#include "mlir/Dialect/Tensor/Transforms/BufferizableOpInterfaceImpl.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Bufferization/IR/BufferizableOpInterface.h"
#include "mlir/Dialect/Bufferization/IR/Bufferization.h"
#include "mlir/Dialect/Bufferization/IR/DstBufferizableOpInterfaceImpl.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Tensor/Transforms/SubsetInsertionOpInterfaceImpl.h"
#include "mlir/Dialect/Utils/StaticValueUtils.h"
#include "mlir/IR/Dialect.h"
#include "mlir/IR/Operation.h"
using namespace mlir;
using namespace mlir::bufferization;
using namespace mlir::tensor;
namespace mlir {
namespace tensor {
namespace {
struct CastOpInterface
: public BufferizableOpInterface::ExternalModel<CastOpInterface,
tensor::CastOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {{op->getResult(0), BufferRelation::Equivalent}};
}
FailureOr<BaseMemRefType>
getBufferType(Operation *op, Value value, const BufferizationOptions &options,
SmallVector<Value> &invocationStack) const {
auto castOp = cast<tensor::CastOp>(op);
auto maybeSrcBufferType = bufferization::getBufferType(
castOp.getSource(), options, invocationStack);
if (failed(maybeSrcBufferType))
return failure();
Attribute memorySpace = maybeSrcBufferType->getMemorySpace();
if (isa<UnrankedTensorType>(castOp.getSource().getType())) {
return getMemRefTypeWithFullyDynamicLayout(castOp.getType(), memorySpace);
}
if (isa<UnrankedTensorType>(castOp.getType())) {
return getMemRefTypeWithFullyDynamicLayout(castOp.getType(), memorySpace);
}
auto rankedResultType = cast<RankedTensorType>(castOp.getType());
return MemRefType::get(
rankedResultType.getShape(), rankedResultType.getElementType(),
llvm::cast<MemRefType>(*maybeSrcBufferType).getLayout(), memorySpace);
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto castOp = cast<tensor::CastOp>(op);
FailureOr<Value> resultBuffer =
getBuffer(rewriter, castOp.getSource(), options);
if (failed(resultBuffer))
return failure();
auto resultMemRefType =
bufferization::getBufferType(castOp.getResult(), options);
if (failed(resultMemRefType))
return failure();
if (resultBuffer->getType() == *resultMemRefType) {
replaceOpWithBufferizedValues(rewriter, op, *resultBuffer);
return success();
}
assert(memref::CastOp::areCastCompatible(resultBuffer->getType(),
*resultMemRefType) &&
"CallOp::bufferize: cast incompatible");
replaceOpWithNewBufferizedOp<memref::CastOp>(
rewriter, op, *resultMemRefType, *resultBuffer);
return success();
}
};
struct CollapseShapeOpInterface
: public BufferizableOpInterface::ExternalModel<CollapseShapeOpInterface,
tensor::CollapseShapeOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {{op->getOpResult(0), BufferRelation::Equivalent}};
}
FailureOr<BaseMemRefType>
getBufferType(Operation *op, Value value, const BufferizationOptions &options,
SmallVector<Value> &invocationStack) const {
auto collapseShapeOp = cast<tensor::CollapseShapeOp>(op);
auto maybeSrcBufferType = bufferization::getBufferType(
collapseShapeOp.getSrc(), options, invocationStack);
if (failed(maybeSrcBufferType))
return failure();
auto srcBufferType = llvm::cast<MemRefType>(*maybeSrcBufferType);
bool canBeCollapsed = memref::CollapseShapeOp::isGuaranteedCollapsible(
srcBufferType, collapseShapeOp.getReassociationIndices());
if (!canBeCollapsed) {
RankedTensorType tensorResultType = collapseShapeOp.getResultType();
return bufferization::getMemRefTypeWithStaticIdentityLayout(
tensorResultType, srcBufferType.getMemorySpace());
}
return memref::CollapseShapeOp::computeCollapsedType(
srcBufferType, collapseShapeOp.getReassociationIndices());
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto collapseShapeOp = cast<tensor::CollapseShapeOp>(op);
RankedTensorType tensorResultType = collapseShapeOp.getResultType();
FailureOr<Value> maybeBuffer =
getBuffer(rewriter, collapseShapeOp.getSrc(), options);
if (failed(maybeBuffer))
return failure();
Value buffer = *maybeBuffer;
auto bufferType = cast<MemRefType>(buffer.getType());
if (tensorResultType.getRank() == 0) {
MemRefType resultType;
if (bufferType.getLayout().isIdentity()) {
MemRefLayoutAttrInterface layout;
resultType = MemRefType::get({}, tensorResultType.getElementType(),
layout, bufferType.getMemorySpace());
} else {
SmallVector<int64_t> strides;
int64_t offset;
if (failed(getStridesAndOffset(bufferType, strides, offset)))
return failure();
resultType = MemRefType::get(
{}, tensorResultType.getElementType(),
StridedLayoutAttr::get(op->getContext(), offset, {}),
bufferType.getMemorySpace());
}
replaceOpWithNewBufferizedOp<memref::CollapseShapeOp>(
rewriter, op, resultType, buffer, collapseShapeOp.getReassociation());
return success();
}
bool canBeCollapsed = memref::CollapseShapeOp::isGuaranteedCollapsible(
bufferType, collapseShapeOp.getReassociationIndices());
if (!canBeCollapsed) {
AnalysisState analysisState(options);
FailureOr<Value> tensorAlloc = allocateTensorForShapedValue(
rewriter, op->getLoc(), collapseShapeOp.getSrc(), options);
if (failed(tensorAlloc))
return failure();
auto memrefType =
MemRefType::get(collapseShapeOp.getSrcType().getShape(),
collapseShapeOp.getSrcType().getElementType(),
AffineMap(), bufferType.getMemorySpace());
buffer = rewriter.create<bufferization::ToMemrefOp>(
op->getLoc(), memrefType, *tensorAlloc);
}
replaceOpWithNewBufferizedOp<memref::CollapseShapeOp>(
rewriter, op, buffer, collapseShapeOp.getReassociationIndices());
return success();
}
};
struct DimOpInterface
: public BufferizableOpInterface::ExternalModel<DimOpInterface,
tensor::DimOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto dimOp = cast<tensor::DimOp>(op);
FailureOr<Value> v = getBuffer(rewriter, dimOp.getSource(), options);
if (failed(v))
return failure();
replaceOpWithNewBufferizedOp<memref::DimOp>(rewriter, op, *v,
dimOp.getIndex());
return success();
}
};
struct EmptyOpInterface
: public BufferizableOpInterface::ExternalModel<EmptyOpInterface,
tensor::EmptyOp> {
bool bufferizesToAllocation(Operation *op, Value value) const { return true; }
bool resultBufferizesToMemoryWrite(Operation *op, OpResult opResult,
const AnalysisState &state) const {
return false;
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto emptyOp = cast<tensor::EmptyOp>(op);
if (op->getUses().empty()) {
rewriter.eraseOp(op);
return success();
}
FailureOr<Value> allocTensor = allocateTensorForShapedValue(
rewriter, op->getLoc(), emptyOp.getResult(), options, false);
if (failed(allocTensor))
return failure();
rewriter.replaceOp(op, *allocTensor);
return success();
}
};
struct ExpandShapeOpInterface
: public BufferizableOpInterface::ExternalModel<ExpandShapeOpInterface,
tensor::ExpandShapeOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {{op->getOpResult(0), BufferRelation::Equivalent}};
}
FailureOr<BaseMemRefType>
getBufferType(Operation *op, Value value, const BufferizationOptions &options,
SmallVector<Value> &invocationStack) const {
auto expandShapeOp = cast<tensor::ExpandShapeOp>(op);
auto maybeSrcBufferType = bufferization::getBufferType(
expandShapeOp.getSrc(), options, invocationStack);
if (failed(maybeSrcBufferType))
return failure();
auto srcBufferType = llvm::cast<MemRefType>(*maybeSrcBufferType);
auto maybeResultType = memref::ExpandShapeOp::computeExpandedType(
srcBufferType, expandShapeOp.getResultType().getShape(),
expandShapeOp.getReassociationIndices());
if (failed(maybeResultType))
return failure();
return *maybeResultType;
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto expandShapeOp = cast<tensor::ExpandShapeOp>(op);
auto tensorResultType = expandShapeOp.getResultType();
FailureOr<Value> buffer =
getBuffer(rewriter, expandShapeOp.getSrc(), options);
if (failed(buffer))
return failure();
replaceOpWithNewBufferizedOp<memref::ExpandShapeOp>(
rewriter, op, tensorResultType.getShape(), *buffer,
expandShapeOp.getReassociationIndices());
return success();
}
};
struct ExtractSliceOpInterface
: public BufferizableOpInterface::ExternalModel<ExtractSliceOpInterface,
tensor::ExtractSliceOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {{op->getOpResult(0), BufferRelation::Unknown}};
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto extractSliceOp = cast<tensor::ExtractSliceOp>(op);
SmallVector<OpFoldResult> mixedOffsets = extractSliceOp.getMixedOffsets();
SmallVector<OpFoldResult> mixedSizes = extractSliceOp.getMixedSizes();
SmallVector<OpFoldResult> mixedStrides = extractSliceOp.getMixedStrides();
Location loc = extractSliceOp.getLoc();
FailureOr<Value> srcMemref =
getBuffer(rewriter, extractSliceOp.getSource(), options);
if (failed(srcMemref))
return failure();
auto resultMemrefType =
bufferization::getBufferType(extractSliceOp.getResult(), options);
if (failed(resultMemrefType))
return failure();
Value subView = rewriter.create<memref::SubViewOp>(
loc, llvm::cast<MemRefType>(*resultMemrefType), *srcMemref,
mixedOffsets, mixedSizes, mixedStrides);
replaceOpWithBufferizedValues(rewriter, op, subView);
return success();
}
FailureOr<BaseMemRefType>
getBufferType(Operation *op, Value value, const BufferizationOptions &options,
SmallVector<Value> &invocationStack) const {
auto extractSliceOp = cast<tensor::ExtractSliceOp>(op);
assert(value == extractSliceOp.getResult() && "invalid value");
auto srcMemrefType = bufferization::getBufferType(
extractSliceOp.getSource(), options, invocationStack);
if (failed(srcMemrefType))
return failure();
SmallVector<OpFoldResult> mixedOffsets = extractSliceOp.getMixedOffsets();
SmallVector<OpFoldResult> mixedSizes = extractSliceOp.getMixedSizes();
SmallVector<OpFoldResult> mixedStrides = extractSliceOp.getMixedStrides();
return cast<BaseMemRefType>(memref::SubViewOp::inferRankReducedResultType(
extractSliceOp.getType().getShape(),
llvm::cast<MemRefType>(*srcMemrefType), mixedOffsets, mixedSizes,
mixedStrides));
}
};
struct ExtractOpInterface
: public BufferizableOpInterface::ExternalModel<ExtractOpInterface,
tensor::ExtractOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto extractOp = cast<tensor::ExtractOp>(op);
FailureOr<Value> srcMemref =
getBuffer(rewriter, extractOp.getTensor(), options);
if (failed(srcMemref))
return failure();
replaceOpWithNewBufferizedOp<memref::LoadOp>(rewriter, op, *srcMemref,
extractOp.getIndices());
return success();
}
};
static void createStores(RewriterBase &rewriter, Location loc, int dim,
Value buffer, ArrayRef<int64_t> shape,
ArrayRef<Value> constants,
OperandRange::iterator &elementIt,
SmallVectorImpl<Value> &indices) {
if (dim == static_cast<int>(shape.size()) - 1) {
for (int i = 0; i < shape.back(); ++i) {
indices.back() = constants[i];
rewriter.create<memref::StoreOp>(loc, *elementIt, buffer, indices);
++elementIt;
}
return;
}
for (int i = 0; i < shape[dim]; ++i) {
indices[dim] = constants[i];
createStores(rewriter, loc, dim + 1, buffer, shape, constants, elementIt,
indices);
}
}
struct FromElementsOpInterface
: public BufferizableOpInterface::ExternalModel<FromElementsOpInterface,
tensor::FromElementsOp> {
bool bufferizesToAllocation(Operation *op, Value value) const { return true; }
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto fromElementsOp = cast<tensor::FromElementsOp>(op);
auto tensorType = cast<RankedTensorType>(fromElementsOp.getType());
if (options.defaultMemorySpaceFn(tensorType) != Attribute())
return op->emitError("memory space not implemented yet");
Location loc = op->getLoc();
auto shape = tensorType.getShape();
FailureOr<Value> tensorAlloc = allocateTensorForShapedValue(
rewriter, loc, fromElementsOp.getResult(), options,
false);
if (failed(tensorAlloc))
return failure();
auto memrefType =
MemRefType::get(tensorType.getShape(), tensorType.getElementType());
Value buffer = rewriter.create<bufferization::ToMemrefOp>(
op->getLoc(), memrefType, *tensorAlloc);
if (fromElementsOp.getElements().empty()) {
replaceOpWithBufferizedValues(rewriter, op, buffer);
return success();
}
if (shape.empty()) {
rewriter.create<memref::StoreOp>(
loc, fromElementsOp.getElements().front(), buffer);
replaceOpWithBufferizedValues(rewriter, op, buffer);
return success();
}
auto maxDim = *llvm::max_element(shape);
SmallVector<Value, 2> constants;
constants.reserve(maxDim);
for (int i = 0; i < maxDim; ++i)
constants.push_back(rewriter.create<arith::ConstantIndexOp>(loc, i));
auto elementIt = fromElementsOp.getElements().begin();
SmallVector<Value, 2> indices(tensorType.getRank(), constants[0]);
createStores(rewriter, loc, 0, buffer, shape, constants, elementIt,
indices);
replaceOpWithBufferizedValues(rewriter, op, buffer);
return success();
}
};
static Value lowerGenerateLikeOpBody(RewriterBase &rewriter, Location loc,
Value tensorDestination,
ValueRange dynamicSizes,
Region &generateBody) {
assert(generateBody.hasOneBlock() && "expected body with single block");
auto tensorType = cast<RankedTensorType>(tensorDestination.getType());
assert(generateBody.getNumArguments() == tensorType.getRank() &&
"rank mismatch");
OpBuilder::InsertionGuard g(rewriter);
auto linalgOp =
rewriter.create<linalg::MapOp>(loc, tensorType, ValueRange(),
tensorDestination);
Block &linalgBody = linalgOp.getMapper().emplaceBlock();
rewriter.setInsertionPointToStart(&linalgBody);
SmallVector<Value> indices;
for (int64_t dim = 0; dim < tensorType.getRank(); ++dim)
indices.push_back(rewriter.create<linalg::IndexOp>(loc, dim));
rewriter.mergeBlocks(&generateBody.front(), &linalgBody, indices);
auto yieldOp = cast<tensor::YieldOp>(linalgBody.getTerminator());
rewriter.replaceOpWithNewOp<linalg::YieldOp>(yieldOp, yieldOp.getValue());
return linalgOp.getResult()[0];
}
struct GenerateOpInterface
: public BufferizableOpInterface::ExternalModel<GenerateOpInterface,
tensor::GenerateOp> {
bool bufferizesToAllocation(Operation *op, Value value) const { return true; }
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto generateOp = cast<tensor::GenerateOp>(op);
auto type = generateOp.getResult().getType();
if (options.defaultMemorySpaceFn(type) != Attribute())
return op->emitError("memory space not implemented yet");
Location loc = op->getLoc();
FailureOr<Value> tensorAlloc = allocateTensorForShapedValue(
rewriter, loc, generateOp.getResult(), options,
false);
if (failed(tensorAlloc))
return failure();
Value result = lowerGenerateLikeOpBody(rewriter, loc, *tensorAlloc,
generateOp.getDynamicExtents(),
generateOp.getBody());
rewriter.replaceOp(generateOp, result);
return success();
}
};
struct InsertOpInterface
: public DstBufferizableOpInterfaceExternalModel<InsertOpInterface,
tensor::InsertOp> {
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto insertOp = cast<tensor::InsertOp>(op);
FailureOr<Value> destMemref =
getBuffer(rewriter, insertOp.getDest(), options);
if (failed(destMemref))
return failure();
rewriter.create<memref::StoreOp>(insertOp.getLoc(), insertOp.getScalar(),
*destMemref, insertOp.getIndices());
replaceOpWithBufferizedValues(rewriter, op, *destMemref);
return success();
}
};
struct InsertSliceOpInterface
: public DstBufferizableOpInterfaceExternalModel<InsertSliceOpInterface,
tensor::InsertSliceOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
auto insertSliceOp = cast<tensor::InsertSliceOp>(op);
RankedTensorType destType = insertSliceOp.getDestType();
if (opOperand == insertSliceOp.getSourceMutable())
return true;
assert(opOperand == insertSliceOp.getDestMutable() && "expected dest");
bool allOffsetsZero =
llvm::all_of(insertSliceOp.getMixedOffsets(), [](OpFoldResult ofr) {
return isConstantIntValue(ofr, 0);
});
bool sizesMatchDestSizes = llvm::all_of(
llvm::enumerate(insertSliceOp.getMixedSizes()), [&](const auto &it) {
return getConstantIntValue(it.value()) ==
destType.getDimSize(it.index());
});
bool allStridesOne =
llvm::all_of(insertSliceOp.getMixedStrides(), [](OpFoldResult ofr) {
return isConstantIntValue(ofr, 1);
});
return !(allOffsetsZero && sizesMatchDestSizes && allStridesOne);
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto insertSliceOp = cast<tensor::InsertSliceOp>(op);
SmallVector<OpFoldResult> mixedOffsets = insertSliceOp.getMixedOffsets();
SmallVector<OpFoldResult> mixedSizes = insertSliceOp.getMixedSizes();
SmallVector<OpFoldResult> mixedStrides = insertSliceOp.getMixedStrides();
Location loc = insertSliceOp.getLoc();
FailureOr<Value> dstMemref =
getBuffer(rewriter, insertSliceOp.getDest(), options);
if (failed(dstMemref))
return failure();
auto dstMemrefType = cast<MemRefType>(dstMemref->getType());
auto subviewMemRefType =
cast<MemRefType>(memref::SubViewOp::inferRankReducedResultType(
insertSliceOp.getSourceType().getShape(), dstMemrefType,
mixedOffsets, mixedSizes, mixedStrides));
Value subView = rewriter.create<memref::SubViewOp>(
loc, subviewMemRefType, *dstMemref, mixedOffsets, mixedSizes,
mixedStrides);
FailureOr<Value> srcMemref =
getBuffer(rewriter, insertSliceOp.getSource(), options);
if (failed(srcMemref))
return failure();
if (failed(options.createMemCpy(rewriter, loc, *srcMemref, subView)))
return failure();
replaceOpWithBufferizedValues(rewriter, op, *dstMemref);
return success();
}
};
struct PadOpInterface
: public BufferizableOpInterface::ExternalModel<PadOpInterface,
tensor::PadOp> {
bool bufferizesToAllocation(Operation *op, Value value) const { return true; }
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
FailureOr<BaseMemRefType>
getBufferType(Operation *op, Value value, const BufferizationOptions &options,
SmallVector<Value> &invocationStack) const {
auto padOp = cast<tensor::PadOp>(op);
auto maybeSrcBufferType = bufferization::getBufferType(
padOp.getSource(), options, invocationStack);
if (failed(maybeSrcBufferType))
return failure();
MemRefLayoutAttrInterface layout;
return MemRefType::get(padOp.getResultType().getShape(),
padOp.getResultType().getElementType(), layout,
maybeSrcBufferType->getMemorySpace());
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto padOp = cast<tensor::PadOp>(op);
Location loc = padOp.getLoc();
RankedTensorType resultType = padOp.getResultType();
RankedTensorType srcType = padOp.getSourceType();
auto toValue = [&](OpFoldResult ofr) {
if (ofr.is<Value>())
return ofr.get<Value>();
return rewriter
.create<arith::ConstantIndexOp>(loc, *getConstantIntValue(ofr))
.getResult();
};
SmallVector<OpFoldResult> mixedLowPad = padOp.getMixedLowPad();
SmallVector<OpFoldResult> mixedHighPad = padOp.getMixedHighPad();
SmallVector<Value> dynamicSizes;
for (int64_t i = 0; i < resultType.getRank(); ++i) {
if (!resultType.isDynamicDim(i))
continue;
Value srcDim = rewriter.create<tensor::DimOp>(loc, padOp.getSource(), i);
Value lowPad = toValue(mixedLowPad[i]);
Value highPad = toValue(mixedHighPad[i]);
AffineExpr s0, s1, s2;
bindSymbols(op->getContext(), s0, s1, s2);
AffineExpr sumExpr = s0 + s1 + s2;
Value sum = rewriter.create<affine::AffineApplyOp>(
loc, sumExpr, ValueRange{srcDim, lowPad, highPad});
dynamicSizes.push_back(sum);
}
FailureOr<Value> tensorAlloc =
allocateTensorForShapedValue(rewriter, loc, padOp.getResult(), options,
false);
if (failed(tensorAlloc))
return failure();
Value filledBuffer = lowerGenerateLikeOpBody(
rewriter, loc, *tensorAlloc, dynamicSizes, padOp.getBodyRegion());
SmallVector<OpFoldResult> sliceSizes =
getMixedSizes(rewriter, loc, padOp.getSource());
SmallVector<OpFoldResult> sliceStrides(srcType.getRank(),
rewriter.getIndexAttr(1));
rewriter.replaceOpWithNewOp<tensor::InsertSliceOp>(
padOp, padOp.getSource(), filledBuffer,
padOp.getMixedLowPad(), sliceSizes, sliceStrides);
return success();
}
};
struct RankOpInterface
: public BufferizableOpInterface::ExternalModel<RankOpInterface,
tensor::RankOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto rankOp = cast<tensor::RankOp>(op);
FailureOr<Value> v = getBuffer(rewriter, rankOp.getTensor(), options);
if (failed(v))
return failure();
replaceOpWithNewBufferizedOp<memref::RankOp>(rewriter, op, rankOp.getType(),
*v);
return success();
}
};
struct ReshapeOpInterface
: public BufferizableOpInterface::ExternalModel<ReshapeOpInterface,
tensor::ReshapeOp> {
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
auto reshapeOp = cast<tensor::ReshapeOp>(op);
return opOperand == reshapeOp.getShapeMutable();
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return false;
}
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {{op->getOpResult(0), BufferRelation::Equivalent}};
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
auto reshapeOp = cast<tensor::ReshapeOp>(op);
FailureOr<Value> srcBuffer =
getBuffer(rewriter, reshapeOp.getSource(), options);
FailureOr<Value> shapeBuffer =
getBuffer(rewriter, reshapeOp.getShape(), options);
if (failed(srcBuffer) || failed(shapeBuffer))
return failure();
auto maybeResultMemRefType =
bufferization::getBufferType(reshapeOp.getResult(), options);
if (failed(maybeResultMemRefType))
return failure();
auto srcType = llvm::dyn_cast<MemRefType>(srcBuffer->getType());
if (srcType && !srcType.getLayout().isIdentity()) {
FailureOr<Value> tensorAlloc = allocateTensorForShapedValue(
rewriter, op->getLoc(), reshapeOp.getSource(), options);
if (failed(tensorAlloc))
return failure();
auto memrefType = MemRefType::get(
srcType.getShape(), srcType.getElementType(), AffineMap(),
cast<BaseMemRefType>(srcBuffer->getType()).getMemorySpace());
srcBuffer = rewriter
.create<bufferization::ToMemrefOp>(
op->getLoc(), memrefType, *tensorAlloc)
.getResult();
}
replaceOpWithNewBufferizedOp<memref::ReshapeOp>(
rewriter, op, maybeResultMemRefType.value(), *srcBuffer, *shapeBuffer);
return success();
}
FailureOr<BaseMemRefType>
getBufferType(Operation *op, Value value, const BufferizationOptions &options,
SmallVector<Value> &invocationStack) const {
auto reshapeOp = cast<tensor::ReshapeOp>(op);
assert(value == reshapeOp.getResult() && "unexpected value provided");
auto maybeSourceBufferType = bufferization::getBufferType(
reshapeOp.getSource(), options, invocationStack);
if (failed(maybeSourceBufferType))
return failure();
return getMemRefTypeWithStaticIdentityLayout(
reshapeOp.getResult().getType(),
cast<BaseMemRefType>(maybeSourceBufferType.value()).getMemorySpace());
}
};
struct ParallelInsertSliceOpInterface
: public BufferizableOpInterface::ExternalModel<
ParallelInsertSliceOpInterface, ParallelInsertSliceOp> {
AliasingValueList getAliasingValues(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return {};
}
bool bufferizesToMemoryRead(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
return true;
}
bool bufferizesToMemoryWrite(Operation *op, OpOperand &opOperand,
const AnalysisState &state) const {
auto parallelInsertSliceOp = cast<ParallelInsertSliceOp>(op);
return opOperand == parallelInsertSliceOp.getDestMutable();
}
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
OpBuilder::InsertionGuard g(rewriter);
auto parallelInsertSliceOp = cast<ParallelInsertSliceOp>(op);
ParallelCombiningOpInterface parallelCombiningParent =
parallelInsertSliceOp.getParallelCombiningParent();
rewriter.setInsertionPoint(parallelCombiningParent);
FailureOr<Value> destBuffer =
getBuffer(rewriter, parallelInsertSliceOp.getDest(), options);
if (failed(destBuffer))
return failure();
FailureOr<Value> srcBuffer =
getBuffer(rewriter, parallelInsertSliceOp.getSource(), options);
if (failed(srcBuffer))
return failure();
auto destBufferType = cast<MemRefType>(destBuffer->getType());
auto subviewMemRefType =
cast<MemRefType>(memref::SubViewOp::inferRankReducedResultType(
parallelInsertSliceOp.getSourceType().getShape(), destBufferType,
parallelInsertSliceOp.getMixedOffsets(),
parallelInsertSliceOp.getMixedSizes(),
parallelInsertSliceOp.getMixedStrides()));
Value subview = rewriter.create<memref::SubViewOp>(
parallelInsertSliceOp.getLoc(), subviewMemRefType, *destBuffer,
parallelInsertSliceOp.getMixedOffsets(),
parallelInsertSliceOp.getMixedSizes(),
parallelInsertSliceOp.getMixedStrides());
if (failed(options.createMemCpy(rewriter, parallelInsertSliceOp.getLoc(),
*srcBuffer, subview)))
return failure();
for (Operation *user : srcBuffer->getUsers()) {
if (hasEffect<MemoryEffects::Free>(user)) {
if (user->getBlock() == parallelCombiningParent->getBlock())
rewriter.moveOpBefore(user, user->getBlock()->getTerminator());
break;
}
}
rewriter.eraseOp(op);
return success();
}
LogicalResult resolveConflicts(Operation *op, RewriterBase &rewriter,
const AnalysisState &state) const {
return success();
}
};
struct SplatOpInterface
: public BufferizableOpInterface::ExternalModel<SplatOpInterface,
tensor::SplatOp> {
bool bufferizesToAllocation(Operation *op, Value value) const { return true; }
LogicalResult bufferize(Operation *op, RewriterBase &rewriter,
const BufferizationOptions &options) const {
OpBuilder::InsertionGuard g(rewriter);
auto splatOp = cast<tensor::SplatOp>(op);
Location loc = op->getLoc();
FailureOr<Value> tensorAlloc = allocateTensorForShapedValue(
rewriter, loc, splatOp.getResult(), options,
false);
if (failed(tensorAlloc))
return failure();
auto tensorType = cast<RankedTensorType>(tensorAlloc->getType());
if (options.defaultMemorySpaceFn(tensorType) != Attribute())
return op->emitError("memory space not implemented yet");
auto linalgOp =
rewriter.create<linalg::MapOp>(loc, tensorType, ValueRange(),
*tensorAlloc);
Block &linalgBody = linalgOp.getMapper().emplaceBlock();
rewriter.setInsertionPointToStart(&linalgBody);
rewriter.create<linalg::YieldOp>(loc, splatOp.getInput());
rewriter.replaceOp(splatOp, linalgOp.getResult()[0]);
return success();
}
};
}
}
}
void mlir::tensor::registerBufferizableOpInterfaceExternalModels(
DialectRegistry ®istry) {
registry.addExtension(+[](MLIRContext *ctx, tensor::TensorDialect *dialect) {
CastOp::attachInterface<CastOpInterface>(*ctx);
CollapseShapeOp::attachInterface<CollapseShapeOpInterface>(*ctx);
DimOp::attachInterface<DimOpInterface>(*ctx);
EmptyOp::attachInterface<EmptyOpInterface>(*ctx);
ExpandShapeOp::attachInterface<ExpandShapeOpInterface>(*ctx);
ExtractSliceOp::attachInterface<ExtractSliceOpInterface>(*ctx);
ExtractOp::attachInterface<ExtractOpInterface>(*ctx);
FromElementsOp::attachInterface<FromElementsOpInterface>(*ctx);
GenerateOp::attachInterface<GenerateOpInterface>(*ctx);
InsertOp::attachInterface<InsertOpInterface>(*ctx);
InsertSliceOp::attachInterface<InsertSliceOpInterface>(*ctx);
PadOp::attachInterface<PadOpInterface>(*ctx);
ParallelInsertSliceOp::attachInterface<ParallelInsertSliceOpInterface>(
*ctx);
RankOp::attachInterface<RankOpInterface>(*ctx);
ReshapeOp::attachInterface<ReshapeOpInterface>(*ctx);
SplatOp::attachInterface<SplatOpInterface>(*ctx);
ctx->loadDialect<arith::ArithDialect, linalg::LinalgDialect>();
});
tensor::registerSubsetOpInterfaceExternalModels(registry);
}