torch-mlir/lib/Conversion/TorchToLinalg/Linear.cpp

1391 lines
59 KiB
C++

//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// Also available under a BSD-style license. See LICENSE.
//
//===----------------------------------------------------------------------===//
#include "torch-mlir/Conversion/TorchToLinalg/TorchToLinalg.h"
#include "PopulatePatterns.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/ControlFlow/IR/ControlFlowOps.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/IR/Matchers.h"
#include "torch-mlir/Conversion/TorchToLinalg/Utils.h"
#include "torch-mlir/Conversion/Utils/Utils.h"
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
#include "torch-mlir/Dialect/Torch/Utils/TorchUpstream.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include <algorithm>
using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
namespace {
static void getZeroPoint(Value value, Value &zeropoint) {
if (auto make = value.getDefiningOp<Aten_MakePerTensorQuantizedTensorOp>()) {
zeropoint = make.getZeroPoint();
}
}
// for uint8 types, we shift down by 128 so that we can faithfully
// represent the quantization with signed i8 types.
static void signShift(PatternRewriter &rewriter, Location loc, Value &arg,
Value &zp, bool isUnsignedType, int64_t numBits) {
if (!isUnsignedType)
return;
int64_t minSI = -(1 << (numBits - 1));
Value minSIValue = rewriter.create<arith::ConstantIntOp>(
loc, minSI, cast<mlir::IntegerType>(zp.getType()).getWidth());
zp = rewriter.create<arith::AddIOp>(loc, zp, minSIValue);
minSIValue = rewriter.create<arith::ConstantIntOp>(loc, minSI, numBits);
arg = torch_to_linalg::createElementwiseLinalgGeneric(
rewriter, loc, ValueRange{arg},
cast<TensorType>(arg.getType()).getElementType(),
[&](OpBuilder &b, Location loc, ValueRange payloadArgs) {
Value result =
rewriter.create<arith::AddIOp>(loc, payloadArgs[0], minSIValue);
b.create<linalg::YieldOp>(loc, result);
});
}
static Value transposeValue(Location loc, Value value, ArrayRef<int64_t> perms,
PatternRewriter &rewriter) {
auto valueTy = cast<RankedTensorType>(value.getType());
auto inShape = valueTy.getShape();
llvm::SmallVector<int64_t> outShape;
llvm::SmallVector<Value> dynDims;
for (size_t i = 0; i < perms.size(); ++i) {
outShape.push_back(inShape[perms[i]]);
if (ShapedType::isDynamic(inShape[perms[i]])) {
dynDims.push_back(rewriter.create<tensor::DimOp>(loc, value, perms[i]));
}
}
auto outTy = RankedTensorType::get(outShape, valueTy.getElementType());
Value empty = rewriter.create<tensor::EmptyOp>(loc, outTy, dynDims);
Value transpose =
rewriter.create<linalg::TransposeOp>(loc, value, empty, perms)
->getResult(0);
return transpose;
}
class ConvertAtenMmOp : public OpConversionPattern<AtenMmOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenMmOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Value lhs = adaptor.getSelf();
Value rhs = adaptor.getMat2();
// A user can write an errorneous program where `aten.mm` is in fact called
// with operands of invalid rank or dtype. We cannot convert to linalg in
// this case or we will get a verifier error, which corresponds to breaking
// of *internal* compiler invariants, and for a user manifests as a compiler
// crash in the worst case (such as we try to canonicalize/fold/print the
// invalid op before the verifier gets to see it -- also release builds of a
// mature compiler usually have the verifier turned off for compile time
// reasons).
//
// The compiler cannot crash even if the user wrote an erroneous program!
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
RankedTensorType lhsType = cast<RankedTensorType>(lhs.getType());
RankedTensorType rhsType = cast<RankedTensorType>(rhs.getType());
if (lhsType.getRank() != 2 || rhsType.getRank() != 2) {
return rewriter.notifyMatchFailure(
op, "expected both operands to aten.mm to be rank 2");
}
ValueTensorType lhsTorchType =
cast<ValueTensorType>(op.getSelf().getType());
ValueTensorType rhsTorchType =
cast<ValueTensorType>(op.getMat2().getType());
Value lhsZeroPoint, rhsZeroPoint;
getZeroPoint(op.getSelf(), lhsZeroPoint);
getZeroPoint(op.getMat2(), rhsZeroPoint);
if (static_cast<bool>(lhsZeroPoint) != static_cast<bool>(rhsZeroPoint)) {
return rewriter.notifyMatchFailure(
op, "unsupported: aten.mm with mixed quantization");
}
if (lhsTorchType.getDtype() != rhsTorchType.getDtype()) {
if (!lhsZeroPoint) {
return rewriter.notifyMatchFailure(
op, "unsupported: aten.mm with different input element types");
}
// Allows quantized types to mismatch since they will be cast to the same
// type.
}
bool isUnsigned = torch_to_linalg::isUnsignedTorchType(lhsTorchType);
bool isUnsignedR = torch_to_linalg::isUnsignedTorchType(rhsTorchType);
Value lhsDim0 = rewriter.create<tensor::DimOp>(loc, lhs, 0);
Value rhsDim1 = rewriter.create<tensor::DimOp>(loc, rhs, 1);
if (!isAssumingStrictSymbolicShapes(rewriter)) {
Value lhsDim1 = rewriter.create<tensor::DimOp>(loc, lhs, 1);
Value rhsDim0 = rewriter.create<tensor::DimOp>(loc, rhs, 0);
Value contractingDimEqual = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::eq, lhsDim1, rhsDim0);
rewriter.create<cf::AssertOp>(
loc, contractingDimEqual,
rewriter.getStringAttr(
"mismatching contracting dimension for torch.aten.mm"));
}
TensorType resultType =
cast<TensorType>(getTypeConverter()->convertType(op.getType()));
Type elementType = resultType.getElementType();
auto accumulatorDType =
getDefaultAccType(rewriter, lhsType.getElementType());
if (accumulatorDType != resultType.getElementType()) {
elementType = accumulatorDType;
}
Value zeroFill = createZeroInitTensor(
rewriter, loc, ValueRange{lhsDim0, rhsDim1}, elementType);
Value matmul;
if (lhsZeroPoint) {
lhsZeroPoint = typeConverter->materializeTargetConversion(
rewriter, loc,
getTypeConverter()->convertType(lhsZeroPoint.getType()),
lhsZeroPoint);
rhsZeroPoint = typeConverter->materializeTargetConversion(
rewriter, loc,
getTypeConverter()->convertType(rhsZeroPoint.getType()),
rhsZeroPoint);
lhsZeroPoint = rewriter.create<arith::TruncIOp>(
loc, rewriter.getI32Type(), lhsZeroPoint);
rhsZeroPoint = rewriter.create<arith::TruncIOp>(
loc, rewriter.getI32Type(), rhsZeroPoint);
// change uint8 quantization -> int8 quantization
int64_t numBits =
cast<mlir::IntegerType>(lhsType.getElementType()).getWidth();
signShift(rewriter, loc, lhs, lhsZeroPoint, isUnsigned, numBits);
numBits = cast<mlir::IntegerType>(rhsType.getElementType()).getWidth();
signShift(rewriter, loc, rhs, rhsZeroPoint, isUnsignedR, numBits);
matmul =
rewriter
.create<linalg::QuantizedMatmulOp>(
loc, zeroFill.getType(),
ValueRange{lhs, rhs, lhsZeroPoint, rhsZeroPoint}, zeroFill)
.getResult(0);
} else if (isUnsigned) {
auto matmulOp = rewriter.create<linalg::MatmulOp>(
loc, zeroFill.getType(), ValueRange{lhs, rhs}, zeroFill);
matmulOp.setCast(linalg::TypeFn::cast_unsigned);
matmul = matmulOp->getResult(0);
} else {
matmul = rewriter
.create<linalg::MatmulOp>(loc, zeroFill.getType(),
ValueRange{lhs, rhs}, zeroFill)
.getResult(0);
}
if (accumulatorDType != resultType.getElementType()) {
matmul = torch_to_linalg::convertTensorToElementType(
rewriter, loc, matmul, resultType.getElementType());
}
// When constructed with just dynamic sizes, EmptyOp will have a result
// type which has all `?`'s for dimensions, which might not be the result
// type of `op`. The constraints on later linalg ops means that the result
// of the MatmulOp will have this type too. So cast it to the desired type
// so that in the end we have the original result type.
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, matmul);
return success();
}
};
} // namespace
namespace {
class ConvertAtenFlipOp : public OpConversionPattern<AtenFlipOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenFlipOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Value self = adaptor.getSelf();
auto selfRank =
cast<RankedTensorType>(adaptor.getSelf().getType()).getRank();
SmallVector<int64_t> axis;
if (!matchPattern(adaptor.getDims(), m_TorchListOfConstantInts(axis)))
return rewriter.notifyMatchFailure(op,
"only constant dim lists supported");
for (unsigned i = 0, e = axis.size(); i < e; i++) {
axis[i] = toPositiveDim(axis[i], selfRank);
if (!isValidDim(axis[i], selfRank)) {
return rewriter.notifyMatchFailure(op, "axis is statically invalid");
}
}
Value flipped = torch_to_linalg::flipTensor(rewriter, loc, self, axis);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, self.getType(), flipped);
return success();
}
};
} // namespace
namespace {
class ConvertAtenMatmulOp : public OpConversionPattern<AtenMatmulOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenMatmulOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
Value lhs = adaptor.getSelf();
Value rhs = adaptor.getOther();
if (failed(verifyLinalgCompatibleTypes(op, rewriter))) {
return failure();
}
auto lhsType = cast<RankedTensorType>(lhs.getType());
auto rhsType = cast<RankedTensorType>(rhs.getType());
auto lhsTorchType = cast<ValueTensorType>(op.getSelf().getType());
auto rhsTorchType = cast<ValueTensorType>(op.getOther().getType());
// Get the rank of both matrix.
unsigned lhsRank = lhsType.getRank();
unsigned rhsRank = rhsType.getRank();
Value lhsZeroPoint, rhsZeroPoint;
getZeroPoint(op.getSelf(), lhsZeroPoint);
getZeroPoint(op.getOther(), rhsZeroPoint);
if (static_cast<bool>(lhsZeroPoint) != static_cast<bool>(rhsZeroPoint)) {
return rewriter.notifyMatchFailure(
op, "unsupported: aten.matmul with mixed quantization");
}
bool isUnsigned = torch_to_linalg::isUnsignedTorchType(lhsTorchType);
bool isUnsignedR = torch_to_linalg::isUnsignedTorchType(rhsTorchType);
if (!lhsZeroPoint && lhsTorchType.getDtype() != rhsTorchType.getDtype()) {
// Allows quantized types to mismatch
return rewriter.notifyMatchFailure(
op, "unsupported: aten.matmul with different input element types");
}
Type newResultType = getTypeConverter()->convertType(op.getType());
auto resultType = cast<RankedTensorType>(newResultType);
Type elementType = resultType.getElementType();
if (lhsZeroPoint) {
// get each zero point ready to pass to a quantized_matmul
lhsZeroPoint = typeConverter->materializeTargetConversion(
rewriter, loc,
getTypeConverter()->convertType(lhsZeroPoint.getType()),
lhsZeroPoint);
rhsZeroPoint = typeConverter->materializeTargetConversion(
rewriter, loc,
getTypeConverter()->convertType(rhsZeroPoint.getType()),
rhsZeroPoint);
lhsZeroPoint = rewriter.create<arith::TruncIOp>(
loc, rewriter.getI32Type(), lhsZeroPoint);
rhsZeroPoint = rewriter.create<arith::TruncIOp>(
loc, rewriter.getI32Type(), rhsZeroPoint);
// change uint8 quantization -> int8 quantization
int64_t numBits =
cast<mlir::IntegerType>(lhsType.getElementType()).getWidth();
signShift(rewriter, loc, lhs, lhsZeroPoint, isUnsigned, numBits);
numBits = cast<mlir::IntegerType>(rhsType.getElementType()).getWidth();
signShift(rewriter, loc, rhs, rhsZeroPoint, isUnsignedR, numBits);
// for quantized vec-vec, vec-mat, and mat-vec cases, lower to
// expand/collapse + quantized_matmul
bool lhsVec = (lhsRank == 1 && rhsRank <= 2);
bool rhsVec = (lhsRank <= 2 && rhsRank == 1);
if (lhsVec || rhsVec) {
SmallVector<ReassociationIndices> reassociation(1);
reassociation[0].push_back(0);
reassociation[0].push_back(1);
if (lhsVec) {
// unsqueeze lhs to a matrix
int64_t lhsDim = lhsType.getShape()[0];
auto lhsUnsqueezeType = RankedTensorType::get(
ArrayRef<int64_t>{1, lhsDim}, lhsType.getElementType());
lhs = rewriter.create<tensor::ExpandShapeOp>(loc, lhsUnsqueezeType,
lhs, reassociation);
}
if (rhsVec) {
// unsqueeze rhs to a matrix
int64_t rhsDim = rhsType.getShape()[0];
auto rhsUnsqueezeType = RankedTensorType::get(
ArrayRef<int64_t>{rhsDim, 1}, rhsType.getElementType());
rhs = rewriter.create<tensor::ExpandShapeOp>(loc, rhsUnsqueezeType,
rhs, reassociation);
}
// get quantized_matmul and squeeze result
Value lhsDim0 = getDimOp(rewriter, loc, lhs, 0);
Value lhsDim1 = getDimOp(rewriter, loc, lhs, 1);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, 0);
Value rhsDim1 = getDimOp(rewriter, loc, rhs, 1);
checkDimEqualHelper(rewriter, loc, lhsDim1, rhsDim0);
Value zeroTensor = createZeroInitTensor(
rewriter, loc, ValueRange{lhsDim0, rhsDim1}, elementType);
Value matmul = rewriter
.create<linalg::QuantizedMatmulOp>(
loc, zeroTensor.getType(),
ValueRange{lhs, rhs, lhsZeroPoint, rhsZeroPoint},
zeroTensor)
.getResult(0);
int64_t resultRank = resultType.getRank();
if (resultRank == 0) {
// in vec-vec case, need to collapse result to a scalar
reassociation.clear();
}
matmul = rewriter.create<tensor::CollapseShapeOp>(
loc, resultType, matmul, reassociation);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, matmul);
return success();
}
// the remaining quantized cases (Mat-Mat and broadcast -> BMM) are
// covered in the relevant section below
}
// The different cases of torch_matmul op is mentioned here:
// https://pytorch.org/docs/stable/generated/torch.matmul.html
// First Case: Dot Product.
if (lhsRank == 1 && rhsRank == 1) {
Value lhsDim0 = getDimOp(rewriter, loc, lhs, 0);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, 0);
checkDimEqualHelper(rewriter, loc, lhsDim0, rhsDim0);
Value zeroTensor = createZeroInitTensor(rewriter, loc, {}, elementType);
Value dotProd =
rewriter
.create<linalg::DotOp>(loc, zeroTensor.getType(),
ValueRange{lhs, rhs}, zeroTensor)
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, dotProd);
return success();
}
// Second Case: Vec-Mat Multiplication.
if (lhsRank == 1 && rhsRank == 2) {
Value lhsDim0 = getDimOp(rewriter, loc, lhs, 0);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, 0);
Value rhsDim1 = getDimOp(rewriter, loc, rhs, 1);
checkDimEqualHelper(rewriter, loc, lhsDim0, rhsDim0);
Value zeroTensor =
createZeroInitTensor(rewriter, loc, ValueRange{rhsDim1}, elementType);
Value matmul =
rewriter
.create<linalg::VecmatOp>(loc, zeroTensor.getType(),
ValueRange{lhs, rhs}, zeroTensor)
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, matmul);
return success();
}
// Third Case: Matrix-Vec Multiplication.
if (lhsRank == 2 && rhsRank == 1) {
Value lhsDim0 = getDimOp(rewriter, loc, lhs, 0);
Value lhsDim1 = getDimOp(rewriter, loc, lhs, 1);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, 0);
checkDimEqualHelper(rewriter, loc, lhsDim1, rhsDim0);
Value zeroTensor =
createZeroInitTensor(rewriter, loc, ValueRange{lhsDim0}, elementType);
Value matmul =
rewriter
.create<linalg::MatvecOp>(loc, zeroTensor.getType(),
ValueRange{lhs, rhs}, zeroTensor)
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, matmul);
return success();
}
// Fourth Case: Mat-Mat Multiplication.
if (lhsRank == 2 && rhsRank == 2) {
Value lhsDim0 = getDimOp(rewriter, loc, lhs, 0);
Value lhsDim1 = getDimOp(rewriter, loc, lhs, 1);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, 0);
Value rhsDim1 = getDimOp(rewriter, loc, rhs, 1);
checkDimEqualHelper(rewriter, loc, lhsDim1, rhsDim0);
Value zeroTensor = createZeroInitTensor(
rewriter, loc, ValueRange{lhsDim0, rhsDim1}, elementType);
Value matmul;
if (lhsZeroPoint) {
matmul = rewriter
.create<linalg::QuantizedMatmulOp>(
loc, zeroTensor.getType(),
ValueRange{lhs, rhs, lhsZeroPoint, rhsZeroPoint},
zeroTensor)
.getResult(0);
} else {
matmul = rewriter
.create<linalg::MatmulOp>(loc, zeroTensor.getType(),
ValueRange{lhs, rhs}, zeroTensor)
.getResult(0);
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, matmul);
return success();
}
// Fifth Case: Batch-Matrix Multiplication.
// TODO: Handle batch matrix multiplication when one of the matrix is unity
// rank and the other has batch dimension.
if (lhsRank > 1 && rhsRank > 1) {
unsigned maxRank = std::max(lhsRank, rhsRank);
unsigned minRank = std::min(lhsRank, rhsRank);
unsigned batchRank = maxRank - 2;
// At least one of the matrix must have rank greater than 2.
if (batchRank <= 0) {
return rewriter.notifyMatchFailure(op, "expected batch dimensions");
}
// The `broadcastedBatchShape` contains batch dimensions of the resultant
// matrix.
SmallVector<Value> broadcastedBatchShape(batchRank);
Value maxRankMatrix = (lhsRank > rhsRank) ? lhs : rhs;
Value maxDim;
// Compute broadcasted batch dimensions if the batch dimensions of
// the matrices are broadcastable.
for (unsigned i = 1; i <= batchRank; i++) {
if (i <= minRank - 2) {
Value lhsDim = getDimOp(rewriter, loc, lhs, lhsRank - 2 - i);
Value rhsDim = getDimOp(rewriter, loc, rhs, rhsRank - 2 - i);
maxDim = rewriter.createOrFold<arith::MaxUIOp>(loc, lhsDim, rhsDim);
} else {
maxDim = getDimOp(rewriter, loc, maxRankMatrix, maxRank - 2 - i);
}
broadcastedBatchShape[batchRank - i] = maxDim;
}
Value lhsDim0 = getDimOp(rewriter, loc, lhs, lhsRank - 2);
Value lhsDim1 = getDimOp(rewriter, loc, lhs, lhsRank - 1);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, rhsRank - 2);
Value rhsDim1 = getDimOp(rewriter, loc, rhs, rhsRank - 1);
checkDimEqualHelper(rewriter, loc, lhsDim1, rhsDim0);
// Compute broadcasted shape of both the matrices in integer format.
SmallVector<Value> lhsBroadcastToShape(broadcastedBatchShape);
lhsBroadcastToShape.push_back(lhsDim0);
lhsBroadcastToShape.push_back(lhsDim1);
SmallVector<Value> rhsBroadcastToShape(broadcastedBatchShape);
rhsBroadcastToShape.push_back(rhsDim0);
rhsBroadcastToShape.push_back(rhsDim1);
for (unsigned i = 0; i < maxRank; i++) {
lhsBroadcastToShape[i] =
castIndexToInt64(rewriter, loc, lhsBroadcastToShape[i]);
rhsBroadcastToShape[i] =
castIndexToInt64(rewriter, loc, rhsBroadcastToShape[i]);
}
// Broadcast the batch dimensions of both the matrices.
Value broadcastedLhs, broadcastedRhs;
// TODO: Improve usage of static shape information.
SmallVector<int64_t> lhsTargetShape(lhsBroadcastToShape.size(),
ShapedType::kDynamic);
auto lhsBroadcastType = RankedTensorType::get(
lhsTargetShape, lhsType.getElementType(), lhsType.getEncoding());
if (failed(torch_to_linalg::broadcastToGivenShape(
op, rewriter, lhs, lhsBroadcastToShape, lhsBroadcastType,
broadcastedLhs))) {
return rewriter.notifyMatchFailure(
op, "unable to perform broadcast operation");
}
SmallVector<int64_t> rhsTargetShape(rhsBroadcastToShape.size(),
ShapedType::kDynamic);
auto rhsBroadcastType = RankedTensorType::get(
rhsTargetShape, rhsType.getElementType(), rhsType.getEncoding());
if (failed(torch_to_linalg::broadcastToGivenShape(
op, rewriter, rhs, rhsBroadcastToShape, rhsBroadcastType,
broadcastedRhs))) {
return rewriter.notifyMatchFailure(
op, "unable to perform broadcast operation");
}
if (maxRank == 3) {
Value zeroTensor = createZeroInitTensor(
rewriter, loc,
ValueRange{broadcastedBatchShape[0], lhsDim0, rhsDim1},
elementType);
Value matmul;
if (lhsZeroPoint) {
matmul = rewriter
.create<linalg::QuantizedBatchMatmulOp>(
loc, zeroTensor.getType(),
ValueRange{broadcastedLhs, broadcastedRhs,
lhsZeroPoint, rhsZeroPoint},
zeroTensor)
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType,
matmul);
return success();
}
matmul = rewriter
.create<linalg::BatchMatmulOp>(
loc, zeroTensor.getType(),
ValueRange{broadcastedLhs, broadcastedRhs}, zeroTensor)
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, matmul);
return success();
}
// Check if the result of the matrix multiplication has more than one
// dynamic batch dimensions.
SmallVector<int64_t> batchDimsInt =
makeShapeTorchCompatible(resultType.getShape());
batchDimsInt.pop_back();
batchDimsInt.pop_back();
bool multipleDynamicBatchDims =
llvm::count(batchDimsInt, kUnknownSize) > 1;
// TODO: Lowering to `linalg.BatchMatmul` is only possible when there is
// at most one dynamic batch dimension due to limited support of the
// `tensor.ExpandShape` op.
if (!multipleDynamicBatchDims) {
// Collapse the batch dimensions into one dimension. The resultant rank
// will always be 3.
SmallVector<ReassociationIndices> reassociation(3);
for (unsigned i = 0, j = 0; i < maxRank; i++) {
if (i >= batchRank)
j++;
reassociation[j].push_back(i);
}
Value collapsedLhs = rewriter.create<tensor::CollapseShapeOp>(
op->getLoc(), broadcastedLhs, reassociation);
Value collapsedRhs = rewriter.create<tensor::CollapseShapeOp>(
op->getLoc(), broadcastedRhs, reassociation);
// Compute the result shape after collapsing the batch dimensions.
SmallVector<Value> collapsedResultShape;
collapsedResultShape.push_back(broadcastedBatchShape[0]);
for (unsigned i = 1; i < batchRank; i++) {
collapsedResultShape[0] = rewriter.createOrFold<arith::MulIOp>(
loc, collapsedResultShape[0], broadcastedBatchShape[i]);
}
collapsedResultShape.push_back(lhsDim0);
collapsedResultShape.push_back(rhsDim1);
SmallVector<OpFoldResult> updatedCollapseResultShape =
getAsOpFoldResult(collapsedResultShape);
Value initTensor = rewriter.create<tensor::EmptyOp>(
loc, updatedCollapseResultShape, elementType);
Value c0 = rewriter.create<arith::ConstantOp>(
loc, rewriter.getZeroAttr(elementType));
Value zeroTensor =
rewriter.create<linalg::FillOp>(loc, c0, initTensor).getResult(0);
Value batchMatMul;
if (lhsZeroPoint) {
batchMatMul = rewriter
.create<linalg::QuantizedBatchMatmulOp>(
loc, zeroTensor.getType(),
ValueRange{collapsedLhs, collapsedRhs,
lhsZeroPoint, rhsZeroPoint},
zeroTensor)
.getResult(0);
} else {
batchMatMul =
rewriter
.create<linalg::BatchMatmulOp>(
loc, zeroTensor.getType(),
ValueRange{collapsedLhs, collapsedRhs}, zeroTensor)
.getResult(0);
}
Value expandResult = rewriter.create<tensor::ExpandShapeOp>(
loc, resultType, batchMatMul, reassociation);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType,
expandResult);
return success();
}
SmallVector<AffineExpr> lhsExpr;
SmallVector<AffineExpr> rhsExpr;
SmallVector<AffineExpr> outExpr;
SmallVector<utils::IteratorType> iteratorTypes(
batchRank, utils::IteratorType::parallel);
for (unsigned i = 0; i < batchRank; i++) {
lhsExpr.push_back(rewriter.getAffineDimExpr(i));
rhsExpr.push_back(rewriter.getAffineDimExpr(i));
outExpr.push_back(rewriter.getAffineDimExpr(i));
}
lhsExpr.insert(lhsExpr.end(), {rewriter.getAffineDimExpr(batchRank),
rewriter.getAffineDimExpr(batchRank + 1)});
rhsExpr.insert(rhsExpr.end(), {rewriter.getAffineDimExpr(batchRank + 1),
rewriter.getAffineDimExpr(batchRank + 2)});
outExpr.insert(outExpr.end(), {rewriter.getAffineDimExpr(batchRank),
rewriter.getAffineDimExpr(batchRank + 2)});
SmallVector<Value> resultShape(broadcastedBatchShape);
resultShape.insert(resultShape.end(), {lhsDim0, rhsDim1});
Value zeroTensor =
createZeroInitTensor(rewriter, loc, resultShape, elementType);
auto indexingMaps = AffineMap::inferFromExprList(
{lhsExpr, rhsExpr, outExpr}, rewriter.getContext());
iteratorTypes.insert(iteratorTypes.end(),
{utils::IteratorType::parallel,
utils::IteratorType::reduction,
utils::IteratorType::parallel});
Value finalRes =
rewriter
.create<linalg::GenericOp>(
loc, zeroTensor.getType(),
ValueRange{broadcastedLhs, broadcastedRhs}, zeroTensor,
/*indexingMaps=*/indexingMaps,
/*iteratorTypes=*/iteratorTypes,
[&](OpBuilder &b, Location loc, ValueRange args) {
Value l = args[0], r = args[1], res = args[2];
Value mul = b.create<arith::MulFOp>(loc, l, r);
Value add = b.create<arith::AddFOp>(loc, mul, res);
b.create<linalg::YieldOp>(loc, add);
})
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, finalRes);
return success();
}
return failure();
}
};
} // namespace
namespace {
class ConvertAtenBmmOp : public OpConversionPattern<AtenBmmOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenBmmOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op->getLoc();
Value lhs = adaptor.getSelf();
Value rhs = adaptor.getMat2();
RankedTensorType lhsType = cast<RankedTensorType>(lhs.getType());
RankedTensorType rhsType = cast<RankedTensorType>(rhs.getType());
Type newResultType = getTypeConverter()->convertType(op.getType());
Type resultElementType =
cast<RankedTensorType>(newResultType).getElementType();
Type lhsElementType = cast<RankedTensorType>(lhsType).getElementType();
Type rhsElementType = cast<RankedTensorType>(rhsType).getElementType();
if (lhsType.getRank() != 3 || rhsType.getRank() != 3) {
return rewriter.notifyMatchFailure(
op, "expected both operands to aten.bmm to be rank 3");
}
// Convert the inputs element type equivalent to the result' element type.
if (lhsElementType != rhsElementType) {
if (lhsElementType != resultElementType) {
// True if the lhs element type is not equal to the result' element
// type.
lhs = torch_to_linalg::convertTensorToElementType(rewriter, loc, lhs,
resultElementType);
} else {
// True if the rhs element type is not equal to the result' element
// type.
rhs = torch_to_linalg::convertTensorToElementType(rewriter, loc, rhs,
resultElementType);
}
}
Value lhsDim0 = getDimOp(rewriter, loc, lhs, 0);
Value lhsDim1 = getDimOp(rewriter, loc, lhs, 1);
Value lhsDim2 = getDimOp(rewriter, loc, lhs, 2);
Value rhsDim0 = getDimOp(rewriter, loc, rhs, 0);
Value rhsDim1 = getDimOp(rewriter, loc, rhs, 1);
Value rhsDim2 = getDimOp(rewriter, loc, rhs, 2);
// Check the batch numbers are equal.
checkDimEqualHelper(rewriter, loc, lhsDim0, rhsDim0);
// Check the matrixs shapes are valid for mulplication.
checkDimEqualHelper(rewriter, loc, lhsDim2, rhsDim1);
Value initTensor0 = createZeroInitTensor(
rewriter, loc, ValueRange{lhsDim0, lhsDim1, rhsDim2},
resultElementType);
Value bmm =
rewriter
.create<linalg::BatchMatmulOp>(loc, initTensor0.getType(),
ValueRange{lhs, rhs}, initTensor0)
.getResult(0);
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, bmm);
return success();
}
};
} // namespace
namespace {
class ConvertAtenConvolutionOp : public OpConversionPattern<AtenConvolutionOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenConvolutionOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op->getLoc();
MLIRContext *context = op->getContext();
Value input = adaptor.getInput(); /* in form of N*C*H*W */
Value weight = adaptor.getWeight(); /* in form of F*C/G*H*W */
Value bias = adaptor.getBias();
auto resultTy = cast<ValueTensorType>(op.getType());
Value inputZp, weightZp;
bool inputUnsigned = false;
bool weightUnsigned = false;
if (auto make = op.getInput()
.getDefiningOp<Aten_MakePerTensorQuantizedTensorOp>()) {
input = make.getSelf();
inputZp = make.getZeroPoint();
input = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(input.getType()), input);
inputZp = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(inputZp.getType()),
inputZp);
inputZp =
rewriter.create<arith::TruncIOp>(loc, rewriter.getI32Type(), inputZp);
auto torchDtype = cast<ValueTensorType>(make.getType()).getDtype();
inputUnsigned = torch_to_linalg::isUnsignedTorchType(torchDtype);
}
if (auto make = op.getWeight()
.getDefiningOp<Aten_MakePerTensorQuantizedTensorOp>()) {
weight = make.getSelf();
weightZp = make.getZeroPoint();
weight = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(weight.getType()), weight);
weightZp = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(weightZp.getType()),
weightZp);
weightZp = rewriter.create<arith::TruncIOp>(loc, rewriter.getI32Type(),
weightZp);
auto torchDtype = cast<ValueTensorType>(make.getType()).getDtype();
weightUnsigned = torch_to_linalg::isUnsignedTorchType(torchDtype);
}
if (static_cast<bool>(inputZp) != static_cast<bool>(weightZp)) {
return rewriter.notifyMatchFailure(
op, "lhs and rhs of convolution must either be both int or fp");
}
if (inputZp && !isa<Torch::NoneType>(bias.getType())) {
auto biasDTy = cast<RankedTensorType>(bias.getType()).getElementType();
if (!biasDTy.isInteger(32)) {
return rewriter.notifyMatchFailure(
op, "quantized result ty should be i32 accumulator");
}
}
bool transposed = true;
if (!matchPattern(op.getTransposed(), m_TorchConstantBool(&transposed)))
return rewriter.notifyMatchFailure(
op, "unimplemented: only constant transposed supported");
auto inputDTy = cast<RankedTensorType>(input.getType()).getElementType();
auto weightDTy = cast<RankedTensorType>(weight.getType()).getElementType();
auto resultDTy = resultTy.toBuiltinTensor().getElementType();
if (!isa<mlir::FloatType, mlir::IntegerType>(inputDTy) ||
!isa<mlir::FloatType, mlir::IntegerType>(weightDTy) ||
!isa<mlir::FloatType, mlir::IntegerType>(resultDTy))
return op.emitError("unimplemented: non-fp not-int type");
size_t inRank = cast<RankedTensorType>(input.getType()).getRank();
size_t numSpatialDims = inRank - 2;
if (numSpatialDims < 1 || numSpatialDims > 3)
return rewriter.notifyMatchFailure(
op, "unimplemented: only 1d-3d convolution currently supported");
Type intType = IntegerType::get(context, 64);
auto castIndexToInt = [&](Value v) {
return rewriter.createOrFold<arith::IndexCastOp>(loc, intType, v);
};
SmallVector<Value> paddingIntValues;
if (!getListConstructElements(op.getPadding(), paddingIntValues))
return rewriter.notifyMatchFailure(
op, "only support padding from a list construct");
paddingIntValues = getTypeConvertedValues(rewriter, loc, getTypeConverter(),
paddingIntValues);
SmallVector<Value> outputPaddingIntValues;
if (!getListConstructElements(op.getOutputPadding(),
outputPaddingIntValues))
return rewriter.notifyMatchFailure(
op, "only support output_padding from a list construct");
outputPaddingIntValues = getTypeConvertedValues(
rewriter, loc, getTypeConverter(), outputPaddingIntValues);
SmallVector<int64_t> strideInts;
if (!matchPattern(op.getStride(), m_TorchListOfConstantInts(strideInts)))
return rewriter.notifyMatchFailure(op,
"only support constant int strides");
SmallVector<int64_t> dilationInts;
if (!matchPattern(op.getDilation(),
m_TorchListOfConstantInts(dilationInts)))
return rewriter.notifyMatchFailure(op,
"only support constant int dilations");
Value inBatch = getDimOp(rewriter, loc, input, 0);
Value inChannels = getDimOp(rewriter, loc, input, 1);
SmallVector<Value> inDims;
for (size_t i = 2; i < inRank; i++)
inDims.push_back(getDimOp(rewriter, loc, input, i));
Value weightBatch = getDimOp(rewriter, loc, weight, 0);
Value weightChannels = getDimOp(rewriter, loc, weight, 1);
SmallVector<Value> weightDims;
for (size_t i = 2; i < inRank; i++)
weightDims.push_back(getDimOp(rewriter, loc, weight, i));
// Checks for valid group size
int64_t numGroups;
if (!matchPattern(op.getGroups(), m_TorchConstantInt(&numGroups)))
return rewriter.notifyMatchFailure(op,
"only constant group size supported.");
Value groups = castIntToIndex(rewriter, loc, adaptor.getGroups());
auto validate = [&](Value toValidate, std::string err) {
Value c0 =
rewriter.create<arith::ConstantOp>(loc, rewriter.getIndexAttr(0));
Value inputValid = rewriter.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::eq, c0,
rewriter.create<arith::RemSIOp>(loc, toValidate, groups));
rewriter.create<cf::AssertOp>(loc, inputValid,
rewriter.getStringAttr(err));
};
validate(inChannels,
"invalid: groups must divide input channel size evenly.");
validate(weightBatch,
"invalid: groups must divide weight batch size evenly.");
SmallVector<Value> dilationIntValues =
getAsConstantIntValues(rewriter, loc, dilationInts);
SmallVector<Value> strideIntValues =
getAsConstantIntValues(rewriter, loc, strideInts);
// convert any uint8 quantization to int8 quantization
if (auto integerType = dyn_cast<mlir::IntegerType>(inputDTy)) {
int64_t width = integerType.getWidth();
signShift(rewriter, loc, input, inputZp, inputUnsigned, width);
}
if (auto integerType = dyn_cast<mlir::IntegerType>(weightDTy)) {
int64_t width = integerType.getWidth();
signShift(rewriter, loc, weight, weightZp, weightUnsigned, width);
}
// Pad the input tensor according to padding.
SmallVector<Value> outDims{inBatch, weightBatch};
Value paddedInput;
Value pad = inputZp;
if (!pad) {
if (isa<mlir::FloatType>(inputDTy))
pad = rewriter.create<arith::ConstantOp>(
op.getLoc(), rewriter.getFloatAttr(inputDTy, 0.0));
if (isa<mlir::IntegerType>(inputDTy))
pad = rewriter.create<arith::ConstantOp>(
op.getLoc(), rewriter.getIntegerAttr(inputDTy, 0));
}
if (pad.getType() != inputDTy) {
if (isa<mlir::FloatType>(inputDTy))
pad = rewriter.create<arith::TruncFOp>(op.getLoc(), inputDTy, pad);
if (isa<mlir::IntegerType>(inputDTy))
pad = rewriter.create<arith::TruncIOp>(op.getLoc(), inputDTy, pad);
}
if (transposed) {
Value c0 =
rewriter.create<arith::ConstantOp>(loc, rewriter.getIndexAttr(0));
Value c1 =
rewriter.create<arith::ConstantOp>(loc, rewriter.getIndexAttr(1));
Value c2 =
rewriter.create<arith::ConstantOp>(loc, rewriter.getIndexAttr(2));
// Transpose and flip weight
SmallVector<Value> weightInitDims = getTensorSizes(rewriter, loc, weight);
std::iter_swap(weightInitDims.begin(), weightInitDims.begin() + 1);
outDims[1] = weightInitDims[0];
Value weightInitTensor =
createZeroInitTensor(rewriter, loc, weightInitDims, weightDTy);
SmallVector<utils::IteratorType> iteratorTypes(
inRank, utils::IteratorType::parallel);
SmallVector<AffineMap> indexingMaps{
AffineMap::getMultiDimIdentityMap(inRank, context)};
weight = rewriter
.create<linalg::GenericOp>(
loc, weightInitTensor.getType(), ValueRange{},
weightInitTensor, indexingMaps, iteratorTypes,
[&](OpBuilder &b, Location loc, ValueRange args) {
SmallVector<Value> indices;
for (size_t i = 0; i < inRank; i++)
indices.push_back(b.create<linalg::IndexOp>(loc, i));
std::iter_swap(indices.begin(), indices.begin() + 1);
// Flip only the spatial dimensions (from 2 to inRank)
for (size_t flipDim = 2; flipDim < inRank; flipDim++) {
indices[flipDim] = b.create<arith::SubIOp>(
loc,
b.create<arith::SubIOp>(
loc, weightInitDims[flipDim], c1),
indices[flipDim]);
}
Value res =
b.create<tensor::ExtractOp>(loc, weight, indices)
.getResult();
b.create<linalg::YieldOp>(loc, res);
})
.getResult(0);
// Calculate padded input size, allocate tensor
SmallVector<Value> outerSizes{inBatch, inChannels};
SmallVector<Value> innerSizes{inBatch, inChannels};
SmallVector<Value> offsets{c0, c0};
for (size_t i = 0; i < numSpatialDims; i++) {
Value innerSize = rewriter.create<arith::SubIOp>(loc, inDims[i], c1);
innerSize = rewriter.create<arith::MulIOp>(
loc, innerSize, castIntToIndex(rewriter, loc, strideIntValues[i]));
innerSize = rewriter.create<arith::AddIOp>(loc, innerSize, c1);
Value offset = rewriter.create<arith::SubIOp>(loc, weightDims[i], c1);
offset = rewriter.create<arith::MulIOp>(
loc, offset, castIntToIndex(rewriter, loc, dilationIntValues[i]));
offset = rewriter.create<arith::SubIOp>(
loc, offset, castIntToIndex(rewriter, loc, paddingIntValues[i]));
Value outerSize = rewriter.create<arith::MulIOp>(loc, offset, c2);
outerSize = rewriter.create<arith::AddIOp>(loc, outerSize, innerSize);
outerSize = rewriter.create<arith::AddIOp>(
loc, outerSize,
castIntToIndex(rewriter, loc, outputPaddingIntValues[i]));
outerSizes.push_back(outerSize);
offsets.push_back(offset);
}
// Allocate padded input tensor
Value initTensor =
createInitTensor(rewriter, loc, outerSizes, inputDTy, pad);
// Insert input into allocated tensor
SmallVector<Value> strideIndexValues{c1, c1};
for (auto stride : strideIntValues)
strideIndexValues.push_back(castIntToIndex(rewriter, loc, stride));
SmallVector<Value> insertSizes = getTensorSizes(rewriter, loc, input);
paddedInput = rewriter.create<tensor::InsertSliceOp>(
loc, torch_to_linalg::removeSizeInformation(rewriter, loc, input),
initTensor, offsets, insertSizes, strideIndexValues);
// Calculate output dims
for (size_t i = 0; i < numSpatialDims; i++)
outDims.push_back(torch_to_linalg::getOutputDimForConvTransposeOps(
rewriter, loc, inDims[i], paddingIntValues[i], dilationIntValues[i],
castIndexToInt(weightDims[i]), strideIntValues[i],
outputPaddingIntValues[i]));
// Set stride to 1
strideInts.clear();
strideInts.append(numSpatialDims, 1);
} else {
// Pad input
paddedInput = torch_to_linalg::getDynamicZeroPaddedTensor(
op, rewriter, input, paddingIntValues, /*unpaddedDims=*/2, pad);
// Calculate output dims
for (size_t i = 0; i < numSpatialDims; i++)
outDims.push_back(torch_to_linalg::getOutputDimForConvOps(
rewriter, loc, inDims[i], paddingIntValues[i], dilationIntValues[i],
castIndexToInt(weightDims[i]), strideIntValues[i]));
}
Type accumulatorDType = getDefaultAccType(rewriter, inputDTy);
Value initTensor = rewriter.create<tensor::EmptyOp>(
loc, getAsOpFoldResult(outDims), accumulatorDType);
Value outputTensor;
if (accumulatorDType != resultDTy && !isa<Torch::NoneType>(bias.getType()))
bias = torch_to_linalg::convertTensorToElementType(rewriter, loc, bias,
accumulatorDType);
if (isa<Torch::NoneType>(bias.getType())) {
Value c0;
if (isa<mlir::FloatType>(accumulatorDType)) {
c0 = rewriter.create<arith::ConstantOp>(
loc, FloatAttr::get(accumulatorDType, 0.0));
} else if (isa<mlir::IntegerType>(accumulatorDType)) {
c0 = rewriter.create<arith::ConstantOp>(
loc, IntegerAttr::get(accumulatorDType, 0));
}
outputTensor =
rewriter.create<linalg::FillOp>(loc, c0, initTensor).getResult(0);
} else {
auto biasType = cast<RankedTensorType>(bias.getType());
if (biasType.getRank() != 1)
return rewriter.notifyMatchFailure(op, "expect bias to be rank 1");
auto resultRank = cast<RankedTensorType>(initTensor.getType()).getRank();
SmallVector<int64_t, 4> addedDimensions;
// bias is used to initialize the channels - dimension 1 of
// output
for (int i = 0; i < resultRank; ++i)
if (i != 1)
addedDimensions.push_back(i);
outputTensor = rewriter
.create<linalg::BroadcastOp>(loc, bias, initTensor,
addedDimensions)
->getResult(0);
}
auto stridesAttr = rewriter.getI64VectorAttr(strideInts);
auto dilationAttr = rewriter.getI64VectorAttr(dilationInts);
Value inputStride =
rewriter.create<arith::FloorDivSIOp>(loc, inChannels, groups);
Value weightStride =
rewriter.create<arith::FloorDivSIOp>(loc, weightBatch, groups);
SmallVector<Value> zeroOffsets(inRank, rewriter.create<arith::ConstantOp>(
loc, rewriter.getIndexAttr(0)));
SmallVector<Value> unitStrides(inRank, rewriter.create<arith::ConstantOp>(
loc, rewriter.getIndexAttr(1)));
SmallVector<Value> outDimSlice(outDims);
outDimSlice[1] = weightStride;
SmallVector<Value> inputSliceSizes{inBatch, inputStride};
inputSliceSizes.append(inDims);
SmallVector<Value> weightSliceSizes{weightStride, weightChannels};
weightSliceSizes.append(weightDims);
Value conv;
// the code so far is able to respect all numSpatialDims
// the code below this point is numSpatialDims specific and numGroups
// specific
// TODO: factor out the above code into a helper function, and then separate
// convolution into:
// - grouped 1d-3d
// - grouped 1d-3d (quantized)
// - ungrouped 1d-3d
if (numGroups == 1 && !inputZp) {
switch (numSpatialDims) {
case 1:
conv = rewriter
.create<linalg::Conv1DNcwFcwOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight}, outputTensor,
stridesAttr, dilationAttr)
.getResult(0);
break;
case 2:
conv = rewriter
.create<linalg::Conv2DNchwFchwOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight}, outputTensor,
stridesAttr, dilationAttr)
.getResult(0);
break;
case 3:
conv = rewriter
.create<linalg::Conv3DNcdhwFcdhwOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight}, outputTensor,
stridesAttr, dilationAttr)
.getResult(0);
break;
default:
return rewriter.notifyMatchFailure(
op, "unimplemented: only 1D, 2D, and 3D convolution supported");
};
Type newResultType = getTypeConverter()->convertType(op.getType());
if (accumulatorDType != resultDTy) {
Type resultElementType =
cast<RankedTensorType>(newResultType).getElementType();
conv = torch_to_linalg::convertTensorToElementType(rewriter, loc, conv,
resultElementType);
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, conv);
return success();
}
if (numGroups == 1 && inputZp) {
// The quantized version uses a different channel ordering so we need to
// permute the tensors in order to use the existing path. We should
// eventually directly support this channel ordering.
llvm::SmallVector<int64_t> inPerms, weightPerms;
inPerms.push_back(0); // N stays at the front for input.
// Then we expect the spatial dimensions
for (size_t i = 0; i < numSpatialDims; ++i) {
inPerms.push_back(i + 2);
weightPerms.push_back(i + 2);
}
inPerms.push_back(1);
weightPerms.append({1, 0});
paddedInput = transposeValue(op.getLoc(), paddedInput, inPerms, rewriter);
weight = transposeValue(op.getLoc(), weight, weightPerms, rewriter);
outputTensor =
transposeValue(op.getLoc(), outputTensor, inPerms, rewriter);
switch (numSpatialDims) {
case 2:
conv = rewriter
.create<linalg::Conv2DNhwcHwcfQOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight, inputZp, weightZp},
outputTensor, stridesAttr, dilationAttr)
.getResult(0);
break;
case 3:
conv = rewriter
.create<linalg::Conv3DNdhwcDhwcfQOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, weight, inputZp, weightZp},
outputTensor, stridesAttr, dilationAttr)
.getResult(0);
break;
default:
return rewriter.notifyMatchFailure(
op, "unimplemented: only 1D, 2D, and 3D convolution supported");
};
llvm::SmallVector<int64_t> outPerms;
outPerms.push_back(0);
outPerms.push_back(inPerms.size() - 1);
for (size_t i = 0; i < numSpatialDims; ++i) {
outPerms.push_back(i + 1);
}
conv = transposeValue(op.getLoc(), conv, outPerms, rewriter);
Type newResultType = getTypeConverter()->convertType(op.getType());
if (accumulatorDType != resultDTy) {
Type resultElementType =
cast<RankedTensorType>(newResultType).getElementType();
conv = torch_to_linalg::convertTensorToElementType(rewriter, loc, conv,
resultElementType);
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, conv);
return success();
}
// Special depthwise case: Cin = Cout = groups.
// Note: pytorch considers Cin == groups (Cout possibly a non-zero multiple
// of groups) to be depthwise in their documentation, but the linalg ops
// apparently disagree.
auto inShape = makeShapeTorchCompatible(
cast<RankedTensorType>(input.getType()).getShape());
auto weightShape = makeShapeTorchCompatible(
cast<RankedTensorType>(weight.getType()).getShape());
if (inShape[1] == numGroups && weightShape[0] == numGroups &&
weightShape[1] == 1) {
// Collapse weight shape (C/G == 1)
SmallVector<ReassociationIndices> collapsedDims = {{0, 1}};
SmallVector<int64_t> collapsedShape{weightShape[0] * weightShape[1]};
for (unsigned i = 0; i < numSpatialDims; i++) {
collapsedDims.push_back({i + 2});
collapsedShape.push_back(weightShape[i + 2]);
}
Type collapsedType = RankedTensorType::get(
makeShapeLLVMCompatible(collapsedShape), weightDTy);
Value collapsedWeight = rewriter.create<tensor::CollapseShapeOp>(
loc, collapsedType, weight, collapsedDims);
if (!inputZp) {
switch (numSpatialDims) {
case 1:
conv = rewriter
.create<linalg::DepthwiseConv1DNcwCwOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, collapsedWeight}, outputTensor,
stridesAttr, dilationAttr)
.getResult(0);
break;
case 2:
conv = rewriter
.create<linalg::DepthwiseConv2DNchwChwOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, collapsedWeight}, outputTensor,
stridesAttr, dilationAttr)
.getResult(0);
break;
default:
return rewriter.notifyMatchFailure(
op, "unimplemented: only 1D and 2D depthwise convolution "
"supported for special case of group convolution");
};
} else {
if (numSpatialDims != 2)
return rewriter.notifyMatchFailure(
op, "unimplemented: only 2D depthwise quantized convolution "
"supported for special case of group convolution");
// currently, the only named depthwise qconv op is nhwc_hwc
// input: nchw -> nhwc; weight (collapsed): chw -> hwc
// linalg conv result nhwc -> nchw
// inPerms = [0, 2, 3, 1]
// weightPerms = [1, 2, 0]
// resultPerms = [0, 3, 1, 2]
llvm::SmallVector<int64_t> inPerms, weightPerms, resultPerms;
inPerms.push_back(0);
resultPerms.append({0, static_cast<int64_t>(numSpatialDims + 1)});
for (size_t i = 0; i < numSpatialDims; ++i) {
inPerms.push_back(i + 2);
weightPerms.push_back(i + 1);
resultPerms.push_back(i + 1);
}
inPerms.push_back(1);
weightPerms.push_back(0);
paddedInput =
transposeValue(op.getLoc(), paddedInput, inPerms, rewriter);
collapsedWeight =
transposeValue(op.getLoc(), collapsedWeight, weightPerms, rewriter);
outputTensor =
transposeValue(op.getLoc(), outputTensor, inPerms, rewriter);
conv =
rewriter
.create<linalg::DepthwiseConv2DNhwcHwcQOp>(
loc, outputTensor.getType(),
ValueRange{paddedInput, collapsedWeight, inputZp, weightZp},
outputTensor, stridesAttr, dilationAttr)
.getResult(0);
// convert output nhwc -> nchw
conv = transposeValue(op.getLoc(), conv, resultPerms, rewriter);
}
Type newResultType = getTypeConverter()->convertType(op.getType());
if (accumulatorDType != resultDTy) {
Type resultElementType =
cast<RankedTensorType>(newResultType).getElementType();
conv = torch_to_linalg::convertTensorToElementType(rewriter, loc, conv,
resultElementType);
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, conv);
return success();
}
if (numSpatialDims != 2)
return rewriter.notifyMatchFailure(
op, "unimplemented: only 2D grouped convolution supported");
// Grouped case, use the grouped conv linalg op
auto expandGroups = [&](Value tensor, size_t dim) {
auto inType = cast<RankedTensorType>(tensor.getType());
auto inShape = makeShapeTorchCompatible(inType.getShape());
SmallVector<int64_t> outShape;
for (auto i = 0; i < (long)inShape.size(); i++) {
if (i == 1) {
outShape.push_back(numGroups);
}
if (i == (long)dim) {
outShape.push_back(inShape[i] == kUnknownSize
? kUnknownSize
: inShape[i] / numGroups);
} else {
outShape.push_back(inShape[i]);
}
}
SmallVector<ReassociationIndices> indices;
for (auto i = 0; i <= (long)inShape.size(); i++) {
if (i == (long)dim) {
indices.push_back({i, ++i});
continue;
}
indices.push_back({i});
}
auto retType = inType.clone(makeShapeLLVMCompatible(outShape));
return rewriter.create<tensor::ExpandShapeOp>(loc, retType, tensor,
indices);
};
// expand F,C,H,W -> G,F/G,C,H,W
auto expandWeight = [&](Value tensor) {
auto inType = cast<RankedTensorType>(tensor.getType());
auto inShape = makeShapeTorchCompatible(inType.getShape());
SmallVector<int64_t> outShape{
numGroups,
(inShape[0] == kUnknownSize ? kUnknownSize : inShape[0] / numGroups)};
outShape.append(inShape.begin() + 1, inShape.end());
SmallVector<ReassociationIndices> indices{{0, 1}};
for (auto i = 2; i <= (long)inShape.size(); i++)
indices.push_back({i});
auto retType = inType.clone(makeShapeLLVMCompatible(outShape));
return rewriter.create<tensor::ExpandShapeOp>(loc, retType, tensor,
indices);
};
Value paddedInputExpanded = expandGroups(paddedInput, 1);
Value weightExpanded = expandWeight(weight);
auto expandOutputTensor = expandGroups(outputTensor, 1);
// TODO: add 1D and 3D case
if (!inputZp) {
conv = rewriter
.create<linalg::Conv2DNgchwGfchwOp>(
loc, expandOutputTensor.getResultType(),
ValueRange{paddedInputExpanded, weightExpanded},
expandOutputTensor.getResult(), stridesAttr, dilationAttr)
.getResult(0);
} else {
conv = rewriter
.create<linalg::Conv2DNgchwGfchwQOp>(
loc, expandOutputTensor.getResultType(),
ValueRange{paddedInputExpanded, weightExpanded, inputZp,
weightZp},
expandOutputTensor.getResult(), stridesAttr, dilationAttr)
.getResult(0);
}
conv = rewriter.create<tensor::CollapseShapeOp>(
loc, outputTensor.getType(), conv,
expandOutputTensor.getReassociationIndices());
Type newResultType = getTypeConverter()->convertType(op.getType());
if (accumulatorDType != resultDTy) {
Type resultElementType =
cast<RankedTensorType>(newResultType).getElementType();
conv = torch_to_linalg::convertTensorToElementType(rewriter, loc, conv,
resultElementType);
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, conv);
return success();
}
};
} // namespace
void mlir::torch::torch_to_linalg::populateLinearPatternsAndLegality(
TypeConverter &typeConverter, RewritePatternSet &patterns,
ConversionTarget &target) {
MLIRContext *context = patterns.getContext();
target.addIllegalOp<AtenMmOp>();
patterns.add<ConvertAtenMmOp>(typeConverter, context);
target.addIllegalOp<AtenFlipOp>();
patterns.add<ConvertAtenFlipOp>(typeConverter, context);
target.addIllegalOp<AtenMatmulOp>();
patterns.add<ConvertAtenMatmulOp>(typeConverter, context);
target.addIllegalOp<AtenBmmOp>();
patterns.add<ConvertAtenBmmOp>(typeConverter, context);
target.addIllegalOp<AtenConvolutionOp>();
patterns.add<ConvertAtenConvolutionOp>(typeConverter, context);
}