torch-mlir/lib/Conversion/TorchToTMTensor/TorchToTMTensor.cpp

1654 lines
70 KiB
C++
Raw Normal View History

//===----------------------------------------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
// Also available under a BSD-style license. See LICENSE.
//
//===----------------------------------------------------------------------===//
#include "torch-mlir/Conversion/TorchToTMTensor/TorchToTMTensor.h"
#include "../PassDetail.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/Math/IR/Math.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
2022-09-09 12:36:40 +08:00
#include "mlir/IR/Builders.h"
#include "mlir/IR/BuiltinTypeInterfaces.h"
2022-09-09 12:36:40 +08:00
#include "mlir/IR/BuiltinTypes.h"
#include "mlir/IR/MLIRContext.h"
#include "mlir/IR/Matchers.h"
#include "mlir/IR/ValueRange.h"
#include "torch-mlir-dialects/Dialect/TMTensor/IR/TMTensorDialect.h"
#include "torch-mlir-dialects/Dialect/TMTensor/IR/TMTensorOps.h"
#include "torch-mlir/Conversion/Utils/Utils.h"
#include "torch-mlir/Dialect/Torch/IR/TorchDialect.h"
#include "torch-mlir/Dialect/Torch/IR/TorchOps.h"
#include "torch-mlir/Dialect/Torch/IR/TorchTypes.h"
#include "torch-mlir/Dialect/Torch/Utils/TorchUpstream.h"
#include "torch-mlir/Dialect/Torch/Utils/Utils.h"
#include "torch-mlir/Dialect/TorchConversion/Transforms/BackendTypeConversion.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/APInt.h"
#include "llvm/Support/ErrorHandling.h"
using namespace mlir;
using namespace mlir::torch;
using namespace mlir::torch::Torch;
using namespace mlir::torch::TorchConversion;
using namespace mlir::torch::TMTensor;
// -----------------------------------------------------------------------------
// Patterns (as this grows, it should be organized into multiple files)
// -----------------------------------------------------------------------------
// This is going to eventually be O(#aten ops), which is in the 100s.
//
// Most of these patterns consist of:
// 1. Checking that the operand/result types and other static properties are
// good-enough to create a valid linalg op (such as operands being of
// ranks/dtypes acceptable to the linalg op).
// 2. Creating dynamic error guards, usually checking a predicate on the
// compatibility of operand shapes.
// 3. Creating init tensors for the computation op. Usually this involves
// reifying IR for a shape transfer function based on the operand shapes.
// 4. Creating a named linalg op to replace the original op.
//
// TODO: Use linalg OpDSL to autogenerate at least 1)/2)/3) such
// that these patterns become mostly mechanical associations of
// "aten.foo -> linalg.foo".
static TypedAttr getNumericLimit(PatternRewriter &rewriter, Type elementType,
bool getMin = true) {
auto bitWidth = elementType.getIntOrFloatBitWidth();
if (llvm::isa<mlir::IntegerType>(elementType)) {
if (getMin) {
return rewriter.getIntegerAttr(elementType,
APInt::getSignedMinValue(bitWidth));
} else {
return rewriter.getIntegerAttr(elementType,
APInt::getSignedMaxValue(bitWidth));
}
} else if (mlir::FloatType floatType =
llvm::dyn_cast<mlir::FloatType>(elementType)) {
return rewriter.getFloatAttr(
elementType,
APFloat::getLargest(floatType.getFloatSemantics(), getMin));
} else {
llvm_unreachable("Only float/integer types are supported!");
}
}
// This function will reformat the `index` and `src` from torch operations
// like `torch.scatter` or `torch.scatter_reduce` to match the expected
// input for the TMScatterOp. It will return the reformated `index` and `src`
// as a pair of mlir::Value that can be used as inputs for the TMScatterOp.
static std::pair<Value, Value>
convertTorchScatterIndexAndSrcToTMScatterIndexAndSrc(PatternRewriter &rewriter,
Value indices, Value src,
int64_t dim) {
// Get information on types for inputs
RankedTensorType indexType = indices.getType().cast<RankedTensorType>();
RankedTensorType srcSelf = src.getType().cast<RankedTensorType>();
// Store location for insertions
Location loc = src.getLoc();
Value indexSize = getTensorSize(rewriter, loc, indices);
indexSize = castIntToIndex(rewriter, loc, indexSize);
SmallVector<Value> indexShape = getTensorSizes(rewriter, loc, indices);
Value cstOne = rewriter.create<arith::ConstantIndexOp>(loc, 1);
// We flatten the `src` values from (i, j, k, ...) -> (i * j * k * ...)
SmallVector<Value> indSliceShape({indexSize, cstOne});
Value indSlice =
createZeroInitTensor(rewriter, loc, indSliceShape, rewriter.getI32Type());
// New output shape will be equal to the product of the dimensions of the
// updates
SmallVector<Value> outputs(indexType.getRank(), indSlice);
outputs.push_back(createZeroInitTensor(rewriter, loc, {indexSize},
srcSelf.getElementType()));
SmallVector<Type> outputsType(indexType.getRank(), indSlice.getType());
outputsType.push_back(outputs[indexType.getRank()].getType());
// Create mapping over flattened iteration space
SmallVector<AffineExpr> indSliceExpr = {rewriter.getAffineDimExpr(0),
rewriter.getAffineConstantExpr(0)};
SmallVector<AffineMap> mapping(
indexType.getRank(), AffineMap::get(/*dimCount=*/1, /*symbolCount=*/0,
indSliceExpr, src.getContext()));
// Mapping for updates
mapping.push_back(rewriter.getDimIdentityMap());
SmallVector<utils::IteratorType> iteratorTypes(
{utils::IteratorType::parallel});
// This function goes over the flattened iteration space of the `indices`
// and `src`. It will reconstruct the original induction variables based
// on the current flattened index. The flattened iteration space is required
// because TMTensorScatterOp expects a list of single element updates.
auto flattenedUpdates =
rewriter
.create<linalg::GenericOp>(
loc, outputsType, ValueRange(), outputs, mapping, iteratorTypes,
[&](OpBuilder &b, Location loc, ValueRange args) {
SmallVector<Value> indexValues(indexType.getRank());
Value ind = b.create<linalg::IndexOp>(loc, 0);
for (int i = indexType.getRank() - 1; i >= 0; i--) {
indexValues[i] =
b.create<arith::RemSIOp>(loc, ind, indexShape[i]);
ind = b.create<arith::DivSIOp>(loc, ind, indexShape[i]);
}
// Extract the scatter index and update value
Value extractIndexValue =
b.create<tensor::ExtractOp>(loc, indices, indexValues);
Value extractSrcValue =
b.create<tensor::ExtractOp>(loc, src, indexValues);
SmallVector<Value> yieldVals;
for (Value v : indexValues) {
Value scalar = castIndexToInt64(b, loc, v);
yieldVals.push_back(b.create<arith::TruncIOp>(
loc, rewriter.getI32Type(), scalar));
}
// Replace the original index with the index specified
// by the scatter.
yieldVals[dim] = b.create<arith::TruncIOp>(
loc, rewriter.getI32Type(), extractIndexValue);
yieldVals.push_back(extractSrcValue);
b.create<linalg::YieldOp>(loc, yieldVals);
})
.getResultTensors();
auto toOpFoldResult = [](Value v) -> OpFoldResult {
auto op = v.getDefiningOp<arith::ConstantIndexOp>();
if (!op)
return v;
return op.getValue();
};
// The result of the linalg::Generic operation gives us (rank(`src`) + 1)
// 1D-tensors where each contains a number of elements equal to the total
// number of elements in the `src` tensor. The indices must now be
// constructed by concatanating the first rank(`src`) tensors together. The
// new `src` tensor is the last tensor returned from the linalg::Generic
// operation.
SmallVector<Value> offsets = {
rewriter.create<arith::ConstantIndexOp>(loc, 0),
rewriter.create<arith::ConstantIndexOp>(loc, 0)};
SmallVector<Value> strides = {
rewriter.create<arith::ConstantIndexOp>(loc, 1),
rewriter.create<arith::ConstantIndexOp>(loc, 1)};
Value indicesRank =
rewriter.create<arith::ConstantIndexOp>(loc, indexType.getRank());
Value flattenedIndices = createZeroInitTensor(
rewriter, loc, SmallVector<Value>({indexSize, indicesRank}),
rewriter.getI32Type());
SmallVector<Value> scatterInputsVector(flattenedUpdates);
for (auto const slice : ArrayRef(scatterInputsVector).drop_back()) {
SmallVector<Value> sizes = getTensorSizes(rewriter, loc, slice);
flattenedIndices = rewriter.createOrFold<tensor::InsertSliceOp>(
loc, slice, flattenedIndices,
llvm::to_vector(llvm::map_range(offsets, toOpFoldResult)),
llvm::to_vector(llvm::map_range(sizes, toOpFoldResult)),
llvm::to_vector(llvm::map_range(strides, toOpFoldResult)));
// Increment offset to insert into next column
offsets[1] = rewriter.createOrFold<arith::AddIOp>(loc, offsets[1], cstOne);
}
return std::make_pair(flattenedIndices,
scatterInputsVector[indexType.getRank()]);
}
static llvm::SmallVector<int64_t> createDefaultDimMap(Value indices) {
llvm::SmallVector<int64_t> dmap;
if (auto iTy = dyn_cast<BaseTensorType>(indices.getType()))
dmap.resize(iTy.getSizes()[1]);
if (auto iTy = dyn_cast<RankedTensorType>(indices.getType()))
dmap.resize(iTy.getDimSize(1));
for (int i = 0, s = dmap.size(); i < s; ++i)
dmap[i] = i;
return dmap;
}
static Value createTMTensorScatterOp(
OpBuilder &b, Location loc, Value updates, Value indices, Value original,
llvm::ArrayRef<int64_t> dimensionsMap, bool uniqueIndices,
function_ref<void(OpBuilder &, Location, Value, Value)> bodyBuild) {
auto dimensionsMapAttr = b.getDenseI64ArrayAttr(dimensionsMap);
auto originalTensorType = original.getType().cast<RankedTensorType>();
Type originalElementType = originalTensorType.getElementType();
auto scatterOp = b.create<TMTensor::ScatterOp>(
loc, originalTensorType, ValueRange{updates, indices},
ValueRange{original}, dimensionsMapAttr, uniqueIndices);
Region &scatterOpRegion = scatterOp.getRegion();
auto &scatterOpBlock = scatterOpRegion.emplaceBlock();
scatterOpBlock.addArguments({originalElementType, originalElementType},
{loc, loc});
OpBuilder regionBuilder(scatterOpRegion);
auto blockArgs = scatterOpBlock.getArguments();
Value updatesElement = blockArgs[0];
Value originalElement = blockArgs[1];
bodyBuild(regionBuilder, loc, updatesElement, originalElement);
return scatterOp->getResult(0);
}
2022-09-09 12:36:40 +08:00
static Value createTMTensorScanOp(
OpBuilder &b, Location loc, Value input, Value output, Value accumulator,
int64_t dim, bool inclusive,
function_ref<void(OpBuilder &, Location, Value, Value)> bodyBuild) {
auto inputType = input.getType().cast<RankedTensorType>();
auto accType = accumulator.getType().cast<RankedTensorType>();
Type elementType = inputType.getElementType();
auto scanOp = b.create<TMTensor::ScanOp>(
loc, TypeRange{inputType, accType}, input,
ValueRange{output, accumulator}, b.getI64IntegerAttr(dim),
b.getBoolAttr(inclusive));
Region &scanOpRegion = scanOp.getRegion();
2022-09-09 12:36:40 +08:00
auto &scanOpBlock = scanOpRegion.emplaceBlock();
scanOpBlock.addArguments({elementType, elementType}, {loc, loc});
OpBuilder regionBuilder(scanOpRegion);
auto blockArgs = scanOpBlock.getArguments();
Value inputElement = blockArgs[0];
Value accElement = blockArgs[1];
bodyBuild(regionBuilder, loc, inputElement, accElement);
return scanOp->getResult(0);
}
// Utility function to create a TMTensor::SortOp.
static FailureOr<SmallVector<Value>>
createTMTensorSortOp(PatternRewriter &rewriter, Location sortOpLoc,
llvm::ArrayRef<Value> operands,
llvm::ArrayRef<Type> elementTypes, int64_t dimension,
bool isStable, bool isDescending) {
// Step 1. Create TMTensor::SortOp structure.
SmallVector<Type> sortResultTypes;
for (Value val : operands) {
sortResultTypes.push_back(val.getType());
}
ValueRange inputs;
auto sortOp = rewriter.create<TMTensor::SortOp>(
sortOpLoc, sortResultTypes, inputs, operands,
rewriter.getI64IntegerAttr(dimension));
// Step 2. Add two arguments for each element type in the SortOp's block.
Region *body = &sortOp.getRegion();
Block *block = rewriter.createBlock(body);
Location loc = body->getLoc();
for (Type elementType : elementTypes) {
block->addArguments({elementType, elementType},
SmallVector<Location, 2>(2, loc));
}
// Step 3. Create comparison op which will be used as the sorting predicate.
Value compareOp;
if (auto intType = elementTypes[0].dyn_cast<mlir::IntegerType>()) {
// Case for using arith::CmpIOp.
arith::CmpIPredicate ge = arith::CmpIPredicate::sge;
arith::CmpIPredicate le = arith::CmpIPredicate::sle;
if (intType.isUnsignedInteger()) {
ge = arith::CmpIPredicate::uge;
le = arith::CmpIPredicate::ule;
}
arith::CmpIPredicate predicate = isDescending ? ge : le;
compareOp = rewriter.create<arith::CmpIOp>(
loc, predicate, block->getArgument(0), block->getArgument(1));
} else if (elementTypes[0].isa<mlir::FloatType>()) {
// Case for using arith::CmpFOp.
arith::CmpFPredicate predicate =
isDescending ? arith::CmpFPredicate::OGE : arith::CmpFPredicate::OLE;
compareOp = rewriter.create<arith::CmpFOp>(
loc, predicate, block->getArgument(0), block->getArgument(1));
} else {
return rewriter.notifyMatchFailure(
sortOpLoc, "Only Integer and Floating element type expected.");
}
// Step 4. Create yield op for yielding the sorting predicate.
rewriter.create<TMTensor::YieldOp>(loc, compareOp);
return SmallVector<Value>(sortOp.getResults());
}
namespace {
class ConvertAtenScatterSrcOp : public OpConversionPattern<AtenScatterSrcOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenScatterSrcOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op.getLoc();
const TypeConverter *typeConverter = getTypeConverter();
Value self = adaptor.getSelf();
Value index = adaptor.getIndex();
Value src = adaptor.getSrc();
RankedTensorType selfType = self.getType().cast<RankedTensorType>();
RankedTensorType indexType = index.getType().cast<RankedTensorType>();
RankedTensorType srcType = src.getType().cast<RankedTensorType>();
if (selfType.getRank() != indexType.getRank() ||
indexType.getRank() != srcType.getRank())
return rewriter.notifyMatchFailure(op,
"'self', 'index' and 'src' should all"
"have the same number of dimensions.");
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(op,
"unimplemented: dim is not constant");
// Get the inputs reformatted for the TMScatterOp
auto [indices, updates] =
convertTorchScatterIndexAndSrcToTMScatterIndexAndSrc(rewriter, index,
src, dim);
Value scatterOp = createTMTensorScatterOp(
rewriter, loc, updates, indices, self,
/*dimensionsMap=*/createDefaultDimMap(indices), /*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value updatesElement,
Value inputElement) {
b.create<TMTensor::YieldOp>(loc, updatesElement);
});
auto resultType = typeConverter->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, scatterOp);
return success();
}
};
} // namespace
namespace {
// aten::bincount op counts the frequency of each value in a 1-d input tensor of
// non-negative ints.
class ConvertAtenBincountOp : public OpConversionPattern<AtenBincountOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenBincountOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op.getLoc();
MLIRContext *context = op->getContext();
const TypeConverter *typeConverter = getTypeConverter();
Value input = adaptor.getSelf();
Value torchTypeInput = op.getSelf();
Value minlength = adaptor.getMinlength();
Value weights = adaptor.getWeights();
// TODO: Add a check to verify that the input tensor elements are all
// non-negative.
// Check whether the input is a 1-d tensor of integer type or not.
RankedTensorType inputType = input.getType().cast<RankedTensorType>();
if (inputType.getRank() != 1 ||
!inputType.getElementType().isa<mlir::IntegerType>())
return rewriter.notifyMatchFailure(
op,
"Input tensor has to be a one-dimensional tensor of integer type.");
// Check whether the input tensor element type is i64 or not.
IntegerType inputIntegerType =
inputType.getElementType().cast<IntegerType>();
if (inputIntegerType.getWidth() != 64)
return rewriter.notifyMatchFailure(
op,
"Unimplemented: Integer width not equal to 64 are not supported.");
// TODO: Incorporate the weight argument.
if (!weights.getType().isa<mlir::torch::Torch::NoneType>())
return rewriter.notifyMatchFailure(
op, "Unimplemented: the weights operand is not incorporated.");
// Finding the maximum value in the input tensor.
SmallVector<int64_t> maxTensorSizes;
ValueTensorType maxTensorType = ValueTensorType::get(
context, llvm::ArrayRef(maxTensorSizes),
torchTypeInput.getType().cast<ValueTensorType>().getDtype());
Value maxTensor =
rewriter.create<AtenMaxOp>(loc, maxTensorType, torchTypeInput);
maxTensor = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(maxTensor.getType()),
maxTensor);
// `maxTensor` is a 0-d tensor, extracting its only element and
// storing it in `maxInput`.
Value maxInput = rewriter.create<tensor::ExtractOp>(loc, maxTensor);
// Creating a tm_tensor.scatter op with the following mapping:
// 1.) `input` tensor maps to the indices in scatter op. `input` is
// expanded from 1-d to 2-d, and its element type is set to i32 as required
// for the scatter op.
// 2.) `updates` is a 1-d dummy tensor with the size equivalent to the
// `input`.
// 3.) `bincount` a 1-d tensor maps to the original in scatter op
// with size equal to the max(max(input) + 1, minlength).
SmallVector<int64_t> expandedInputSizes{
makeShapeTorchCompatible(inputType.getShape())[0], 1};
ValueTensorType expandInputType = ValueTensorType::get(
context, llvm::ArrayRef(expandedInputSizes),
torchTypeInput.getType().cast<ValueTensorType>().getDtype());
Value torchCstOne = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(1));
Value expandedInputTensor = rewriter.create<AtenUnsqueezeOp>(
loc, expandInputType, torchTypeInput, torchCstOne);
// Converting the input element type to i32.
Value indices = convertTensorToDtype(
rewriter, loc, expandedInputTensor,
mlir::IntegerType::get(context, 32, mlir::IntegerType::Signed));
indices = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(indices.getType()), indices);
auto resultType = typeConverter->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
Type resultElemType = resultType.getElementType();
SmallVector<Value, 1> inputSizeDynamic =
getTensorSizesUntilDim(rewriter, loc, input, 0);
Value updatesTensor = rewriter.create<tensor::EmptyOp>(
loc, getAsOpFoldResult(inputSizeDynamic), resultElemType);
Value constantZero = rewriter.create<arith::ConstantOp>(
loc, rewriter.getZeroAttr(resultElemType));
Value constantOne = rewriter.create<arith::ConstantIntOp>(
loc, 1, resultElemType.getIntOrFloatBitWidth());
// Bincount size = max(max(input) + 1, minlength)
Value maxInputPlusOne =
rewriter.create<arith::AddIOp>(loc, maxInput, constantOne);
Value bincountSize =
rewriter.create<arith::MaxSIOp>(loc, maxInputPlusOne, minlength);
bincountSize = castIntToIndex(rewriter, loc, bincountSize);
Value bincountTensor = createInitTensor(rewriter, loc, {bincountSize},
resultElemType, constantZero);
Value scatterOp = createTMTensorScatterOp(
rewriter, loc, updatesTensor, indices, bincountTensor,
/*dimensionsMap=*/createDefaultDimMap(indices), /*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value _, Value bincountElem) {
Value add = b.create<arith::AddIOp>(loc, bincountElem, constantOne);
b.create<TMTensor::YieldOp>(loc, add);
});
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, scatterOp);
return success();
}
};
} // namespace
namespace {
Value combinePutIndices(Location loc, llvm::ArrayRef<Value> indicesRef,
OpBuilder b) {
llvm::SmallVector<Value> indices(indicesRef);
// Declare commonly used constants up front:
Value torchCstZero =
b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(0));
Value torchCstOne =
b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(1));
Value torchCstNegOne =
b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(-1));
// Determine the broadcast sizes and materialize missing implicit end
// dimensions:
int64_t indicesRank = 0;
for (auto index : indices) {
auto indexTy = cast<Torch::ValueTensorType>(index.getType());
int64_t rank = indexTy.getSizes().size();
indicesRank = std::max(rank, indicesRank);
}
auto maxDim = [](int64_t dim0, int64_t dim1) {
if (dim0 == Torch::kUnknownSize || dim1 == Torch::kUnknownSize)
return Torch::kUnknownSize;
return std::max(dim0, dim1);
};
llvm::SmallVector<Value> broadcastSizes(indicesRank, torchCstOne);
llvm::SmallVector<int64_t> broadcastShape(indicesRank, 0);
for (auto index : indices) {
auto indexTy = cast<Torch::ValueTensorType>(index.getType());
auto shape = indexTy.getSizes();
int32_t rank = shape.size();
for (int32_t j = 0; j < rank; ++j) {
Value dim = b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(j));
auto sizeOp = b.create<Torch::AtenSizeIntOp>(loc, index, dim);
auto size = shape[j];
int32_t idx = broadcastShape.size() - rank + j;
broadcastSizes[idx] =
b.create<Torch::PrimMaxIntOp>(loc, sizeOp, broadcastSizes[idx]);
broadcastShape[idx] = maxDim(size, broadcastShape[idx]);
}
}
auto mulDim = [](int64_t dim0, int64_t dim1) {
if (dim0 == Torch::kUnknownSize || dim1 == Torch::kUnknownSize)
return Torch::kUnknownSize;
return dim0 * dim1;
};
int64_t scatterBatchCount = 1;
for (auto dim : broadcastShape) {
scatterBatchCount = mulDim(scatterBatchCount, dim);
}
// Broadcast together and flatten to batch values:
Value broadcastSizeList = b.create<PrimListConstructOp>(
loc, Torch::ListType::get(b.getType<Torch::IntType>()), broadcastSizes);
for (Value &index : indices) {
auto indexTy = cast<Torch::ValueTensorType>(index.getType());
auto expandTy = b.getType<Torch::ValueTensorType>(
broadcastShape, indexTy.getOptionalDtype());
index = b.create<Torch::AtenBroadcastToOp>(loc, expandTy, index,
broadcastSizeList);
auto flattenTy = b.getType<Torch::ValueTensorType>(
scatterBatchCount, indexTy.getOptionalDtype());
index = b.create<Torch::AtenFlattenUsingIntsOp>(
loc, flattenTy, index, torchCstZero, torchCstNegOne);
}
// Unsqueeze so we have a 1 dim to concat along:
for (Value &tensor : indices) {
auto btt = cast<Torch::BaseTensorType>(tensor.getType());
if (!btt.hasSizes())
return nullptr;
llvm::SmallVector<int64_t> shape(btt.getSizes());
shape.push_back(1);
auto unsqueezeTy = b.getType<Torch::ValueTensorType>(shape, btt.getDtype());
Value unsqueezed =
b.create<AtenUnsqueezeOp>(loc, unsqueezeTy, tensor, torchCstOne);
tensor = unsqueezed;
}
BaseTensorType unsqueezedTensorType =
indices[0].getType().cast<BaseTensorType>();
Value indicesTorchList = b.create<PrimListConstructOp>(
loc, Torch::ListType::get(unsqueezedTensorType), indices);
llvm::SmallVector<int64_t, 2> concatShape{
unsqueezedTensorType.getSizes()[0], static_cast<int64_t>(indices.size())};
ValueTensorType concatIndicesType = b.getType<ValueTensorType>(
llvm::ArrayRef(concatShape), unsqueezedTensorType.getDtype());
return b.create<AtenCatOp>(loc, concatIndicesType, indicesTorchList,
torchCstOne);
}
// Helper that collapses the batch dimensions together and moves it to the front
// of the array.
static Value collapseAndMoveBatchDims(Location loc, Value values, int64_t batch,
int64_t count, OpBuilder b) {
if (batch == 0 && count == 1)
return values;
auto valuesTy = cast<Torch::ValueTensorType>(values.getType());
auto inShape = valuesTy.getSizes();
llvm::SmallVector<int64_t> outShape;
llvm::SmallVector<Value> outDims;
// We need a length-1 dim at the start to transpose the batch to:
if (batch != 0) {
outDims.push_back(b.create<Torch::ConstantIntOp>(loc, 1));
outShape.push_back(1);
}
// Dimensions before the batch stay the same:
for (int i = 0; i <= batch; i++) {
auto k = b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(i));
auto dim = b.create<Torch::AtenSizeIntOp>(loc, values, k);
outDims.push_back(dim);
outShape.push_back(inShape[i]);
}
auto mulI = [](int64_t dim0, int64_t dim1) {
if (dim0 == Torch::kUnknownSize || dim1 == Torch::kUnknownSize)
return Torch::kUnknownSize;
return dim0 * dim1;
};
// Determine the collapse size of the batch dimension:
for (int i = 1; i < count; i++) {
outShape.back() = mulI(outShape.back(), inShape[batch + i]);
auto k =
b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(batch + i));
auto dim = b.create<Torch::AtenSizeIntOp>(loc, values, k);
outDims.back() = b.create<Torch::AtenMulIntOp>(loc, dim, outDims.back());
}
// Add the dimensions after the batch dims:
for (int i = batch + count, s = inShape.size(); i < s; ++i) {
auto k = b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(i));
auto dim = b.create<Torch::AtenSizeIntOp>(loc, values, k);
outDims.push_back(dim);
outShape.push_back(inShape[i]);
}
Value outDimsList = b.create<PrimListConstructOp>(
loc, Torch::ListType::get(b.getType<Torch::IntType>()), outDims);
valuesTy =
b.getType<Torch::ValueTensorType>(outShape, valuesTy.getOptionalDtype());
values = b.create<AtenViewOp>(loc, valuesTy, values, outDimsList);
if (batch == 0)
return values;
// Batch is already at the front, no need to transpose:
std::swap(outDims[0], outDims[batch + 1]);
std::swap(outShape[0], outShape[batch + 1]);
Value dim0 = b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(0));
Value dimB =
b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(batch + 1));
valuesTy =
b.getType<Torch::ValueTensorType>(outShape, valuesTy.getOptionalDtype());
values =
b.create<Torch::AtenTransposeIntOp>(loc, valuesTy, values, dim0, dimB);
outDims.clear();
outShape.clear();
auto transposeShape = valuesTy.getSizes();
int64_t transposeRank = transposeShape.size();
for (int i = 0; i < transposeRank; ++i) {
if (i == batch + 1)
continue;
Value k = b.create<Torch::ConstantIntOp>(loc, b.getI64IntegerAttr(i));
outDims.push_back(b.create<AtenSizeIntOp>(loc, values, k));
outShape.push_back(transposeShape[i]);
}
valuesTy =
b.getType<Torch::ValueTensorType>(outShape, valuesTy.getOptionalDtype());
outDimsList = b.create<PrimListConstructOp>(
loc, Torch::ListType::get(b.getType<Torch::IntType>()), outDims);
return b.create<AtenViewOp>(loc, valuesTy, values, outDimsList);
}
class ConvertAten_IndexPutImplOp
: public OpConversionPattern<Aten_IndexPutImplOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(Aten_IndexPutImplOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op.getLoc();
MLIRContext *context = op->getContext();
Value input = op.getSelf();
Value values = op.getValues();
auto inputType = cast<ValueTensorType>(input.getType());
auto valuesType = cast<ValueTensorType>(values.getType());
int64_t inputRank = inputType.getSizes().size();
auto valuesTensorType = op.getValues().getType().cast<BaseTensorType>();
auto resultType = typeConverter->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
if (!valuesTensorType.hasSizes())
return rewriter.notifyMatchFailure(
op, "unimplemented: the values tensor type must have sizes.");
// The unsafe should be either `False` or `none`.
if (!op.getUnsafe().getType().isa<Torch::NoneType>()) {
bool unsafe;
if (!matchPattern(op.getUnsafe(), m_TorchConstantBool(&unsafe)))
return rewriter.notifyMatchFailure(
op, "unimplemented: unsafe must be a constant");
else if (unsafe)
return rewriter.notifyMatchFailure(
op, "unimplemented: unsafe is expected to be false");
}
// The accumulate should be a torch constant of boolean type.
bool accumulate;
if (!matchPattern(op.getAccumulate(), m_TorchConstantBool(&accumulate)))
return rewriter.notifyMatchFailure(
op, "Expected accumulate to be constant bool.");
// The element type of the `input` and `values` should be same.
if (inputType.getDtype() != valuesType.getDtype())
return rewriter.notifyMatchFailure(
op, "Input element type should be same as the values element type.");
SmallVector<Value> optionalIndicesList;
getListConstructElements(op.getIndices(), optionalIndicesList);
int64_t optionalIndicesCount = optionalIndicesList.size();
// The size of the list of the index tensors should not be greater than the
// input rank.
if (optionalIndicesCount > inputRank)
return rewriter.notifyMatchFailure(
op, "Indices list size should not be greater than the input rank.");
if (optionalIndicesCount == 0)
return rewriter.notifyMatchFailure(op, "Indices list must not be empty.");
// Filter to available indices and get the indicesMap:
SmallVector<Value> indicesList;
SmallVector<int64_t> indicesMap;
int64_t numBatchDims = 0;
for (int i = 0, s = optionalIndicesList.size(); i < s; ++i) {
if (isa<Torch::NoneType>(optionalIndicesList[i].getType()))
continue;
indicesList.push_back(optionalIndicesList[i]);
indicesMap.push_back(i);
auto indexTy = cast<ValueTensorType>(indicesList.back().getType());
numBatchDims = std::max(static_cast<int64_t>(indexTy.getSizes().size()),
numBatchDims);
}
// Value broadcasting semantics require batch dimensions to be up front if
// the indices are not sequential, otherwise they are sequentially at their
// location:
int64_t batchDim = 0;
for (int s = optionalIndicesList.size(); batchDim < s; ++batchDim)
if (!isa<Torch::NoneType>(optionalIndicesList[batchDim].getType()))
break;
int64_t nextNone = batchDim;
for (int s = optionalIndicesList.size(); nextNone < s; ++nextNone)
if (isa<Torch::NoneType>(optionalIndicesList[nextNone].getType()))
break;
for (int s = optionalIndicesList.size(); nextNone < s; ++nextNone)
if (!isa<Torch::NoneType>(optionalIndicesList[nextNone].getType()))
batchDim = 0;
// Indices are extended, catted, and collapsed into a [batch, depth] tensor:
Value indices = combinePutIndices(loc, indicesList, rewriter);
// Bove batch dimensions to the front and collapse into a single dim:
values =
collapseAndMoveBatchDims(loc, values, batchDim, numBatchDims, rewriter);
valuesType = cast<Torch::ValueTensorType>(values.getType());
// Materialize out the length-1 dimensions:
Value zero = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(0));
Value one = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(1));
llvm::SmallVector<int64_t> valuesShape;
llvm::SmallVector<Value> valuesDims;
int vDim = 0;
if (optionalIndicesCount + valuesType.getSizes().size() >
inputType.getSizes().size()) {
valuesShape.push_back(valuesType.getSizes().front());
valuesDims.push_back(
rewriter.create<Torch::AtenSizeIntOp>(loc, values, zero));
vDim++;
}
for (int i = 0, s = inputType.getSizes().size(); i < s; ++i) {
if (i < optionalIndicesCount &&
!isa<Torch::NoneType>(optionalIndicesList[i].getType())) {
valuesDims.push_back(one);
valuesShape.push_back(1);
continue;
}
Value k = rewriter.create<Torch::ConstantIntOp>(
loc, rewriter.getI64IntegerAttr(vDim));
valuesDims.push_back(
rewriter.create<Torch::AtenSizeIntOp>(loc, values, k));
valuesShape.push_back(inputType.getSizes()[i]);
vDim++;
}
Value valuesDimsList = rewriter.create<PrimListConstructOp>(
loc, Torch::ListType::get(rewriter.getType<Torch::IntType>()),
valuesDims);
valuesType = rewriter.getType<Torch::ValueTensorType>(
valuesShape, valuesType.getOptionalDtype());
values =
rewriter.create<AtenViewOp>(loc, valuesType, values, valuesDimsList);
// `TMTensor::ScatterOp` expects indices of element type i32.
indices = convertTensorToDtype(
rewriter, loc, indices,
mlir::IntegerType::get(context, 32, mlir::IntegerType::Signed));
input = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(input.getType()), input);
values = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(values.getType()), values);
indices = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(indices.getType()), indices);
// Creating a tm_tensor.scatter op with the following mapping:
// 1.) Index tensor from the `indicesList` maps to the indices in scatter
// op.
// 2.) `values` is mapped to `updates` in scatter op.
// 3.) `input` is mapped to `original` in scatter op.
bool invalidInputTypeFound = false;
Value scatterOp = createTMTensorScatterOp(
rewriter, loc, values, indices, input, indicesMap,
/*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value valuesElement,
Value inputElement) {
Value yieldValue = valuesElement;
if (accumulate) {
if (inputElement.getType().isa<mlir::IntegerType>()) {
yieldValue =
b.create<arith::AddIOp>(loc, inputElement, valuesElement);
} else if (inputElement.getType().isa<mlir::FloatType>()) {
yieldValue =
b.create<arith::AddFOp>(loc, inputElement, valuesElement);
} else {
invalidInputTypeFound = true;
return;
}
}
b.create<TMTensor::YieldOp>(loc, yieldValue);
});
if (invalidInputTypeFound) {
return rewriter.notifyMatchFailure(
op,
"unimplemented: input tensor must be of integer type or float type");
}
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, scatterOp);
return success();
}
};
} // namespace
namespace {
// The original implementation of the op is as follows:
//
// Indices and GradOutput Layout: [N, C, H, W] or [C, H, W]
// Input Layout: [N, C, Hin, Win] or [C, Hin, Win]
//
// for i in range(N):
// for j in range(C):
// for k in range(H):
// for l in range(W):
// index = indices[i, j, k, l]
// result[i, j, index/Win, index%Win] += gradOutput[i, j, k, l]
//
// OR
//
// for i in range(C):
// for j in range(H):
// for k in range(W):
// index = indices[i, j, k]
// result[i, index/Win, index%Win] += gradOutput[i, j, k]
//
class ConvertAtenMaxPool2dWithIndicesBackwardOp
: public OpConversionPattern<AtenMaxPool2dWithIndicesBackwardOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenMaxPool2dWithIndicesBackwardOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op.getLoc();
MLIRContext *context = op->getContext();
Value gradOutput = adaptor.getGradOutput();
Value input = adaptor.getSelf();
RankedTensorType gradOutputType =
gradOutput.getType().cast<RankedTensorType>();
Type gradOutputElemType = gradOutputType.getElementType();
RankedTensorType inputType = input.getType().cast<RankedTensorType>();
Type inputElemType = inputType.getElementType();
int64_t tensorOperandRank = inputType.getRank();
// `TMTensor::ScatterOp` expects indices of element type i32.
Value indices = convertTensorToDtype(
rewriter, loc, op.getIndices(),
mlir::IntegerType::get(context, 32, mlir::IntegerType::Signed));
indices = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(indices.getType()), indices);
RankedTensorType indicesType = indices.getType().cast<RankedTensorType>();
Type indicesElemType = indicesType.getElementType();
// The element type of the `input` and `grad_output` should be same.
if (inputElemType != gradOutputElemType)
return rewriter.notifyMatchFailure(
op,
"Input element type should be same as the grad_output element type.");
// Since the scatter op requires indices to be a 2-d tensor, we create a new
// 5-d/4-d tensor (depending on the original indices layout) comprising the
// index values. We will collapse this tensor into a 2-d tensor. The
// algorithm for the creation of updated indices tensor is as follows:
//
// for i in range(N):
// for j in range(C):
// for k in range(H):
// for l in range(W):
// for m in range(4):
// if m == 0:
// updatedIndices[N][C][H][W][0] = i
// if m == 1:
// updatedIndices[N][C][H][W][1] = j
// if m == 2:
// updatedIndices[N][C][H][W][2] =
// originalIndices[i, j, k, l] / Win
// if m == 3:
// updatedIndices[N][C][H][W][3] =
// originalIndices[i, j, k, l] % Win
//
// OR
//
// for j in range(C):
// for k in range(H):
// for l in range(W):
// for m in range(3):
// if m == 0:
// updatedIndices[C][H][W][0] = i
// if m == 1:
// updatedIndices[C][H][W][1] = originalIndices[i, j, k, l] / Win
// if m == 2:
// updatedIndices[C][H][W][2] = originalIndices[i, j, k, l] % Win
SmallVector<Value> inputShape = getTensorSizes(rewriter, loc, input);
SmallVector<AffineExpr> originalIndicesDimExprs, updatedIndicesDimExprs;
for (int64_t i = 0; i < tensorOperandRank; i++) {
originalIndicesDimExprs.push_back(rewriter.getAffineDimExpr(i));
updatedIndicesDimExprs.push_back(rewriter.getAffineDimExpr(i));
}
updatedIndicesDimExprs.push_back(
rewriter.getAffineDimExpr(tensorOperandRank));
SmallVector<AffineMap> indexingMaps = AffineMap::inferFromExprList(
{originalIndicesDimExprs, updatedIndicesDimExprs},
rewriter.getContext());
SmallVector<utils::IteratorType> iteratorTypes(
tensorOperandRank + 1, utils::IteratorType::parallel);
SmallVector<OpFoldResult> updatedIndicesShape =
getAsOpFoldResult(getTensorSizes(rewriter, loc, indices));
updatedIndicesShape.push_back(rewriter.getIndexAttr(tensorOperandRank));
Value initTensor = rewriter.create<tensor::EmptyOp>(
loc, updatedIndicesShape, indicesElemType);
Value wIn = inputShape[tensorOperandRank - 1];
SmallVector<Value> cstValues;
for (int64_t i = 0; i < tensorOperandRank; i++)
cstValues.push_back(rewriter.create<arith::ConstantIndexOp>(loc, i));
Value updatedIndices =
rewriter
.create<linalg::GenericOp>(
loc, initTensor.getType(), indices, initTensor, indexingMaps,
iteratorTypes,
[tensorOperandRank, wIn, cstValues,
indicesElemType](OpBuilder &b, Location loc, ValueRange args) {
Value index = castIntToIndex(b, loc, args[0]);
Value updatedIndex = cstValues[0];
Value lastDim =
b.create<linalg::IndexOp>(loc, tensorOperandRank);
for (int64_t i = tensorOperandRank - 1; i >= 0; i--) {
Value result;
if (i == tensorOperandRank - 1)
result = b.create<arith::RemSIOp>(loc, index, wIn);
if (i == tensorOperandRank - 2)
result = b.create<arith::FloorDivSIOp>(loc, index, wIn);
if (i == tensorOperandRank - 3 ||
i == tensorOperandRank - 4)
result = b.create<linalg::IndexOp>(loc, i);
Value pred = b.create<arith::CmpIOp>(
loc, arith::CmpIPredicate::eq, lastDim, cstValues[i]);
Value addAmount = b.create<arith::SelectOp>(
loc, pred, result, cstValues[0]);
updatedIndex =
b.create<arith::AddIOp>(loc, updatedIndex, addAmount);
}
updatedIndex = b.create<arith::IndexCastOp>(
loc, indicesElemType, updatedIndex);
b.create<linalg::YieldOp>(loc, updatedIndex);
})
.getResult(0);
// Creating a new tensor initialized with zeros and size same as the input
// tensor.
Value outputTensor =
createZeroInitTensor(rewriter, loc, inputShape, inputElemType);
// Collapsing `gradOutput` into a 1-d tensor.
SmallVector<ReassociationIndices> reassociationCollapse(1);
for (auto i = 0; i < gradOutputType.getRank(); i++)
reassociationCollapse[0].push_back(i);
RankedTensorType gradOutputFlattenedType;
int64_t numelGradOutput = getNumberOfElements(gradOutputType);
gradOutputFlattenedType = RankedTensorType::get(
makeShapeLLVMCompatible({numelGradOutput}), gradOutputElemType);
Value gradOutputFlattened = rewriter.create<tensor::CollapseShapeOp>(
loc, gradOutputFlattenedType, gradOutput, reassociationCollapse);
// Collapsing updated indices into a 2-d tensor.
SmallVector<ReassociationIndices> reassociationCollapseIndices(2);
for (auto i = 0; i < tensorOperandRank; i++)
reassociationCollapseIndices[0].push_back(i);
reassociationCollapseIndices[1].push_back(tensorOperandRank);
int64_t numelIndices = getNumberOfElements(indicesType);
Value indicesCollapsed = rewriter.create<tensor::CollapseShapeOp>(
loc,
RankedTensorType::get(
makeShapeLLVMCompatible({numelIndices, tensorOperandRank}),
indicesElemType),
updatedIndices, reassociationCollapseIndices);
bool invalidInputTypeFound = false;
Value scatterOp = createTMTensorScatterOp(
rewriter, loc, /*updates=*/gradOutputFlattened,
/*indices=*/indicesCollapsed, /*original=*/outputTensor,
/*dimensionsMap=*/createDefaultDimMap(indicesCollapsed),
/*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value valuesElement,
Value inputElement) {
Value yieldValue = valuesElement;
if (inputElement.getType().isa<mlir::IntegerType>()) {
yieldValue =
b.create<arith::AddIOp>(loc, inputElement, valuesElement);
} else if (inputElement.getType().isa<mlir::FloatType>()) {
yieldValue =
b.create<arith::AddFOp>(loc, inputElement, valuesElement);
} else {
invalidInputTypeFound = true;
return;
}
b.create<TMTensor::YieldOp>(loc, yieldValue);
});
if (invalidInputTypeFound) {
return rewriter.notifyMatchFailure(
op,
"unimplemented: input tensor must be of integer type or float type");
}
Type newResultType = getTypeConverter()->convertType(op.getType());
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, newResultType, scatterOp);
return success();
}
};
} // namespace
namespace {
class ConvertAtenScatterReduceTwoOp
: public OpConversionPattern<AtenScatterReduceTwoOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenScatterReduceTwoOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
if (failed(verifyLinalgCompatibleTypes(op, rewriter)))
return failure();
Location loc = op.getLoc();
RankedTensorType selfType =
adaptor.getSelf().getType().cast<RankedTensorType>();
RankedTensorType indexType =
adaptor.getIndex().getType().cast<RankedTensorType>();
RankedTensorType srcType =
adaptor.getSrc().getType().cast<RankedTensorType>();
Value self = adaptor.getSelf();
if (selfType.getRank() != indexType.getRank() ||
indexType.getRank() != srcType.getRank())
return rewriter.notifyMatchFailure(op,
"'self', 'index' and 'src' should all "
"have the same number of dimensions.");
std::string reduceType;
if (!matchPattern(op.getReduce(), m_TorchConstantStr(reduceType)))
return rewriter.notifyMatchFailure(op,
"'reduce' must be a costant string");
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(op, "'dim' is not constant");
bool includeSelf;
if (!matchPattern(op.getIncludeSelf(), m_TorchConstantBool(&includeSelf)))
return rewriter.notifyMatchFailure(op, "'include_self' is not constant");
// Get reduce string as the equivalent enum
auto reduceEnum = torch_upstream::get_reduction_enum(reduceType);
// Get the inputs reformatted for the TMScatterOp
auto [indices, updates] =
convertTorchScatterIndexAndSrcToTMScatterIndexAndSrc(
rewriter, adaptor.getIndex(), adaptor.getSrc(), dim);
// Value 'counts' will be used to tally the number of reductions into
// each unique index. The tally is used to calculate the average of the
// values scattered per index.
Value counts = nullptr;
if (reduceEnum == torch_upstream::ReductionType::MEAN) {
SmallVector<Value> selfShape =
getTensorSizes(rewriter, loc, adaptor.getSelf());
TypedAttr initAttr;
if (llvm::isa<mlir::FloatType>(srcType.getElementType())) {
initAttr = rewriter.getFloatAttr(srcType.getElementType(), 1);
} else if (llvm::isa<mlir::IntegerType>(srcType.getElementType())) {
initAttr = rewriter.getIntegerAttr(srcType.getElementType(), 1);
} else {
llvm_unreachable("Only integer/float types supported!");
}
Value initElement = rewriter.create<arith::ConstantOp>(loc, initAttr);
counts = createInitTensor(rewriter, loc, selfShape,
selfType.getElementType(), initElement);
}
// If the original values shouldn't be included, normalize the
// input tensor where the scatters take place.
if (!includeSelf) {
Value normalizationValue;
if (reduceEnum == torch_upstream::ReductionType::SUM ||
reduceEnum == torch_upstream::ReductionType::MEAN) {
// Set the values in the input tensor to '0' so they are not included
normalizationValue = rewriter.create<arith::ConstantOp>(
loc, rewriter.getZeroAttr(srcType.getElementType()));
} else if (reduceEnum == torch_upstream::ReductionType::PROD) {
// Set the values in the input tensor to '1' (multiplication identity)
if (llvm::isa<mlir::FloatType>(srcType.getElementType())) {
normalizationValue = rewriter.create<arith::ConstantOp>(
loc, rewriter.getFloatAttr(srcType.getElementType(), 1.0));
} else if (llvm::isa<mlir::IntegerType>(srcType.getElementType())) {
normalizationValue = rewriter.create<arith::ConstantOp>(
loc, rewriter.getIntegerAttr(srcType.getElementType(), 1));
} else {
llvm_unreachable("Only integer/float types supported!");
}
} else if (reduceEnum == torch_upstream::ReductionType::MAX) {
// Set the values in the input tensor to the smallest element of that
// type
TypedAttr minAttr = getNumericLimit(rewriter, srcType.getElementType(),
/*getMin=*/true);
normalizationValue = rewriter.create<arith::ConstantOp>(loc, minAttr);
} else if (reduceEnum == torch_upstream::ReductionType::MIN) {
// Set the values in the input tensor to the largest element of that
// type
TypedAttr maxAttr = getNumericLimit(rewriter, srcType.getElementType(),
/*getMin=*/false);
normalizationValue = rewriter.create<arith::ConstantOp>(loc, maxAttr);
}
// Scatter the normalizations into the input tensor
Value indexSize = getTensorSize(rewriter, loc, adaptor.getIndex());
indexSize = castIntToIndex(rewriter, loc, indexSize);
Value normalizations = createInitTensor(
rewriter, loc, SmallVector<Value>({indexSize}),
srcType.getElementType(), /*init_element=*/normalizationValue);
self = createTMTensorScatterOp(
rewriter, loc, normalizations, indices, self,
/*dimensionsMap=*/createDefaultDimMap(indices),
/*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value update, Value current) {
b.create<TMTensor::YieldOp>(loc, update);
});
if (reduceEnum == torch_upstream::ReductionType::MEAN) {
counts = createTMTensorScatterOp(
rewriter, loc, normalizations, indices, counts,
/*dimensionsMap=*/createDefaultDimMap(indices),
/*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value update, Value current) {
b.create<TMTensor::YieldOp>(loc, update);
});
}
}
// Create final operation
Value scatterOp = createTMTensorScatterOp(
rewriter, loc, updates, indices, self,
/*dimensionsMap=*/createDefaultDimMap(indices), /*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value update, Value current) {
Value result;
if (reduceEnum == torch_upstream::ReductionType::SUM ||
reduceEnum == torch_upstream::ReductionType::MEAN) {
if (update.getType().isa<mlir::IntegerType>()) {
result = b.create<arith::AddIOp>(loc, update, current);
} else if (update.getType().isa<mlir::FloatType>()) {
result = b.create<arith::AddFOp>(loc, update, current);
} else {
llvm_unreachable("Only integer/float types supported!");
}
} else if (reduceEnum == torch_upstream::ReductionType::PROD) {
if (update.getType().isa<mlir::IntegerType>()) {
result = b.create<arith::MulIOp>(loc, update, current);
} else if (update.getType().isa<mlir::FloatType>()) {
result = b.create<arith::MulFOp>(loc, update, current);
} else {
llvm_unreachable("Only integer/float types supported!");
}
} else if (reduceEnum == torch_upstream::ReductionType::MAX) {
if (update.getType().isa<mlir::IntegerType>()) {
result = b.create<arith::MaxSIOp>(loc, update, current);
} else if (update.getType().isa<mlir::FloatType>()) {
result = b.create<arith::MaximumFOp>(loc, update, current);
} else {
llvm_unreachable("Only integer/float types supported!");
}
} else if (reduceEnum == torch_upstream::ReductionType::MIN) {
if (update.getType().isa<mlir::IntegerType>()) {
result = b.create<arith::MinSIOp>(loc, update, current);
} else if (update.getType().isa<mlir::FloatType>()) {
result = b.create<arith::MinimumFOp>(loc, update, current);
} else {
llvm_unreachable("Only integer/float types supported!");
}
}
b.create<TMTensor::YieldOp>(loc, result);
});
// Special case for the mean
if (reduceEnum == torch_upstream::ReductionType::MEAN) {
counts = createTMTensorScatterOp(
rewriter, loc, updates, indices, counts,
/*dimensionsMap=*/createDefaultDimMap(indices),
/*uniqueIndices=*/false,
[&](OpBuilder &b, Location loc, Value update, Value current) {
Value result;
if (mlir::IntegerType intType =
llvm::dyn_cast<mlir::IntegerType>(current.getType())) {
Value constantUpdate = b.create<arith::ConstantOp>(
loc, b.getIntegerAttr(intType, 1));
result = b.create<arith::AddIOp>(loc, constantUpdate, current);
} else if (mlir::FloatType floatType =
llvm::dyn_cast<mlir::FloatType>(current.getType())) {
Value constantUpdate = b.create<arith::ConstantOp>(
loc, b.getFloatAttr(floatType, 1.0));
result = b.create<arith::AddFOp>(loc, constantUpdate, current);
} else {
llvm_unreachable("Only integer/float types supported!");
}
b.create<TMTensor::YieldOp>(loc, result);
});
Value output = rewriter.create<tensor::EmptyOp>(
loc, tensor::getMixedSizes(rewriter, loc, self),
selfType.getElementType());
// Finally divide the result
scatterOp =
rewriter
.create<linalg::MapOp>(
loc, ValueRange{scatterOp, counts}, output,
[&](OpBuilder &b, Location loc, ValueRange args) {
Value result;
if (llvm::isa<mlir::IntegerType>(args[0].getType())) {
result = b.create<arith::DivSIOp>(loc, args[0], args[1]);
} else if (llvm::isa<mlir::FloatType>(args[0].getType())) {
result = b.create<arith::DivFOp>(loc, args[0], args[1]);
} else {
llvm_unreachable("Only integer/float types supported!");
}
b.create<linalg::YieldOp>(loc, result);
})
.getResult()[0];
}
auto resultType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, scatterOp);
return success();
}
};
} // namespace
namespace {
class ConvertAtenSortOp : public OpConversionPattern<AtenSortOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenSortOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op.getLoc();
// Step 1. Fetch Input to sort.
Value inputTensor = adaptor.getSelf();
auto inputType = inputTensor.getType().cast<RankedTensorType>();
unsigned inputRank = inputType.getRank();
// Step 2. Fetch dimension to perform sort in.
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
return rewriter.notifyMatchFailure(
op, "unimplemented: only constant dim value is supported");
dim = toPositiveDim(dim, inputRank);
if (!isValidDim(dim, inputRank)) {
return rewriter.notifyMatchFailure(op, "dim is statically invalid");
}
// Step 3. Fetch the order of sorting.
bool descending;
if (!matchPattern(op.getDescending(), m_TorchConstantBool(&descending)))
return rewriter.notifyMatchFailure(
op, "unimplemented: only constant descending value is supported");
// Step 4. Form a RankedTensorType with same shape as that of the input's
// but with elemental type i64.
RankedTensorType indicesType =
RankedTensorType::get(inputType.getShape(), rewriter.getI64Type());
// Step 5. Generate indices tensor.
SmallVector<Value> dynDims;
for (unsigned i = 0; i < inputType.getRank(); i++) {
if (inputType.isDynamicDim(i)) {
dynDims.push_back(rewriter.create<tensor::DimOp>(loc, inputTensor, i));
}
}
Value initEmptyTensor = rewriter.create<tensor::EmptyOp>(
loc, inputType.getShape(), rewriter.getI64Type(), dynDims);
SmallVector<AffineMap> indexingMaps = {
AffineMap::getMultiDimIdentityMap(inputRank, op.getContext())};
SmallVector<utils::IteratorType> iteratorTypes(
inputRank, utils::IteratorType::parallel);
Value indicesTensor =
rewriter
.create<linalg::GenericOp>(
loc, initEmptyTensor.getType(), ValueRange{}, initEmptyTensor,
indexingMaps, iteratorTypes,
[&](OpBuilder &b, Location loc, ValueRange args) {
Value index = b.create<linalg::IndexOp>(loc, dim);
index = castIndexToInt64(b, loc, index);
b.create<linalg::YieldOp>(loc, index);
})
.getResult(0);
// Step 6. Create TMTensor::SortOp.
SmallVector<Value> operands;
operands.push_back(inputTensor);
operands.push_back(indicesTensor);
SmallVector<Type> elementTypes;
elementTypes.push_back(inputType.getElementType());
elementTypes.push_back(indicesType.getElementType());
// The default value for aten.sort op's `stable` parameter is `false`.
// Refer: https://pytorch.org/docs/stable/generated/torch.sort.html
FailureOr<SmallVector<Value>> sortOpValues =
createTMTensorSortOp(rewriter, loc, operands, elementTypes,
/*dimension=*/dim, /*isStable=*/false,
/*isDescending=*/descending);
if (failed(sortOpValues))
return rewriter.notifyMatchFailure(
loc, "Only Integer and Floating element type expected.");
auto sortOpVal = *sortOpValues;
rewriter.replaceOp(op, sortOpVal);
return success();
}
};
} // namespace
2022-09-09 12:36:40 +08:00
namespace {
class ConvertAtenCumsumOp : public OpConversionPattern<AtenCumsumOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenCumsumOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Location loc = op.getLoc();
Value input = adaptor.getSelf();
auto resultType = getTypeConverter()
->convertType(op->getResult(0).getType())
.cast<RankedTensorType>();
2022-09-09 12:36:40 +08:00
Type elementType = resultType.getElementType();
Type inputElementType =
input.getType().cast<RankedTensorType>().getElementType();
// Converting the input element type to the result's element type.
// The only possible mismatch would be when the input element type is an
// integer but not `si64`. Therefore, we directly convert the input to
// `si64`. Rest all cases are handled in the dtype definition for this op.
if (elementType != inputElementType) {
Value torchInput = convertTensorToDtype(
rewriter, loc, op.getSelf(),
rewriter.getIntegerType(64, IntegerType::Signed));
input = typeConverter->materializeTargetConversion(
rewriter, loc, typeConverter->convertType(torchInput.getType()),
torchInput);
}
2022-09-09 12:36:40 +08:00
int64_t inputRank = resultType.getRank();
Value dtype = op.getDtype();
2022-09-09 12:36:40 +08:00
if (!dtype.getType().isa<Torch::NoneType>())
return rewriter.notifyMatchFailure(
op, "unsupported: dtype argument not supported");
int64_t dim;
if (!matchPattern(op.getDim(), m_TorchConstantInt(&dim)))
2022-09-09 12:36:40 +08:00
return rewriter.notifyMatchFailure(
op, "unimplemented: only constant dim value is supported");
dim = toPositiveDim(dim, inputRank);
if (!isValidDim(dim, inputRank))
return rewriter.notifyMatchFailure(op, "invalid dim");
SmallVector<Value> sizes = getTensorSizes(rewriter, loc, input);
Value output = createZeroInitTensor(rewriter, loc, sizes, elementType);
output = rewriter.create<tensor::CastOp>(loc, resultType, output);
SmallVector<Value> accSizes(sizes);
accSizes.erase(accSizes.begin() + dim);
SmallVector<int64_t> accStatic(
makeShapeTorchCompatible(resultType.getShape()));
2022-09-09 12:36:40 +08:00
accStatic.erase(accStatic.begin() + dim);
Value acc = createZeroInitTensor(rewriter, loc, accSizes, elementType);
Type accType =
RankedTensorType::get(makeShapeLLVMCompatible(accStatic), elementType);
2022-09-09 12:36:40 +08:00
acc = rewriter.create<tensor::CastOp>(loc, accType, acc);
Value result = createTMTensorScanOp(
rewriter, loc, input, output, acc, dim, /*inclusive=*/true,
[](OpBuilder &b, Location loc, Value input, Value acc) {
Value sum =
(input.getType().isa<mlir::FloatType>()
? b.create<arith::AddFOp>(loc, input, acc)->getResult(0)
: b.create<arith::AddIOp>(loc, input, acc)->getResult(0));
2022-09-09 12:36:40 +08:00
b.create<TMTensor::YieldOp>(loc, sum);
});
rewriter.replaceOpWithNewOp<tensor::CastOp>(op, resultType, result);
return success();
}
};
} // namespace
namespace {
class ConvertAtenScaledDotProductAttentionOp
: public OpConversionPattern<AtenScaledDotProductAttentionOp> {
public:
using OpConversionPattern::OpConversionPattern;
LogicalResult
matchAndRewrite(AtenScaledDotProductAttentionOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
Value mask = op.getAttnMask();
Value dropoutP = op.getDropoutP();
Value isCausal = op.getIsCausal();
Value scale = op.getScale();
Type elementType =
adaptor.getQuery().getType().cast<ShapedType>().getElementType();
// Verify inputs (only support defaults)
if (!mask.getType().isa<Torch::NoneType>())
return rewriter.notifyMatchFailure(op.getLoc(),
"attention masking not supported");
double dropout;
if (!matchPattern(dropoutP, m_TorchConstantFloat(&dropout)) ||
dropout > 0.0)
return rewriter.notifyMatchFailure(op.getLoc(), "dropout not supported");
bool causal;
if (!matchPattern(isCausal, m_TorchConstantBool(&causal)) || causal)
return rewriter.notifyMatchFailure(
op.getLoc(), "causal attention masking not supported");
if (!scale.getType().isa<Torch::NoneType>()) {
double scaleFloat;
if (!matchPattern(scale, m_TorchConstantFloat(&scaleFloat)) ||
scaleFloat != 1.0)
return rewriter.notifyMatchFailure(op.getLoc(),
"only default scale supported");
}
auto opTy = cast<ValueTensorType>(op.getType()).toBuiltinTensor();
auto query = adaptor.getQuery();
auto value = adaptor.getValue();
auto key = adaptor.getKey();
auto queryTy = cast<ShapedType>(query.getType());
auto valueTy = cast<ShapedType>(value.getType());
auto keyTy = cast<ShapedType>(key.getType());
if (queryTy.getRank() != valueTy.getRank() ||
queryTy.getRank() != keyTy.getRank())
return rewriter.notifyMatchFailure(op, "operand ranks do not match");
if (queryTy.getRank() < 3)
return rewriter.notifyMatchFailure(op, "missing batch dimension");
llvm::SmallVector<ReassociationIndices, 3> reassociation(3);
for (int i = 0, s = valueTy.getRank() - 2; i < s; ++i)
reassociation.front().push_back(i);
reassociation[1].push_back(valueTy.getRank() - 2);
reassociation[2].push_back(valueTy.getRank() - 1);
auto loc = op.getLoc();
auto collapseBatch = [&rewriter, &reassociation,
loc](Value value) -> Value {
auto valueTy = cast<ShapedType>(value.getType());
if (valueTy.getRank() == 3)
return value;
llvm::SmallVector<int64_t, 3> newShape(3, 1);
newShape[1] = valueTy.getDimSize(valueTy.getRank() - 2);
newShape[2] = valueTy.getDimSize(valueTy.getRank() - 1);
for (int i = 0, s = valueTy.getRank() - 2; i < s; ++i) {
if (valueTy.isDynamicDim(i)) {
newShape[0] = ShapedType::kDynamic;
break;
}
newShape[0] = newShape[0] * valueTy.getDimSize(i);
}
auto collapseTy = valueTy.clone(newShape);
return rewriter.create<tensor::CollapseShapeOp>(loc, collapseTy, value,
reassociation);
};
query = collapseBatch(query);
key = collapseBatch(key);
value = collapseBatch(value);
SmallVector<int64_t> outSizes(
query.getType().cast<ShapedType>().getShape());
SmallVector<int64_t> valueSizes(
value.getType().cast<ShapedType>().getShape());
outSizes[outSizes.size() - 1] = valueSizes[valueSizes.size() - 1];
SmallVector<Value> outSizesDynamic(
getTensorSizes(rewriter, op.getLoc(), query));
outSizesDynamic[outSizesDynamic.size() - 1] =
getTensorSizes(rewriter, op.getLoc(), value)[valueSizes.size() - 1];
Type outType = RankedTensorType::get(outSizes, elementType);
Value output = createZeroInitTensor(rewriter, op.getLoc(), outSizesDynamic,
elementType);
// Overwrite with tm_tensor::attention
Value attention =
rewriter
.create<AttentionOp>(loc, outType,
SmallVector<Value>{query, key, value},
SmallVector<Value>{output})
.getResult()[0];
if (opTy != outType) {
attention = rewriter.create<tensor::ExpandShapeOp>(loc, opTy, attention,
reassociation);
}
rewriter.replaceOp(op, attention);
return success();
}
};
} // namespace
// -----------------------------------------------------------------------------
// The pass
// -----------------------------------------------------------------------------
namespace {
class ConvertTorchToTMTensor
: public ConvertTorchToTMTensorBase<ConvertTorchToTMTensor> {
public:
void getDependentDialects(DialectRegistry &registry) const override {
registry.insert<linalg::LinalgDialect>();
registry.insert<func::FuncDialect>();
registry.insert<tensor::TensorDialect>();
registry.insert<arith::ArithDialect>();
registry.insert<TMTensorDialect>();
TorchConversion::getBackendTypeConversionDependentDialects(registry);
}
void runOnOperation() override {
MLIRContext *context = &getContext();
ConversionTarget target(*context);
target.addLegalDialect<linalg::LinalgDialect, func::FuncDialect,
tensor::TensorDialect, arith::ArithDialect,
math::MathDialect, Torch::TorchDialect,
TMTensorDialect>();
TypeConverter typeConverter;
typeConverter.addConversion([](Type type) { return type; });
TorchConversion::setupBackendTypeConversion(target, typeConverter);
RewritePatternSet patterns(context);
target.addIllegalOp<AtenBincountOp>();
patterns.add<ConvertAtenBincountOp>(typeConverter, context);
target.addIllegalOp<Aten_IndexPutImplOp>();
patterns.add<ConvertAten_IndexPutImplOp>(typeConverter, context);
target.addIllegalOp<AtenMaxPool2dWithIndicesBackwardOp>();
patterns.add<ConvertAtenMaxPool2dWithIndicesBackwardOp>(typeConverter,
context);
target.addIllegalOp<AtenScatterReduceTwoOp>();
patterns.add<ConvertAtenScatterReduceTwoOp>(typeConverter, context);
target.addIllegalOp<AtenSortOp>();
patterns.add<ConvertAtenSortOp>(typeConverter, context);
2022-09-09 12:36:40 +08:00
target.addIllegalOp<AtenCumsumOp>();
patterns.add<ConvertAtenCumsumOp>(typeConverter, context);
target.addIllegalOp<AtenScaledDotProductAttentionOp>();
patterns.add<ConvertAtenScaledDotProductAttentionOp>(typeConverter,
context);
target.addIllegalOp<AtenScatterSrcOp>();
patterns.add<ConvertAtenScatterSrcOp>(typeConverter, context);
if (failed(applyPartialConversion(getOperation(), target,
std::move(patterns))))
return signalPassFailure();
}
};
} // namespace
std::unique_ptr<OperationPass<func::FuncOp>>
mlir::torch::createConvertTorchToTMTensorPass() {
return std::make_unique<ConvertTorchToTMTensor>();
}