr::get(resultType.toBuiltinTensor().clone(dtype),
rewriter.getFloatAttr(dtype, floatValue));
rewriter.replaceOpWithNewOp<Torch::ValueTensorLiteralOp>(
binder.op, resultType, splatAttr);
return success();
}
int64_t intValue;
if (binder.op->hasAttr("torch.onnx.value_int") &&
!binder.s64IntegerAttr(intValue, "value_int", 0)) {
auto splatAttr =
SplatElementsAttr::get(resultType.toBuiltinTensor().clone(dtype),
rewriter.getIntegerAttr(dtype, intValue));
rewriter.replaceOpWithNewOp<Torch::ValueTensorLiteralOp>(
binder.op, resultType, splatAttr);
return success();
}
if (DenseResourceElementsAttr attr =
binder.op->getAttr("torch.onnx.value")
.dyn_cast_or_null<DenseResourceElementsAttr>()) {
// Bytes are stored in little endian order. Big endian support will
// require swizzling.
if (!Endian::little) {
binder.op->emitError(
"unimplemented: importing on big endian systems");
return failure();
}
auto ty = cast<ShapedType>(attr.getType());
ElementsAttr denseAttr;
auto ptr = attr.getRawHandle().getBlob()->getData();
if (cast<ShapedType>(attr.getType()).getElementType().isInteger(1)) {
llvm::SmallVector<APInt> newContents;
for (auto val : ptr) {
APInt apval(1, val);
newContents.push_back(apval);
}
denseAttr = DenseElementsAttr::get(ty, newContents);
} else {
denseAttr = DenseElementsAttr::getFromRawBuffer(ty, ptr);
}
rewriter.replaceOpWithNewOp<Torch::ValueTensorLiteralOp>(
binder.op, resultType, denseAttr);
return success();
}
if (ElementsAttr attr = binder.op->getAttr("torch.onnx.value")
.dyn_cast_or_null<ElementsAttr>()) {
rewriter.replaceOpWithNewOp<Torch::ValueTensorLiteralOp>(
binder.op, resultType, attr);
return success();
}
llvm::SmallVector<int64_t> intValues;
if (!binder.s64IntegerArrayAttr(intValues, "value_ints", {}) &&
!intValues.empty()) {
llvm::SmallVector<APInt> apValues;
for (auto intVal : intValues) {
apValues.push_back(APInt(dtype.getIntOrFloatBitWidth(), intVal));
}
auto attr = DenseElementsAttr::get(
resultType.toBuiltinTensor().clone(dtype), apValues);
rewriter.replaceOpWithNewOp<Torch::ValueTensorLiteralOp>(
binder.op, resultType, attr);
return success();
}
return failure();
});
patterns.onOp(
"Conv", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
std::string autoPad;
if (binder.customOpNameStringAttr(autoPad, "auto_pad", "NOTSET"))
return failure();
if (autoPad != "NOTSET") {
// TODO: Add support for `auto_pad` != "NOTSET"
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: auto_pad != NOTSET");
}
Torch::ValueTensorType resultType;
Value input, weight;
int64_t group;
if (binder.tensorOperandAtIndex(input, 0) ||
binder.tensorOperandAtIndex(weight, 1) ||
binder.s64IntegerAttr(group, "group", 1) ||
binder.tensorResultType(resultType))
return failure();
auto weightTensorType = weight.getType().cast<Torch::ValueTensorType>();
if (!weightTensorType || !weightTensorType.hasSizes()) {
return rewriter.notifyMatchFailure(
binder.op, "Expected weight type having sizes");
}
ArrayRef<int64_t> weightShape = weightTensorType.getSizes();
SmallVector<int64_t> kernelShape;
if (binder.s64IntegerArrayAttr(kernelShape, "kernel_shape", {}))
return failure();
if (kernelShape.size()) {
if (kernelShape.size() != weightShape.size() - 2) {
return rewriter.notifyMatchFailure(
binder.op,
"unsupported conversion: kernel_shape list size should have "
"number of values equal to weight_rank - 2");
} else {
for (unsigned i = 0; i < kernelShape.size(); i++) {
if (weightShape[i + 2] != kernelShape[i]) {
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: kernel_shape value "
"should be equal to the weight tensor shape");
}
}
}
}
// Determine the rank of input tensor.
std::optional<unsigned> maybeRank = Torch::getTensorRank(input);
if (!maybeRank)
return rewriter.notifyMatchFailure(binder.op,
"Unimplemented: unranked tensor");
unsigned rank = *maybeRank;
SmallVector<int64_t> padding, strides, dilations;
SmallVector<int64_t> defaultPadding, defaultStrides, defaultDilations;
for (unsigned i = 0; i < rank - 2; i++) {
defaultPadding.push_back(0);
defaultStrides.push_back(1);
defaultDilations.push_back(1);
}
// Padding for the beginning and ending along each spatial axis, it can
// take any value greater than or equal to 0. The value represent the
// number of pixels added to the beginning and end part of the
// corresponding axis. pads format should be as follow [x1_begin,
// x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added
// at the beginning of axis i and xi_end, the number of pixels added at
// the end of axis i.
if (binder.s64IntegerArrayAttr(padding, "pads", defaultPadding)) {
return failure();
}
if (padding.size() != rank - 2 && padding.size() != 2 * (rank - 2)) {
return rewriter.notifyMatchFailure(
binder.op, "padding list size does not match the number of axes");
}
if (binder.s64IntegerArrayAttr(dilations, "dilations",
defaultDilations)) {
return failure();
}
if (dilations.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op,
"dilations list size does not match the number of axes");
}
if (binder.s64IntegerArrayAttr(strides, "strides", defaultStrides)) {
return failure();
}
if (strides.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op, "strides list size does not match the number of axes");
}
SmallVector<Value> cstPadding, cstStrides, cstDilations,
cstOutputPadding;
if (padding.size() != 2 * (rank - 2)) {
for (int64_t i : padding) {
cstPadding.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
} else {
for (unsigned i = 0; i < padding.size() / 2; i++) {
if (padding[i] != padding[i + (padding.size() / 2)]) {
// TODO: Add support for different padding values for the
// beginning and ending along each spatial axis
return rewriter.notifyMatchFailure(
binder.op,
"unsupported conversion: padding values for the beginning "
"and ending along each spatial axis must be equal");
}
cstPadding.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(padding[i])));
}
}
for (int64_t i : dilations) {
cstDilations.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
for (int64_t i : strides) {
cstStrides.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
Value cstZero = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(0));
cstOutputPadding = {cstZero, cstZero};
Value paddingList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstPadding);
Value dilationsList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstDilations);
Value stridesList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstStrides);
Value outputPaddingList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstOutputPadding);
Value transposed =
rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), false);
Value bias;
if (binder.op->getNumOperands() == 3) {
if (binder.tensorOperandAtIndex(bias, 2)) {
return failure();
}
} else {
bias = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
}
Value cstGroup = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(group));
rewriter.replaceOpWithNewOp<Torch::AtenConvolutionOp>(
binder.op, resultType, input, weight, bias, stridesList,
paddingList, dilationsList, transposed, outputPaddingList,
cstGroup);
return success();
});
patterns.onOp(
"ConvTranspose", 11,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
std::string autoPad;
if (binder.customOpNameStringAttr(autoPad, "auto_pad", "NOTSET"))
return failure();
if (autoPad != "NOTSET") {
// TODO: Add support for `auto_pad` != "NOTSET"
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: auto_pad != NOTSET");
}
SmallVector<int64_t> outputShape;
if (binder.s64IntegerArrayAttr(outputShape, "output_shape", {}))
return failure();
if (outputShape.size()) {
// TODO: Add support for non-None output_shape value.
return rewriter.notifyMatchFailure(
binder.op,
"unsupported conversion: output_shape should be absent");
}
Torch::ValueTensorType resultType;
Value input, weight;
int64_t group;
if (binder.tensorOperandAtIndex(input, 0) ||
binder.tensorOperandAtIndex(weight, 1) ||
binder.s64IntegerAttr(group, "group", 1) ||
binder.tensorResultType(resultType))
return failure();
auto weightTensorType = weight.getType().cast<Torch::ValueTensorType>();
if (!weightTensorType || !weightTensorType.hasSizes()) {
return rewriter.notifyMatchFailure(
binder.op, "Expected weight type having sizes");
}
ArrayRef<int64_t> weightShape = weightTensorType.getSizes();
SmallVector<int64_t> kernelShape;
if (binder.s64IntegerArrayAttr(kernelShape, "kernel_shape", {}))
return failure();
if (kernelShape.size()) {
if (kernelShape.size() != weightShape.size() - 2) {
return rewriter.notifyMatchFailure(
binder.op,
"unsupported conversion: kernel_shape list size should have "
"number of values equal to weight_rank - 2");
} else {
for (unsigned i = 0; i < kernelShape.size(); i++) {
if (weightShape[i + 2] != kernelShape[i]) {
return rewriter.notifyMatchFailure(
binder.op, "unsupported conversion: kernel_shape value "
"should be equal to the weight tensor shape");
}
}
}
}
// Determine the rank of input tensor.
std::optional<unsigned> maybeRank = Torch::getTensorRank(input);
if (!maybeRank)
return rewriter.notifyMatchFailure(binder.op,
"Unimplemented: unranked tensor");
unsigned rank = *maybeRank;
SmallVector<int64_t> padding, strides, dilations, outputPadding;
SmallVector<int64_t> defaultPadding, defaultStrides, defaultDilations,
defaultOutputPadding;
for (unsigned i = 0; i < rank - 2; i++) {
defaultPadding.push_back(0);
defaultStrides.push_back(1);
defaultDilations.push_back(1);
defaultOutputPadding.push_back(0);
}
// Padding for the beginning and ending along each spatial axis, it can
// take any value greater than or equal to 0. The value represent the
// number of pixels added to the beginning and end part of the
// corresponding axis. pads format should be as follow [x1_begin,
// x2_begin…x1_end, x2_end,…], where xi_begin the number of pixels added
// at the beginning of axis i and xi_end, the number of pixels added at
// the end of axis i.
if (binder.s64IntegerArrayAttr(padding, "pads", defaultPadding)) {
return failure();
}
if (padding.size() != rank - 2 && padding.size() != 2 * (rank - 2)) {
return rewriter.notifyMatchFailure(
binder.op, "padding list size does not match the number of axes");
}
if (binder.s64IntegerArrayAttr(dilations, "dilations",
defaultDilations)) {
return failure();
}
if (dilations.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op,
"dilations list size does not match the number of axes");
}
if (binder.s64IntegerArrayAttr(strides, "strides", defaultStrides)) {
return failure();
}
if (strides.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op, "strides list size does not match the number of axes");
}
if (binder.s64IntegerArrayAttr(outputPadding, "output_padding",
defaultOutputPadding)) {
return failure();
}
if (outputPadding.size() != rank - 2) {
return rewriter.notifyMatchFailure(
binder.op,
"output_padding list size does not match the number of axes");
}
SmallVector<Value> cstPadding, cstStrides, cstDilations,
cstOutputPadding;
if (padding.size() != 2 * (rank - 2)) {
for (int64_t i : padding) {
cstPadding.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
} else {
for (unsigned i = 0; i < padding.size() / 2; i++) {
if (padding[i] != padding[i + (padding.size() / 2)]) {
// TODO: Add support for different padding values for the
// beginning and ending along each spatial axis
return rewriter.notifyMatchFailure(
binder.op,
"unsupported conversion: padding values for the beginning "
"and ending along each spatial axis must be equal");
}
cstPadding.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(padding[i])));
}
}
for (int64_t i : dilations) {
cstDilations.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
for (int64_t i : strides) {
cstStrides.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
for (int64_t i : outputPadding) {
cstOutputPadding.push_back(rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(i)));
}
Value paddingList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstPadding);
Value dilationsList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstDilations);
Value stridesList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstStrides);
Value outputPaddingList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
cstOutputPadding);
Value transposed =
rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), true);
Value bias;
if (binder.op->getNumOperands() == 3) {
if (binder.tensorOperandAtIndex(bias, 2)) {
return failure();
}
} else {
bias = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
}
Value cstGroup = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(group));
rewriter.replaceOpWithNewOp<Torch::AtenConvolutionOp>(
binder.op, resultType, input, weight, bias, stridesList,
paddingList, dilationsList, transposed, outputPaddingList,
cstGroup);
return success();
});
patterns.onOp("Cos", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenCosOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Cosh", 9,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenCoshOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp(
"CumSum", 11, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Location loc = binder.getLoc();
Torch::ValueTensorType resultType;
Value operand;
Value axisTensor;
if (binder.tensorOperands(operand, axisTensor) ||
binder.tensorResultType(resultType))
return failure();
int64_t exclusive;
int64_t reverse;
// if bind succeeds and either is set, fail because not implemented
if (!binder.s64IntegerAttr(exclusive, "exclusive", 0))
if (exclusive != 0)
return rewriter.notifyMatchFailure(
binder.op, "unsupported onnx.CumSum conversion: exclusive");
if (!binder.s64IntegerAttr(reverse, "reverse", 0))
if (reverse != 0)
return rewriter.notifyMatchFailure(
binder.op, "unsupported onnx.CumSum conversion: reverse");
// deal with neg axis: if (axis < 0) axis += rank
int64_t rank =
cast<Torch::ValueTensorType>(operand.getType()).getSizes().size();
Value rankVal = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), rank));
Value zero = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(0));
Value axisScalar = rewriter.create<Torch::AtenItemOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(), axisTensor);
Value isNegative = rewriter.create<Torch::AtenLtIntOp>(
binder.getLoc(), axisScalar, zero);
isNegative =
rewriter.create<Torch::AtenIntBoolOp>(binder.getLoc(), isNegative);
Value finalOffset = rewriter.create<Torch::AtenMulIntOp>(
binder.getLoc(), isNegative, rankVal);
Value dim = rewriter.create<Torch::AtenAddIntOp>(
binder.getLoc(), axisScalar, finalOffset);
Torch::BaseTensorType resultTensorType =
cast<Torch::BaseTensorType>(resultType);
if (!resultTensorType.hasDtype()) {
return rewriter.notifyMatchFailure(
binder.op, "expected result type to have a dtype");
}
// resultTensorType.print(llvm::outs());
Value none = rewriter.create<Torch::ConstantNoneOp>(loc);
rewriter.replaceOpWithNewOp<Torch::AtenCumsumOp>(binder.op, resultType,
operand, dim, none);
return success();
});
patterns.onOp(
"DepthToSpace", 1,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value input;
int64_t blockSize;
std::string mode;
if (binder.tensorOperand(input) ||
binder.s64IntegerAttr(blockSize, "blocksize") ||
binder.customOpNameStringAttr(mode, "mode", "DCR") ||
binder.tensorResultType(resultType))
return failure();
auto inputTy = input.getType().dyn_cast<Torch::BaseTensorType>();
if (!inputTy || !inputTy.hasSizes()) {
return rewriter.notifyMatchFailure(
binder.op, "Expected input type having sizes");
}
SmallVector<int64_t> inputSizes{inputTy.getSizes()};
if (inputSizes.size() != 4) {
return rewriter.notifyMatchFailure(binder.op,
"Expected input rank to be 4");
}
Value b = rewriter.create<Torch::AtenSizeIntOp>(
binder.getLoc(), input,
rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(0)));
Value c = rewriter.create<Torch::AtenSizeIntOp>(
binder.getLoc(), input,
rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(1)));
Value h = rewriter.create<Torch::AtenSizeIntOp>(
binder.getLoc(), input,
rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(2)));
Value w = rewriter.create<Torch::AtenSizeIntOp>(
binder.getLoc(), input,
rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(3)));
Value cstBlockSize = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(blockSize));
Value cstBlockSizeSquare = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(blockSize * blockSize));
Value cDivBlockSizeSquare = rewriter.create<Torch::AtenDivIntOp>(
binder.getLoc(), c, cstBlockSizeSquare);
cDivBlockSizeSquare = rewriter.create<Torch::AtenIntFloatOp>(
binder.getLoc(), cDivBlockSizeSquare);
Value reshapeSizesList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(input.getContext())),
llvm::SmallVector<Value>{b, cstBlockSize, cstBlockSize,
cDivBlockSizeSquare, h, w});
int64_t cDivBlockSizeSquareInt =
inputSizes[1] == Torch::kUnknownSize
? Torch::kUnknownSize
: inputSizes[1] / (blockSize * blockSize);
SmallVector<int64_t, 6> reshapeSizesInt{
inputSizes[0], blockSize, blockSize,
cDivBlockSizeSquareInt, inputSizes[2], inputSizes[3]};
Value reshapedInput = rewriter.create<Torch::AtenReshapeOp>(
binder.getLoc(),
inputTy.getWithSizesAndDtype(reshapeSizesInt,
inputTy.getOptionalDtype()),
input, reshapeSizesList);
Value transposedInput;
if (mode == "DCR") {
if (failed(createTorchTransposeOp(
rewriter, binder.getLoc(), reshapedInput,
/*dimA=*/1, /*dimB=*/3, transposedInput)))
return rewriter.notifyMatchFailure(
binder.op, "Failed to create TorchTranspose op");
if (failed(createTorchTransposeOp(
rewriter, binder.getLoc(), transposedInput,
/*dimA=*/2, /*dimB=*/4, transposedInput)))
return rewriter.notifyMatchFailure(
binder.op, "Failed to create TorchTranspose op");
} else {
// mode == "CRD"
if (failed(createTorchTransposeOp(
rewriter, binder.getLoc(), reshapedInput,
/*dimA=*/2, /*dimB=*/4, transposedInput)))
return rewriter.notifyMatchFailure(
binder.op, "Failed to create TorchTranspose op");
if (failed(createTorchTransposeOp(
rewriter, binder.getLoc(), transposedInput,
/*dimA=*/3, /*dimB=*/4, transposedInput)))
return rewriter.notifyMatchFailure(
binder.op, "Failed to create TorchTranspose op");
}
if (failed(createTorchTransposeOp(
rewriter, binder.getLoc(), transposedInput,
/*dimA=*/4, /*dimB=*/5, transposedInput)))
return rewriter.notifyMatchFailure(
binder.op, "Failed to create TorchTranspose op");
Value hMulBlockSize = rewriter.create<Torch::AtenMulIntOp>(
binder.getLoc(), h, cstBlockSize);
Value wMulBlockSize = rewriter.create<Torch::AtenMulIntOp>(
binder.getLoc(), w, cstBlockSize);
reshapeSizesList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(input.getContext())),
llvm::SmallVector<Value>{b, cDivBlockSizeSquare, hMulBlockSize,
wMulBlockSize});
rewriter.replaceOpWithNewOp<Torch::AtenReshapeOp>(
binder.op, resultType, transposedInput, reshapeSizesList);
return success();
});
patterns.onOp(
"DequantizeLinear", 1,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
llvm::SmallVector<Value> operands;
if (binder.tensorOperands(operands, 3) ||
binder.tensorResultType(resultType))
return failure();
Value operand = operands[0];
Value scale = operands[1];
Value zeropoint = operands[2];
auto operandTy = operand.getType().cast<Torch::ValueTensorType>();
auto scaleTy = scale.getType().dyn_cast<Torch::ValueTensorType>();
if (!scaleTy || !scaleTy.hasSizes())
return rewriter.notifyMatchFailure(binder.op, "requires known rank");
if (!resultType.hasDtype())
return rewriter.notifyMatchFailure(binder.op,
"requires known result dtype");
if (scaleTy.getSizes().size() == 0 ||
(scaleTy.getSizes().size() == 1 && scaleTy.getSizes()[0] == 1)) {
Type qTy = operandTy.getDtype();
if (qTy.isUnsignedInteger(8)) {
qTy = rewriter.getType<Torch::QUInt8Type>();
} else if (qTy.isSignedInteger(8)) {
qTy = rewriter.getType<Torch::QInt8Type>();
} else if (qTy.isSignedInteger(32)) {
qTy = rewriter.getType<Torch::QInt32Type>();
} else {
return rewriter.notifyMatchFailure(binder.op,
"unsupported result dtype");
}
auto qTensorTy = rewriter.getType<Torch::ValueTensorType>(
resultType.getOptionalSizes(), qTy);
scale = rewriter.create<Torch::AtenItemOp>(
binder.getLoc(), rewriter.getType<Torch::FloatType>(), scale);
zeropoint = rewriter.create<Torch::AtenItemOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(), zeropoint);
auto quantize =
rewriter.create<Torch::Aten_MakePerTensorQuantizedTensorOp>(
binder.getLoc(), qTensorTy, operand, scale, zeropoint);
rewriter.replaceOpWithNewOp<Torch::AtenDequantizeSelfOp>(
binder.op, resultType, quantize);
return success();
}
return rewriter.notifyMatchFailure(binder.op,
"unimplemented: non-scalar scale");
});
patterns.onOp("Div", 7,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenDivTensorOp>(
binder.op, resultType, lhs, rhs);
return success();
});
patterns.onOp(
"Dropout", 12, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Location loc = binder.getLoc();
Torch::ValueTensorType resultType;
int64_t numOperands = binder.op->getNumOperands();
SmallVector<Value> operands;
int64_t seed;
if (binder.tensorOperands(operands, numOperands) ||
binder.s64IntegerAttr(seed, "seed", 0) ||
binder.tensorResultTypeAtIndex(resultType, 0))
return failure();
// Global Seed value is 0.
if (seed != 0) {
return rewriter.notifyMatchFailure(binder.op,
"expected seed value to be 0");
}
Value ratio, trainingMode;
if (numOperands == 3) {
ratio = rewriter.create<Torch::AtenFloatImplicitOp>(loc, operands[1]);
Value trainVal = operands[2];
auto trainTensorType =
trainVal.getType().dyn_cast<Torch::BaseTensorType>();
if (!trainTensorType)
return rewriter.notifyMatchFailure(binder.op,
"train tensor must have a type");
Type inputDtype = trainTensorType.getOptionalDtype();
if (!inputDtype || !inputDtype.isInteger(1))
return rewriter.notifyMatchFailure(
binder.op,
"train tensor must have an integer dtype of width 1");
std::optional<unsigned> inputRank = Torch::getTensorRank(trainVal);
if (!inputRank || *inputRank != 0)
return rewriter.notifyMatchFailure(binder.op,
"train tensor must have rank 0");
if (auto valueTensorLiteralOp =
trainVal.getDefiningOp<Torch::ValueTensorLiteralOp>()) {
auto val = valueTensorLiteralOp.getValue()
.cast<DenseElementsAttr>()
.getSplatValue<bool>();
trainingMode = rewriter.create<Torch::ConstantBoolOp>(loc, val);
} else {
Value trainingModeScalar =
rewriter.create<Torch::AtenIntImplicitOp>(loc, operands[2]);
Value cstOne = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(1));
trainingMode = rewriter.create<Torch::AtenEqIntOp>(
loc, trainingModeScalar, cstOne);
}
} else if (numOperands == 2) {
ratio = rewriter.create<Torch::AtenFloatImplicitOp>(loc, operands[1]);
trainingMode = rewriter.create<Torch::ConstantBoolOp>(loc, false);
} else {
ratio = rewriter.create<Torch::ConstantFloatOp>(
loc, rewriter.getF64FloatAttr(0.5));
trainingMode = rewriter.create<Torch::ConstantBoolOp>(loc, false);
}
Value dropout = rewriter.create<Torch::AtenDropoutOp>(
loc, resultType, /*input=*/operands[0], ratio, trainingMode);
if (binder.op->getNumResults() == 1) {
rewriter.replaceOp(binder.op, dropout);
return success();
}
Torch::ValueTensorType maskType;
if (binder.tensorResultTypeAtIndex(maskType, 1))
return failure();
Value dtype = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(
(int64_t)torch_upstream::ScalarType::Bool));
Value none = rewriter.create<Torch::ConstantNoneOp>(loc);
Value mask = rewriter.create<Torch::AtenOnesLikeOp>(
loc, maskType, operands[0], dtype, /*layout=*/none,
/*device=*/none, /*pin_memory=*/none, /*memory_format=*/none);
rewriter.replaceOp(binder.op, {dropout, mask});
return success();
});
patterns.onOp(
"DynamicQuantizeLinear", 11,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Location loc = binder.getLoc();
Value input;
Torch::ValueTensorType resultType, scaleType, zeroPointType;
if (binder.tensorOperand(input) ||
binder.tensorResultTypeAtIndex(resultType, 0) ||
binder.tensorResultTypeAtIndex(scaleType, 1) ||
binder.tensorResultTypeAtIndex(zeroPointType, 2))
return failure();
Value scale, zeroPoint;
// scale = ( max(0, max(input)) - min(0, min(input)) ) / 255
Value inputMax =
rewriter.create<Torch::AtenMaxOp>(loc, scaleType, input);
Value inputMin =
rewriter.create<Torch::AtenMinOp>(loc, scaleType, input);
Value constantZero = rewriter.create<Torch::ConstantFloatOp>(
loc, rewriter.getF64FloatAttr(0));
Value constantOne = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(1));
Value zeroTensor =
createRank0Tensor(rewriter, loc, scaleType, constantZero);
Value inputMaxW0 = rewriter.create<Torch::AtenMaximumOp>(
loc, scaleType, inputMax, zeroTensor);
Value inputMinW0 = rewriter.create<Torch::AtenMinimumOp>(
loc, scaleType, inputMin, zeroTensor);
Value scaleTensor = rewriter.create<Torch::AtenSubTensorOp>(
loc, scaleType, inputMaxW0, inputMinW0, constantOne);
// Note: the following is hard-coded for ui8
Value width = rewriter.create<Torch::ConstantFloatOp>(
loc, rewriter.getF64FloatAttr(255));
Value widthTensor = createRank0Tensor(rewriter, loc, scaleType, width);
scaleTensor = rewriter.create<Torch::AtenDivTensorOp>(
loc, scaleType, scaleTensor, widthTensor);
// compute the preZeroPoint = 0 - (inputMin/scale)
// compute the zeroPoint = cast ( round (clip or saturate
// (preZeroPoint)))
Value preZeroPoint = rewriter.create<Torch::AtenDivTensorOp>(
loc, scaleType, inputMin, scaleTensor);
preZeroPoint = rewriter.create<Torch::AtenSubTensorOp>(
loc, scaleType, zeroTensor, preZeroPoint, constantOne);
// saturate to interval [0, 255]
preZeroPoint = rewriter.create<Torch::AtenClampOp>(
loc, scaleType, preZeroPoint, /*min=*/constantZero, /*max=*/width);
// round, then cast to uint8
preZeroPoint =
rewriter.create<Torch::AtenRoundOp>(loc, scaleType, preZeroPoint);
Type qTy = rewriter.getType<Torch::QUInt8Type>();
auto qTensorTy = rewriter.getType<Torch::ValueTensorType>(
resultType.getOptionalSizes(), qTy);
auto torchqTy = Torch::getScalarTypeForType(qTy);
Value tyConst = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64),
static_cast<int64_t>(torchqTy)));
Value none = rewriter.create<Torch::ConstantNoneOp>(loc);
Value cstFalse = rewriter.create<Torch::ConstantBoolOp>(loc, false);
Value zeroPointTensor = rewriter.create<Torch::AtenToDtypeOp>(
loc, zeroPointType, preZeroPoint, tyConst,
/*non_blocking=*/cstFalse, /*copy=*/cstFalse,
/*memory_format=*/none);
// extract scale and zeroPoint scalars to pass to
// AtenQuantizePerTensorOp
zeroPoint = rewriter.create<Torch::AtenItemOp>(
loc, rewriter.getType<Torch::IntType>(), zeroPointTensor);
scale = rewriter.create<Torch::AtenItemOp>(
loc, rewriter.getType<Torch::FloatType>(), scaleTensor);
Value quantizedTensor = rewriter.create<Torch::AtenQuantizePerTensorOp>(
loc, qTensorTy, input, scale, zeroPoint, tyConst);
// get uint8 tensor output
Value output = rewriter.create<Torch::AtenIntReprOp>(loc, resultType,
quantizedTensor);
rewriter.replaceOp(binder.op, {output, scaleTensor, zeroPointTensor});
return success();
});
patterns.onOp("Equal", 1,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value lhs, rhs;
std::string direction;
if (binder.tensorOperands(lhs, rhs) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenEqTensorOp>(
binder.op, resultType, lhs, rhs);
return success();
});
patterns.onOp("Elu", 6,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Location loc = binder.getLoc();
Torch::ValueTensorType resultType;
Value input;
float alpha;
if (binder.tensorOperand(input) ||
binder.f32FloatAttr(alpha, "alpha") ||
binder.tensorResultType(resultType))
return failure();
Value cstAlpha = rewriter.create<Torch::ConstantFloatOp>(
loc, rewriter.getF64FloatAttr(alpha));
Value cstOne = rewriter.create<Torch::ConstantFloatOp>(
loc, rewriter.getF64FloatAttr(1.0));
rewriter.replaceOpWithNewOp<Torch::AtenEluOp>(
binder.op, resultType, input, cstAlpha, /*scale=*/cstOne,
/*input_scale=*/cstOne);
return success();
});
patterns.onOp("Erf", 13,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
std::string direction;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenErfOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp("Exp", 6,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenExpOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp(
"Expand", 1, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
// uses ideas and code from onnx.Reshape
auto loc = binder.getLoc();
Torch::ValueTensorType resultType;
Value data, shape;
if (binder.tensorOperands(data, shape) ||
binder.tensorResultType(resultType))
return failure();
auto dataType = cast<Torch::BaseTensorType>(data.getType());
auto shapeType = cast<Torch::BaseTensorType>(shape.getType());
if (!dataType.hasSizes() || !shapeType.hasSizes())
return failure();
auto shapeSizes = shapeType.getSizes();
int64_t dataRank = dataType.getSizes().size();
int64_t shapeRank = shapeSizes.size();
if (shapeRank != 1 || shapeSizes[0] == Torch::kUnknownSize)
return failure();
auto rankDifference = dataRank - shapeSizes[0];
SmallVector<int64_t> selectSizes;
Type selectResultType = shapeType.getWithSizesAndDtype(
llvm::ArrayRef(selectSizes), shapeType.getOptionalDtype());
// Variable to store 1-D onnx shape tensor, shapeSizes[0] has the
// dimension size
// A constant zero value
Value zero = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(0));
// Variable to store pytorch int list of shape (dimension)
SmallVector<Value> dimList;
// Convert the shape tensor from vector of int64_t to torch int list as
// we are using torch implementation Torch::AtenBroadcastToOp which
// takes list of int
for (int i = 0; i < shapeSizes[0]; i++) {
Value selectIndex = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), i));
Value extract = rewriter.create<Torch::AtenSelectIntOp>(
loc, selectResultType, shape, zero, selectIndex);
Value dim = rewriter.create<Torch::AtenItemOp>(
loc, rewriter.getType<Torch::IntType>(), extract);
if (i + rankDifference >= 0) {
Value iv =
rewriter.create<Torch::ConstantIntOp>(loc, i + rankDifference);
auto sz = rewriter.create<Torch::AtenSizeIntOp>(
loc, rewriter.getType<Torch::IntType>(), data, iv);
dim = rewriter.create<Torch::PrimMaxIntOp>(loc, dim, sz);
}
dimList.push_back(dim);
}
Value dimValueList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
dimList);
rewriter.replaceOpWithNewOp<Torch::AtenBroadcastToOp>(
binder.op, resultType, data, dimValueList);
return success();
});
patterns.onOp(
"EyeLike", 9, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
int64_t dtypeIntOnnx, diagonalIndex;
if (binder.tensorOperand(operand) ||
binder.s64IntegerAttr(dtypeIntOnnx, "dtype", 1) ||
binder.s64IntegerAttr(diagonalIndex, "k", 0) ||
binder.tensorResultType(resultType))
return failure();
auto operandTy = cast<Torch::ValueTensorType>(operand.getType());
SmallVector<int64_t> shape(operandTy.getSizes());
for (unsigned i = 0; i < shape.size(); i++) {
if (shape[i] == ShapedType::kDynamic)
shape[i] = Torch::kUnknownSize;
}
Value cst0 = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(0));
Value cst1 = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(1));
Value nVal = rewriter.create<Torch::AtenSizeIntOp>(binder.getLoc(),
operand, cst0);
Value mVal = rewriter.create<Torch::AtenSizeIntOp>(binder.getLoc(),
operand, cst1);
Value noneVal = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
int64_t dtypeIntTorch = onnxDtypeIntToTorchDtypeInt(dtypeIntOnnx);
Value dtypeVal = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(dtypeIntTorch));
// diagonalIndex = 0 populates the main diagonal
// diagonalIndex > 0 populates an upper diagonal
// diagonalIndex < 0 populates a lower diagonal
if (diagonalIndex == 0) {
rewriter.replaceOpWithNewOp<Torch::AtenEyeMOp>(
binder.op, resultType, nVal, mVal, dtypeVal, noneVal, noneVal,
noneVal);
return success();
}
Value diagVal = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(),
rewriter.getI64IntegerAttr(std::abs(diagonalIndex)));
Value newN, newM, dimVal, startVal;
// get shapes of main diag eye op and zeros op
if (diagonalIndex > 0) {
newN = nVal;
newM = rewriter.create<Torch::AtenSubIntOp>(binder.getLoc(), mVal,
diagVal);
if (shape[1] != Torch::kUnknownSize) {
shape[1] -= diagonalIndex;
}
dimVal = cst1;
startVal = mVal;
} else {
newN = rewriter.create<Torch::AtenSubIntOp>(binder.getLoc(), nVal,
diagVal);
newM = mVal;
if (shape[0] != Torch::kUnknownSize) {
shape[0] += diagonalIndex;
}
dimVal = cst0;
startVal = nVal;
}
// create main diag eye op
auto eyeResultType = rewriter.getType<Torch::ValueTensorType>(
shape, resultType.getOptionalDtype());
Value eyeOp = rewriter.create<Torch::AtenEyeMOp>(
binder.getLoc(), eyeResultType, newN, newM, dtypeVal, noneVal,
noneVal, noneVal);
// create zeros op
SmallVector<Value> zerosShapeValues = {nVal, mVal};
Value zerosShapeList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
rewriter.getType<Torch::ListType>(
rewriter.getType<Torch::IntType>()),
zerosShapeValues);
Value zerosOp = rewriter.create<Torch::AtenZerosOp>(
binder.getLoc(), resultType, zerosShapeList, dtypeVal, noneVal,
noneVal, noneVal);
// embeds the values of the eye matrix into zeros
rewriter.replaceOpWithNewOp<Torch::AtenSliceScatterOp>(
binder.op, resultType, zerosOp, eyeOp, dimVal,
/*start=*/diagVal, /*end=*/startVal, /*step=*/cst1);
return success();
});
patterns.onOp(
"Flatten", 13, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
// Flatten means to partition the input tensor's dimensions
// into a "left range" spanning 0 to axis - 1 and a "right range"
// spanning axis to rank - 1. Each range is then collapsed
// into a single dimension, resulting in a 2-D tensor.
// If either range is empty, it is replaced with a single
// dimension of size 1.
//
// For example, for a 4-D input tensor of shape (a, b, c, d)
// and axis==2, flatten produces a 2-D tensor of shape
// (a*b, c*d).
//
// If instead axis==0, the left range is empty, and the result
// is (1, a*b*c*d).
Torch::ValueTensorType resultType;
Value operand;
int64_t axis;
if (binder.tensorOperand(operand) ||
binder.s64IntegerAttr(axis, "axis", 1) ||
binder.tensorResultType(resultType))
return failure();
auto operandTy = cast<Torch::ValueTensorType>(operand.getType());
llvm::SmallVector<int64_t> shape(operandTy.getSizes());
int64_t rank = shape.size();
// If axis is negative, count from the right instead of left
if (axis < 0)
axis = rank + axis;
// We collapse in the dimensions to the right of the axis.
for (int i = axis + 1; i < rank; ++i) {
bool dynamic = shape[axis] == Torch::kUnknownSize ||
shape[i] == Torch::kUnknownSize;
if (dynamic) {
shape[axis] = Torch::kUnknownSize;
} else {
shape[axis] = shape[axis] * shape[i];
}
}
shape.resize(axis + 1, 1);
auto baseType = rewriter.getType<Torch::ValueTensorType>(
shape, operandTy.getDtype());
Value collapsedRight;
if (axis >= rank) {
// If the right range is empty, add a dim of size 1 to the
// right side of the shape:
// cr = torch.unsqueeze(x, x.ndim)
Value rankConst = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(rank));
collapsedRight = rewriter.create<Torch::AtenUnsqueezeOp>(
binder.getLoc(), baseType, operand, rankConst);
} else {
// Otherwise, collapse the right range into a single dimension:
// cr = torch._prims.collapse(x, axis, x.ndim - 1)
Value axisConst = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(axis));
Value rankLess1Const = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(rank - 1));
collapsedRight = rewriter.create<Torch::PrimsCollapseOp>(
binder.getLoc(), baseType, operand, axisConst, rankLess1Const);
}
Value zero = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(0));
if (axis <= 0) {
// If the left range is empty, add a dim of size 1 to the
// left side of the shape:
// torch.unsqueeze(cr, 0)
rewriter.replaceOpWithNewOp<Torch::AtenUnsqueezeOp>(
binder.op, resultType, collapsedRight, zero);
return success();
}
// Otherwise, collapse the left range into a single dimension:
// torch._prims.collapse(cr, 0, axis - 1)
Value axisLess1Const = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getI64IntegerAttr(axis - 1));
rewriter.replaceOpWithNewOp<Torch::PrimsCollapseOp>(
binder.op, resultType, collapsedRight, zero, axisLess1Const);
return success();
});
patterns.onOp("Floor", 13,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value operand;
if (binder.tensorOperand(operand) ||
binder.tensorResultType(resultType))
return failure();
rewriter.replaceOpWithNewOp<Torch::AtenFloorOp>(
binder.op, resultType, operand);
return success();
});
patterns.onOp(
"ConstantOfShape", 1,
[](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
Value shape;
if (binder.tensorOperand(shape) || binder.tensorResultType(resultType))
return failure();
// convert shape tensor to list of ints
auto shapeSizes =
dyn_cast<Torch::ValueTensorType>(shape.getType()).getSizes();
SmallVector<Value> dimList;
Torch::BaseTensorType shapeType =
shape.getType().cast<Torch::BaseTensorType>();
Type selectResultType = rewriter.getType<Torch::ValueTensorType>(
ArrayRef<int64_t>({}), shapeType.getOptionalDtype());
Value zero = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), 0));
for (int i = 0; i < shapeSizes[0]; i++) {
Value selectIndex = rewriter.create<Torch::ConstantIntOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(),
rewriter.getIntegerAttr(rewriter.getIntegerType(64), i));
Value extract = rewriter.create<Torch::AtenSelectIntOp>(
binder.getLoc(), selectResultType, shape, zero, selectIndex);
Value dim = rewriter.create<Torch::AtenItemOp>(
binder.getLoc(), rewriter.getType<Torch::IntType>(), extract);
dimList.push_back(dim);
}
Value dimValueList = rewriter.create<Torch::PrimListConstructOp>(
binder.getLoc(),
Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
dimList);
Value noneVal = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
// Get fill_value if it is present.
// Assumption : resultDType and value attr type match.
auto attr = binder.op->getAttr("torch.onnx.value");
auto resultDType = resultType.getDtype();
// Extract the fill value and dtype
// ONNX requires value attr to be a tensor
if (!attr) {
attr = DenseElementsAttr::get(
resultType.toBuiltinTensor().clone(resultDType),
rewriter.getFloatAttr(resultDType, 0.0));
}
// If its a dense resource attr we need to convert to a dense type:
if (DenseResourceElementsAttr rattr =
dyn_cast_or_null<DenseResourceElementsAttr>(attr)) {
// Bytes are stored in little endian order. Big endian support will
// require swizzling.
if (!Endian::little) {
binder.op->emitError(
"unimplemented: importing on big endian systems");
return failure();
}
auto ty = cast<ShapedType>(rattr.getType());
auto ptr = rattr.getRawHandle().getBlob()->getData();
auto denseAttr = DenseElementsAttr::getFromRawBuffer(ty, ptr);
attr = dyn_cast_or_null<SplatElementsAttr>(denseAttr);
}
Attribute splattr;
if (isa<SplatElementsAttr>(attr)) {
auto denseAttr = cast<DenseElementsAttr>(attr);
splattr = denseAttr.getSplatValue<Attribute>();
}
if (!isa<FloatAttr, IntegerAttr>(splattr)) {
return rewriter.notifyMatchFailure(
binder.op,
"`value` attr tensor only supports types int and float for now.");
}
Value splatvalue;
if (auto intattr = dyn_cast<IntegerAttr>(splattr)) {
IntegerType intty = cast<IntegerType>(intattr.getType());
int64_t value;
if (intty.isUnsignedInteger()) {
value = intattr.getUInt();
} else if (intty.isSignedInteger()) {
value = intattr.getSInt();
} else {
value = intattr.getInt();
}
splatvalue =
rewriter.create<Torch::ConstantIntOp>(binder.getLoc(), value);
}
if (auto fpattr = dyn_cast<FloatAttr>(splattr))
splatvalue = rewriter.create<Torch::ConstantFloatOp>(
binder.getLoc(),
rewriter.getF64FloatAttr(fpattr.getValueAsDouble()));
rewriter.replaceOpWithNewOp<Torch::AtenFullOp>(
binder.op, resultType, dimValueList, splatvalue, /*dtype=*/noneVal,
/*layout=*/noneVal, /*device=*/noneVal, /*pin_memory=*/noneVal);
return success();
});
patterns.onOp(
"Einsum", 12, [](OpBinder binder, ConversionPatternRewriter &rewriter) {
Torch::ValueTensorType resultType;
SmallVector<Value> tensors;
std::string equation;
if (binder.tensorOperands(tensors, binder.op->getNumOperands()) ||
binder.customOpNameStringAttr(equation, "equation") ||
binder.tensorResultType(resultType))
return failure();
Type listElemType =
tensors[0]
.getType()
.cast<Torch::BaseTensorType>()
.getWithSizesAndDtype(/*optionalSizes=*/std::nullopt,
/*optionalDtype=*/nullptr);
Type listType = Torch::ListType::get(listElemType);
Value tensorList = rewriter.create<Torch::PrimListConstructOp>(
binder.op->getLoc(), listType, tensors);
Value cstEquation = rewriter.create<Torch::ConstantStrOp>(
binder.getLoc(), rewriter.getType<Torch::StringType>(),
rewriter.getStringAttr(equation));
Value cstNone = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
rewriter.replaceOpWithNewOp<Torch::AtenEinsumOp>(
binder.op, resultType, cstEquation, tensorList, /*path=*/cstNone);
return success();
});
}