diff options
22 files changed, 123 insertions, 118 deletions
diff --git a/mlir/include/mlir/Analysis/Utils.h b/mlir/include/mlir/Analysis/Utils.h index 8c8f73da409..c32583ebc6e 100644 --- a/mlir/include/mlir/Analysis/Utils.h +++ b/mlir/include/mlir/Analysis/Utils.h @@ -81,7 +81,7 @@ struct MemRefRegion { /// bounds major to minor. We use int64_t instead of uint64_t since index /// types can be at most int64_t. Optional<int64_t> getConstantBoundingSizeAndShape( - SmallVectorImpl<int> *shape = nullptr, + SmallVectorImpl<int64_t> *shape = nullptr, std::vector<SmallVector<int64_t, 4>> *lbs = nullptr, SmallVectorImpl<int64_t> *lbDivisors = nullptr) const; diff --git a/mlir/include/mlir/Analysis/VectorAnalysis.h b/mlir/include/mlir/Analysis/VectorAnalysis.h index e34e3433f2f..dfb1164750a 100644 --- a/mlir/include/mlir/Analysis/VectorAnalysis.h +++ b/mlir/include/mlir/Analysis/VectorAnalysis.h @@ -48,7 +48,7 @@ class VectorType; /// - shapeRatio({3, 4, 4, 8}, {2, 5, 2}) returns None /// - shapeRatio({1, 2, 10, 32}, {2, 5, 2}) returns {1, 1, 2, 16} llvm::Optional<llvm::SmallVector<unsigned, 4>> -shapeRatio(ArrayRef<int> superShape, ArrayRef<int> subShape); +shapeRatio(ArrayRef<int64_t> superShape, ArrayRef<int64_t> subShape); /// Computes and returns the multi-dimensional ratio of the shapes of /// `superVector` to `subVector`. If integral division is not possible, returns diff --git a/mlir/include/mlir/IR/Builders.h b/mlir/include/mlir/IR/Builders.h index 3fa2e3ff4a9..3814e5fc6d9 100644 --- a/mlir/include/mlir/IR/Builders.h +++ b/mlir/include/mlir/IR/Builders.h @@ -80,11 +80,11 @@ public: IntegerType getI1Type(); IntegerType getIntegerType(unsigned width); FunctionType getFunctionType(ArrayRef<Type> inputs, ArrayRef<Type> results); - MemRefType getMemRefType(ArrayRef<int> shape, Type elementType, + MemRefType getMemRefType(ArrayRef<int64_t> shape, Type elementType, ArrayRef<AffineMap> affineMapComposition = {}, unsigned memorySpace = 0); - VectorType getVectorType(ArrayRef<int> shape, Type elementType); - RankedTensorType getTensorType(ArrayRef<int> shape, Type elementType); + VectorType getVectorType(ArrayRef<int64_t> shape, Type elementType); + RankedTensorType getTensorType(ArrayRef<int64_t> shape, Type elementType); UnrankedTensorType getTensorType(Type elementType); /// Get or construct an instance of the type 'ty' with provided arguments. diff --git a/mlir/include/mlir/IR/StandardTypes.h b/mlir/include/mlir/IR/StandardTypes.h index 823ea75d403..3a12f8016f1 100644 --- a/mlir/include/mlir/IR/StandardTypes.h +++ b/mlir/include/mlir/IR/StandardTypes.h @@ -186,11 +186,11 @@ public: /// If this is ranked tensor or vector type, return the rank. If it is an /// unranked tensor, return -1. - int getRank() const; + int64_t getRank() const; /// If this is ranked tensor or vector type, return the shape. If it is an /// unranked tensor, abort. - ArrayRef<int> getShape() const; + ArrayRef<int64_t> getShape() const; /// If this is unranked tensor or any dimension has unknown size (<0), /// it doesn't have static shape. If all dimensions have known size (>= 0), @@ -200,7 +200,7 @@ public: /// If this is ranked tensor or vector type, return the size of the specified /// dimension. It aborts if the tensor is unranked (this can be checked by /// the getRank call method). - int getDimSize(unsigned i) const; + int64_t getDimSize(unsigned i) const; /// Get the total amount of bits occupied by a value of this type. This does /// not take into account any memory layout or widening constraints, e.g. a @@ -208,7 +208,7 @@ public: /// it will likely be stored as in a 4xi64 vector register. Fail an assertion /// if the size cannot be computed statically, i.e. if the tensor has a /// dynamic shape or if its elemental type does not have a known bit width. - long getSizeInBits() const; + int64_t getSizeInBits() const; /// Methods for support type inquiry through isa, cast, and dyn_cast. static bool kindof(unsigned kind) { @@ -227,26 +227,26 @@ public: /// Get or create a new VectorType of the provided shape and element type. /// Assumes the arguments define a well-formed VectorType. - static VectorType get(ArrayRef<int> shape, Type elementType); + static VectorType get(ArrayRef<int64_t> shape, Type elementType); /// Get or create a new VectorType of the provided shape and element type /// declared at the given, potentially unknown, location. If the VectorType /// defined by the arguments would be ill-formed, emit errors and return /// nullptr-wrapping type. - static VectorType getChecked(ArrayRef<int> shape, Type elementType, + static VectorType getChecked(ArrayRef<int64_t> shape, Type elementType, Location location); /// Verify the construction of a vector type. static bool verifyConstructionInvariants(llvm::Optional<Location> loc, MLIRContext *context, - ArrayRef<int> shape, + ArrayRef<int64_t> shape, Type elementType); /// Returns true of the given type can be used as an element of a vector type. /// In particular, vectors can consist of integer or float primitives. static bool isValidElementType(Type t) { return t.isIntOrFloat(); } - ArrayRef<int> getShape() const; + ArrayRef<int64_t> getShape() const; /// Methods for support type inquiry through isa, cast, and dyn_cast. static bool kindof(unsigned kind) { return kind == StandardTypes::Vector; } @@ -290,22 +290,22 @@ public: /// Get or create a new RankedTensorType of the provided shape and element /// type. Assumes the arguments define a well-formed type. - static RankedTensorType get(ArrayRef<int> shape, Type elementType); + static RankedTensorType get(ArrayRef<int64_t> shape, Type elementType); /// Get or create a new RankedTensorType of the provided shape and element /// type declared at the given, potentially unknown, location. If the /// RankedTensorType defined by the arguments would be ill-formed, emit errors /// and return a nullptr-wrapping type. - static RankedTensorType getChecked(ArrayRef<int> shape, Type elementType, + static RankedTensorType getChecked(ArrayRef<int64_t> shape, Type elementType, Location location); /// Verify the construction of a ranked tensor type. static bool verifyConstructionInvariants(llvm::Optional<Location> loc, MLIRContext *context, - ArrayRef<int> shape, + ArrayRef<int64_t> shape, Type elementType); - ArrayRef<int> getShape() const; + ArrayRef<int64_t> getShape() const; static bool kindof(unsigned kind) { return kind == StandardTypes::RankedTensor; @@ -338,7 +338,7 @@ public: MLIRContext *context, Type elementType); - ArrayRef<int> getShape() const { return ArrayRef<int>(); } + ArrayRef<int64_t> getShape() const { return llvm::None; } static bool kindof(unsigned kind) { return kind == StandardTypes::UnrankedTensor; @@ -361,7 +361,7 @@ public: /// map composition, and memory space. Assumes the arguments define a /// well-formed MemRef type. Use getChecked to gracefully handle MemRefType /// construction failures. - static MemRefType get(ArrayRef<int> shape, Type elementType, + static MemRefType get(ArrayRef<int64_t> shape, Type elementType, ArrayRef<AffineMap> affineMapComposition, unsigned memorySpace) { auto result = getImpl(shape, elementType, affineMapComposition, memorySpace, @@ -376,7 +376,7 @@ public: /// UnknownLoc. If the MemRefType defined by the arguments would be /// ill-formed, emits errors (to the handler registered with the context or to /// the error stream) and returns nullptr. - static MemRefType getChecked(ArrayRef<int> shape, Type elementType, + static MemRefType getChecked(ArrayRef<int64_t> shape, Type elementType, ArrayRef<AffineMap> affineMapComposition, unsigned memorySpace, Location location) { return getImpl(shape, elementType, affineMapComposition, memorySpace, @@ -386,10 +386,10 @@ public: unsigned getRank() const { return getShape().size(); } /// Returns an array of memref shape dimension sizes. - ArrayRef<int> getShape() const; + ArrayRef<int64_t> getShape() const; /// Return the size of the specified dimension, or -1 if unspecified. - int getDimSize(unsigned i) const { return getShape()[i]; } + int64_t getDimSize(unsigned i) const { return getShape()[i]; } /// Returns the elemental type for this memref shape. Type getElementType() const; @@ -404,6 +404,10 @@ public: /// Returns the number of dimensions with dynamic size. unsigned getNumDynamicDims() const; + /// If any dimension of the shape has unknown size (<0), it doesn't have + /// static shape. + bool hasStaticShape() const { return getNumDynamicDims() == 0; } + static bool kindof(unsigned kind) { return kind == StandardTypes::MemRef; } /// Unique identifier for this type class. @@ -413,7 +417,7 @@ private: /// Get or create a new MemRefType defined by the arguments. If the resulting /// type would be ill-formed, return nullptr. If the location is provided, /// emit detailed error messages. - static MemRefType getImpl(ArrayRef<int> shape, Type elementType, + static MemRefType getImpl(ArrayRef<int64_t> shape, Type elementType, ArrayRef<AffineMap> affineMapComposition, unsigned memorySpace, Optional<Location> location); }; diff --git a/mlir/lib/Analysis/Utils.cpp b/mlir/lib/Analysis/Utils.cpp index 79d1b696612..6c33bdee6aa 100644 --- a/mlir/lib/Analysis/Utils.cpp +++ b/mlir/lib/Analysis/Utils.cpp @@ -56,7 +56,7 @@ unsigned MemRefRegion::getRank() const { } Optional<int64_t> MemRefRegion::getConstantBoundingSizeAndShape( - SmallVectorImpl<int> *shape, std::vector<SmallVector<int64_t, 4>> *lbs, + SmallVectorImpl<int64_t> *shape, std::vector<SmallVector<int64_t, 4>> *lbs, SmallVectorImpl<int64_t> *lbDivisors) const { auto memRefType = memref->getType().cast<MemRefType>(); unsigned rank = memRefType.getRank(); @@ -289,7 +289,7 @@ bool mlir::boundCheckLoadOrStoreOp(LoadOrStoreOpPointer loadOrStoreOp, // of upper and out of lower), and check if the constraint system is // feasible. If it is, there is at least one point out of bounds. SmallVector<int64_t, 4> ineq(rank + 1, 0); - int dimSize = loadOrStoreOp->getMemRefType().getDimSize(r); + int64_t dimSize = loadOrStoreOp->getMemRefType().getDimSize(r); // TODO(bondhugula): handle dynamic dim sizes. if (dimSize == -1) continue; diff --git a/mlir/lib/Analysis/VectorAnalysis.cpp b/mlir/lib/Analysis/VectorAnalysis.cpp index bc43e2ca5eb..37eed71508f 100644 --- a/mlir/lib/Analysis/VectorAnalysis.cpp +++ b/mlir/lib/Analysis/VectorAnalysis.cpp @@ -37,8 +37,8 @@ using namespace mlir; using llvm::SetVector; -Optional<SmallVector<unsigned, 4>> mlir::shapeRatio(ArrayRef<int> superShape, - ArrayRef<int> subShape) { +Optional<SmallVector<unsigned, 4>> +mlir::shapeRatio(ArrayRef<int64_t> superShape, ArrayRef<int64_t> subShape) { if (superShape.size() < subShape.size()) { return Optional<SmallVector<unsigned, 4>>(); } @@ -55,8 +55,8 @@ Optional<SmallVector<unsigned, 4>> mlir::shapeRatio(ArrayRef<int> superShape, result.push_back(superSize / subSize); }; functional::zipApply( - divide, SmallVector<int, 8>{superShape.rbegin(), superShape.rend()}, - SmallVector<int, 8>{subShape.rbegin(), subShape.rend()}); + divide, SmallVector<int64_t, 8>{superShape.rbegin(), superShape.rend()}, + SmallVector<int64_t, 8>{subShape.rbegin(), subShape.rend()}); // If integral division does not occur, return and let the caller decide. if (!divides) { diff --git a/mlir/lib/Dialect/Traits.cpp b/mlir/lib/Dialect/Traits.cpp index 6beb8cba41d..0fa57c23364 100644 --- a/mlir/lib/Dialect/Traits.cpp +++ b/mlir/lib/Dialect/Traits.cpp @@ -86,7 +86,7 @@ Type OpTrait::util::getBroadcastedType(Type type1, Type type2) { } // Returns the shape of the given type. - auto getShape = [](Type type) -> ArrayRef<int> { + auto getShape = [](Type type) -> ArrayRef<int64_t> { if (auto vtType = type.dyn_cast<VectorOrTensorType>()) return vtType.getShape(); return {}; @@ -104,7 +104,7 @@ Type OpTrait::util::getBroadcastedType(Type type1, Type type2) { // The result shape has the maximum among the two inputs at every // dimension index. - SmallVector<int, 4> resultShape; + SmallVector<int64_t, 4> resultShape; if (shape1.size() > shape2.size()) { std::copy(shape1.begin(), shape1.end(), std::back_inserter(resultShape)); } else { diff --git a/mlir/lib/IR/Attributes.cpp b/mlir/lib/IR/Attributes.cpp index b3ff25035de..e00297a6bb0 100644 --- a/mlir/lib/IR/Attributes.cpp +++ b/mlir/lib/IR/Attributes.cpp @@ -168,7 +168,7 @@ Attribute DenseElementsAttr::getValue(ArrayRef<uint64_t> index) const { // Reduce the provided multidimensional index into a 1D index. uint64_t valueIndex = 0; uint64_t dimMultiplier = 1; - for (int i = rank - 1; i >= 0; --i) { + for (auto i = rank - 1; i >= 0; --i) { valueIndex += index[i] * dimMultiplier; dimMultiplier *= shape[i]; } @@ -346,7 +346,7 @@ Attribute SparseElementsAttr::getValue(ArrayRef<uint64_t> index) const { // Build a mapping between known indices and the offset of the stored element. llvm::SmallDenseMap<llvm::ArrayRef<uint64_t>, size_t> mappedIndices; - size_t numSparseIndices = sparseIndices.getType().getDimSize(0); + auto numSparseIndices = sparseIndices.getType().getDimSize(0); for (size_t i = 0, e = numSparseIndices; i != e; ++i) mappedIndices.try_emplace( {sparseIndexValues + (i * rank), static_cast<size_t>(rank)}, i); diff --git a/mlir/lib/IR/Builders.cpp b/mlir/lib/IR/Builders.cpp index 2e453f0c6e8..6a513352afc 100644 --- a/mlir/lib/IR/Builders.cpp +++ b/mlir/lib/IR/Builders.cpp @@ -77,17 +77,18 @@ FunctionType Builder::getFunctionType(ArrayRef<Type> inputs, return FunctionType::get(inputs, results, context); } -MemRefType Builder::getMemRefType(ArrayRef<int> shape, Type elementType, +MemRefType Builder::getMemRefType(ArrayRef<int64_t> shape, Type elementType, ArrayRef<AffineMap> affineMapComposition, unsigned memorySpace) { return MemRefType::get(shape, elementType, affineMapComposition, memorySpace); } -VectorType Builder::getVectorType(ArrayRef<int> shape, Type elementType) { +VectorType Builder::getVectorType(ArrayRef<int64_t> shape, Type elementType) { return VectorType::get(shape, elementType); } -RankedTensorType Builder::getTensorType(ArrayRef<int> shape, Type elementType) { +RankedTensorType Builder::getTensorType(ArrayRef<int64_t> shape, + Type elementType) { return RankedTensorType::get(shape, elementType); } diff --git a/mlir/lib/IR/StandardTypes.cpp b/mlir/lib/IR/StandardTypes.cpp index f031a11859f..7b7041fe428 100644 --- a/mlir/lib/IR/StandardTypes.cpp +++ b/mlir/lib/IR/StandardTypes.cpp @@ -112,6 +112,7 @@ unsigned VectorOrTensorType::getNumElements() const { switch (getKind()) { case StandardTypes::Vector: case StandardTypes::RankedTensor: { + assert(hasStaticShape() && "expected type to have static shape"); auto shape = getShape(); unsigned num = 1; for (auto dim : shape) @@ -125,7 +126,7 @@ unsigned VectorOrTensorType::getNumElements() const { /// If this is ranked tensor or vector type, return the rank. If it is an /// unranked tensor, return -1. -int VectorOrTensorType::getRank() const { +int64_t VectorOrTensorType::getRank() const { switch (getKind()) { case StandardTypes::Vector: case StandardTypes::RankedTensor: @@ -137,7 +138,7 @@ int VectorOrTensorType::getRank() const { } } -int VectorOrTensorType::getDimSize(unsigned i) const { +int64_t VectorOrTensorType::getDimSize(unsigned i) const { switch (getKind()) { case StandardTypes::Vector: case StandardTypes::RankedTensor: @@ -150,7 +151,7 @@ int VectorOrTensorType::getDimSize(unsigned i) const { // Get the number of number of bits require to store a value of the given vector // or tensor types. Compute the value recursively since tensors are allowed to // have vectors as elements. -long VectorOrTensorType::getSizeInBits() const { +int64_t VectorOrTensorType::getSizeInBits() const { assert(hasStaticShape() && "cannot get the bit size of an aggregate with a dynamic shape"); @@ -165,7 +166,7 @@ long VectorOrTensorType::getSizeInBits() const { return getNumElements() * elementVectorOrTensorType.getSizeInBits(); } -ArrayRef<int> VectorOrTensorType::getShape() const { +ArrayRef<int64_t> VectorOrTensorType::getShape() const { switch (getKind()) { case StandardTypes::Vector: return cast<VectorType>().getShape(); @@ -179,18 +180,17 @@ ArrayRef<int> VectorOrTensorType::getShape() const { bool VectorOrTensorType::hasStaticShape() const { if (isa<UnrankedTensorType>()) return false; - auto dims = getShape(); - return !std::any_of(dims.begin(), dims.end(), [](int i) { return i < 0; }); + return llvm::none_of(getShape(), [](int64_t i) { return i < 0; }); } /// VectorType -VectorType VectorType::get(ArrayRef<int> shape, Type elementType) { +VectorType VectorType::get(ArrayRef<int64_t> shape, Type elementType) { return Base::get(elementType.getContext(), StandardTypes::Vector, shape, elementType); } -VectorType VectorType::getChecked(ArrayRef<int> shape, Type elementType, +VectorType VectorType::getChecked(ArrayRef<int64_t> shape, Type elementType, Location location) { return Base::getChecked(location, elementType.getContext(), StandardTypes::Vector, shape, elementType); @@ -198,7 +198,7 @@ VectorType VectorType::getChecked(ArrayRef<int> shape, Type elementType, bool VectorType::verifyConstructionInvariants(llvm::Optional<Location> loc, MLIRContext *context, - ArrayRef<int> shape, + ArrayRef<int64_t> shape, Type elementType) { if (shape.empty()) { if (loc) @@ -212,7 +212,7 @@ bool VectorType::verifyConstructionInvariants(llvm::Optional<Location> loc, return true; } - if (any_of(shape, [](int i) { return i < 0; })) { + if (any_of(shape, [](int64_t i) { return i < 0; })) { if (loc) context->emitError(*loc, "vector types must have static shape"); return true; @@ -220,7 +220,7 @@ bool VectorType::verifyConstructionInvariants(llvm::Optional<Location> loc, return false; } -ArrayRef<int> VectorType::getShape() const { +ArrayRef<int64_t> VectorType::getShape() const { return static_cast<ImplType *>(type)->getShape(); } @@ -241,12 +241,13 @@ static inline bool checkTensorElementType(Optional<Location> location, /// RankedTensorType -RankedTensorType RankedTensorType::get(ArrayRef<int> shape, Type elementType) { +RankedTensorType RankedTensorType::get(ArrayRef<int64_t> shape, + Type elementType) { return Base::get(elementType.getContext(), StandardTypes::RankedTensor, shape, elementType); } -RankedTensorType RankedTensorType::getChecked(ArrayRef<int> shape, +RankedTensorType RankedTensorType::getChecked(ArrayRef<int64_t> shape, Type elementType, Location location) { return Base::getChecked(location, elementType.getContext(), @@ -254,16 +255,16 @@ RankedTensorType RankedTensorType::getChecked(ArrayRef<int> shape, } bool RankedTensorType::verifyConstructionInvariants( - llvm::Optional<Location> loc, MLIRContext *context, ArrayRef<int> shape, + llvm::Optional<Location> loc, MLIRContext *context, ArrayRef<int64_t> shape, Type elementType) { return checkTensorElementType(loc, context, elementType); } -ArrayRef<int> RankedTensorType::getShape() const { +ArrayRef<int64_t> RankedTensorType::getShape() const { return static_cast<ImplType *>(type)->getShape(); } -ArrayRef<int> MemRefType::getShape() const { +ArrayRef<int64_t> MemRefType::getShape() const { return static_cast<ImplType *>(type)->getShape(); } @@ -291,7 +292,7 @@ bool UnrankedTensorType::verifyConstructionInvariants( /// type would be ill-formed, return nullptr. If the location is provided, /// emit detailed error messages. To emit errors when the location is unknown, /// pass in an instance of UnknownLoc. -MemRefType MemRefType::getImpl(ArrayRef<int> shape, Type elementType, +MemRefType MemRefType::getImpl(ArrayRef<int64_t> shape, Type elementType, ArrayRef<AffineMap> affineMapComposition, unsigned memorySpace, Optional<Location> location) { @@ -346,12 +347,7 @@ unsigned MemRefType::getMemorySpace() const { } unsigned MemRefType::getNumDynamicDims() const { - unsigned numDynamicDims = 0; - for (int dimSize : getShape()) { - if (dimSize == -1) - ++numDynamicDims; - } - return numDynamicDims; + return llvm::count_if(getShape(), [](int64_t i) { return i < 0; }); } // Define type identifiers. diff --git a/mlir/lib/IR/TypeDetail.h b/mlir/lib/IR/TypeDetail.h index 32a30a6275f..91762df53d6 100644 --- a/mlir/lib/IR/TypeDetail.h +++ b/mlir/lib/IR/TypeDetail.h @@ -131,12 +131,12 @@ struct VectorOrTensorTypeStorage : public TypeStorage { /// Vector Type Storage and Uniquing. struct VectorTypeStorage : public VectorOrTensorTypeStorage { VectorTypeStorage(unsigned shapeSize, Type elementTy, - const int *shapeElements) + const int64_t *shapeElements) : VectorOrTensorTypeStorage(elementTy, shapeSize), shapeElements(shapeElements) {} /// The hash key used for uniquing. - using KeyTy = std::pair<ArrayRef<int>, Type>; + using KeyTy = std::pair<ArrayRef<int64_t>, Type>; bool operator==(const KeyTy &key) const { return key == KeyTy(getShape(), elementType); } @@ -145,28 +145,28 @@ struct VectorTypeStorage : public VectorOrTensorTypeStorage { static VectorTypeStorage *construct(TypeStorageAllocator &allocator, const KeyTy &key) { // Copy the shape into the bump pointer. - ArrayRef<int> shape = allocator.copyInto(key.first); + ArrayRef<int64_t> shape = allocator.copyInto(key.first); // Initialize the memory using placement new. return new (allocator.allocate<VectorTypeStorage>()) VectorTypeStorage(shape.size(), key.second, shape.data()); } - ArrayRef<int> getShape() const { - return ArrayRef<int>(shapeElements, getSubclassData()); + ArrayRef<int64_t> getShape() const { + return ArrayRef<int64_t>(shapeElements, getSubclassData()); } - const int *shapeElements; + const int64_t *shapeElements; }; struct RankedTensorTypeStorage : public VectorOrTensorTypeStorage { RankedTensorTypeStorage(unsigned shapeSize, Type elementTy, - const int *shapeElements) + const int64_t *shapeElements) : VectorOrTensorTypeStorage(elementTy, shapeSize), shapeElements(shapeElements) {} /// The hash key used for uniquing. - using KeyTy = std::pair<ArrayRef<int>, Type>; + using KeyTy = std::pair<ArrayRef<int64_t>, Type>; bool operator==(const KeyTy &key) const { return key == KeyTy(getShape(), elementType); } @@ -175,18 +175,18 @@ struct RankedTensorTypeStorage : public VectorOrTensorTypeStorage { static RankedTensorTypeStorage *construct(TypeStorageAllocator &allocator, const KeyTy &key) { // Copy the shape into the bump pointer. - ArrayRef<int> shape = allocator.copyInto(key.first); + ArrayRef<int64_t> shape = allocator.copyInto(key.first); // Initialize the memory using placement new. return new (allocator.allocate<RankedTensorTypeStorage>()) RankedTensorTypeStorage(shape.size(), key.second, shape.data()); } - ArrayRef<int> getShape() const { - return ArrayRef<int>(shapeElements, getSubclassData()); + ArrayRef<int64_t> getShape() const { + return ArrayRef<int64_t>(shapeElements, getSubclassData()); } - const int *shapeElements; + const int64_t *shapeElements; }; struct UnrankedTensorTypeStorage : public VectorOrTensorTypeStorage { @@ -203,7 +203,7 @@ struct UnrankedTensorTypeStorage : public VectorOrTensorTypeStorage { struct MemRefTypeStorage : public TypeStorage { MemRefTypeStorage(unsigned shapeSize, Type elementType, - const int *shapeElements, const unsigned numAffineMaps, + const int64_t *shapeElements, const unsigned numAffineMaps, AffineMap const *affineMapList, const unsigned memorySpace) : TypeStorage(shapeSize), elementType(elementType), shapeElements(shapeElements), numAffineMaps(numAffineMaps), @@ -212,7 +212,8 @@ struct MemRefTypeStorage : public TypeStorage { /// The hash key used for uniquing. // MemRefs are uniqued based on their shape, element type, affine map // composition, and memory space. - using KeyTy = std::tuple<ArrayRef<int>, Type, ArrayRef<AffineMap>, unsigned>; + using KeyTy = + std::tuple<ArrayRef<int64_t>, Type, ArrayRef<AffineMap>, unsigned>; bool operator==(const KeyTy &key) const { return key == KeyTy(getShape(), elementType, getAffineMaps(), memorySpace); } @@ -221,7 +222,7 @@ struct MemRefTypeStorage : public TypeStorage { static MemRefTypeStorage *construct(TypeStorageAllocator &allocator, const KeyTy &key) { // Copy the shape into the bump pointer. - ArrayRef<int> shape = allocator.copyInto(std::get<0>(key)); + ArrayRef<int64_t> shape = allocator.copyInto(std::get<0>(key)); // Copy the affine map composition into the bump pointer. ArrayRef<AffineMap> affineMapComposition = @@ -234,8 +235,8 @@ struct MemRefTypeStorage : public TypeStorage { affineMapComposition.data(), std::get<3>(key)); } - ArrayRef<int> getShape() const { - return ArrayRef<int>(shapeElements, getSubclassData()); + ArrayRef<int64_t> getShape() const { + return ArrayRef<int64_t>(shapeElements, getSubclassData()); } ArrayRef<AffineMap> getAffineMaps() const { @@ -245,7 +246,7 @@ struct MemRefTypeStorage : public TypeStorage { /// The type of each scalar element of the memref. Type elementType; /// An array of integers which stores the shape dimension sizes. - const int *shapeElements; + const int64_t *shapeElements; /// The number of affine maps in the 'affineMapList' array. const unsigned numAffineMaps; /// List of affine maps in the memref's layout/index map composition. diff --git a/mlir/lib/Parser/Parser.cpp b/mlir/lib/Parser/Parser.cpp index 7766c5900c2..6a96021dcb5 100644 --- a/mlir/lib/Parser/Parser.cpp +++ b/mlir/lib/Parser/Parser.cpp @@ -183,7 +183,7 @@ public: // Type parsing. VectorType parseVectorType(); ParseResult parseXInDimensionList(); - ParseResult parseDimensionListRanked(SmallVectorImpl<int> &dimensions); + ParseResult parseDimensionListRanked(SmallVectorImpl<int64_t> &dimensions); Type parseExtendedType(); Type parseTensorType(); Type parseMemRefType(); @@ -386,13 +386,13 @@ VectorType Parser::parseVectorType() { if (getToken().isNot(Token::integer)) return (emitError("expected dimension size in vector type"), nullptr); - SmallVector<int, 4> dimensions; + SmallVector<int64_t, 4> dimensions; while (getToken().is(Token::integer)) { // Make sure this integer value is in bound and valid. auto dimension = getToken().getUnsignedIntegerValue(); if (!dimension.hasValue()) return (emitError("invalid dimension in vector type"), nullptr); - dimensions.push_back((int)dimension.getValue()); + dimensions.push_back((int64_t)dimension.getValue()); consumeToken(Token::integer); @@ -442,16 +442,17 @@ ParseResult Parser::parseXInDimensionList() { /// dimension-list-ranked ::= (dimension `x`)* /// dimension ::= `?` | integer-literal /// -ParseResult Parser::parseDimensionListRanked(SmallVectorImpl<int> &dimensions) { +ParseResult +Parser::parseDimensionListRanked(SmallVectorImpl<int64_t> &dimensions) { while (getToken().isAny(Token::integer, Token::question)) { if (consumeIf(Token::question)) { dimensions.push_back(-1); } else { // Make sure this integer value is in bound and valid. auto dimension = getToken().getUnsignedIntegerValue(); - if (!dimension.hasValue() || (int)dimension.getValue() < 0) + if (!dimension.hasValue() || (int64_t)dimension.getValue() < 0) return emitError("invalid dimension"); - dimensions.push_back((int)dimension.getValue()); + dimensions.push_back((int64_t)dimension.getValue()); consumeToken(Token::integer); } @@ -540,7 +541,7 @@ Type Parser::parseTensorType() { return nullptr; bool isUnranked; - SmallVector<int, 4> dimensions; + SmallVector<int64_t, 4> dimensions; if (consumeIf(Token::star)) { // This is an unranked tensor type. @@ -580,7 +581,7 @@ Type Parser::parseMemRefType() { if (parseToken(Token::less, "expected '<' in memref type")) return nullptr; - SmallVector<int, 4> dimensions; + SmallVector<int64_t, 4> dimensions; if (parseDimensionListRanked(dimensions)) return nullptr; @@ -706,12 +707,12 @@ public: ArrayRef<Attribute> getValues() const { return storage; } - ArrayRef<int> getShape() const { return shape; } + ArrayRef<int64_t> getShape() const { return shape; } private: /// Parse either a single element or a list of elements. Return the dimensions /// of the parsed sub-tensor in dims. - ParseResult parseElementOrList(llvm::SmallVectorImpl<int> &dims); + ParseResult parseElementOrList(llvm::SmallVectorImpl<int64_t> &dims); /// Parse a list of either lists or elements, returning the dimensions of the /// parsed sub-tensors in dims. For example: @@ -719,11 +720,11 @@ private: /// parseList([[1, 2], [3, 4]]) -> Success, [2, 2] /// parseList([[1, 2], 3]) -> Failure /// parseList([[1, [2, 3]], [4, [5]]]) -> Failure - ParseResult parseList(llvm::SmallVectorImpl<int> &dims); + ParseResult parseList(llvm::SmallVectorImpl<int64_t> &dims); Parser &p; Type eltTy; - SmallVector<int, 4> shape; + SmallVector<int64_t, 4> shape; std::vector<Attribute> storage; }; } // namespace @@ -731,7 +732,7 @@ private: /// Parse either a single element or a list of elements. Return the dimensions /// of the parsed sub-tensor in dims. ParseResult -TensorLiteralParser::parseElementOrList(llvm::SmallVectorImpl<int> &dims) { +TensorLiteralParser::parseElementOrList(llvm::SmallVectorImpl<int64_t> &dims) { switch (p.getToken().getKind()) { case Token::l_square: return parseList(dims); @@ -789,11 +790,12 @@ TensorLiteralParser::parseElementOrList(llvm::SmallVectorImpl<int> &dims) { /// parseList([[1, 2], [3, 4]]) -> Success, [2, 2] /// parseList([[1, 2], 3]) -> Failure /// parseList([[1, [2, 3]], [4, [5]]]) -> Failure -ParseResult TensorLiteralParser::parseList(llvm::SmallVectorImpl<int> &dims) { +ParseResult +TensorLiteralParser::parseList(llvm::SmallVectorImpl<int64_t> &dims) { p.consumeToken(Token::l_square); - auto checkDims = [&](const llvm::SmallVectorImpl<int> &prevDims, - const llvm::SmallVectorImpl<int> &newDims) { + auto checkDims = [&](const llvm::SmallVectorImpl<int64_t> &prevDims, + const llvm::SmallVectorImpl<int64_t> &newDims) { if (prevDims == newDims) return ParseSuccess; return p.emitError("tensor literal is invalid; ranks are not consistent " @@ -801,10 +803,10 @@ ParseResult TensorLiteralParser::parseList(llvm::SmallVectorImpl<int> &dims) { }; bool first = true; - llvm::SmallVector<int, 4> newDims; + llvm::SmallVector<int64_t, 4> newDims; unsigned size = 0; auto parseCommaSeparatedList = [&]() { - llvm::SmallVector<int, 4> thisDims; + llvm::SmallVector<int64_t, 4> thisDims; if (parseElementOrList(thisDims)) return ParseFailure; ++size; diff --git a/mlir/lib/StandardOps/StandardOps.cpp b/mlir/lib/StandardOps/StandardOps.cpp index d9476a7ea9c..328e0c1623b 100644 --- a/mlir/lib/StandardOps/StandardOps.cpp +++ b/mlir/lib/StandardOps/StandardOps.cpp @@ -258,14 +258,14 @@ struct SimplifyAllocConst : public RewritePattern { // Ok, we have one or more constant operands. Collect the non-constant ones // and keep track of the resultant memref type to build. - SmallVector<int, 4> newShapeConstants; + SmallVector<int64_t, 4> newShapeConstants; newShapeConstants.reserve(memrefType.getRank()); SmallVector<Value *, 4> newOperands; SmallVector<Value *, 4> droppedOperands; unsigned dynamicDimPos = 0; for (unsigned dim = 0, e = memrefType.getRank(); dim < e; ++dim) { - int dimSize = memrefType.getDimSize(dim); + int64_t dimSize = memrefType.getDimSize(dim); // If this is already static dimension, keep it. if (dimSize != -1) { newShapeConstants.push_back(dimSize); @@ -794,7 +794,7 @@ Attribute DimOp::constantFold(ArrayRef<Attribute> operands, MLIRContext *context) const { // Constant fold dim when the size along the index referred to is a constant. auto opType = getOperand()->getType(); - int indexSize = -1; + int64_t indexSize = -1; if (auto tensorType = opType.dyn_cast<RankedTensorType>()) { indexSize = tensorType.getShape()[getIndex()]; } else if (auto memrefType = opType.dyn_cast<MemRefType>()) { @@ -1268,7 +1268,7 @@ bool MemRefCastOp::verify() const { return emitOpError("requires input and result ranks to match"); for (unsigned i = 0, e = opType.getRank(); i != e; ++i) { - int opDim = opType.getDimSize(i), resultDim = resType.getDimSize(i); + int64_t opDim = opType.getDimSize(i), resultDim = resType.getDimSize(i); if (opDim != -1 && resultDim != -1 && opDim != resultDim) return emitOpError("requires static dimensions to match"); } @@ -1628,7 +1628,7 @@ bool TensorCastOp::verify() const { return emitOpError("requires input and result ranks to match"); for (unsigned i = 0, e = opRType.getRank(); i != e; ++i) { - int opDim = opRType.getDimSize(i), resultDim = resRType.getDimSize(i); + int64_t opDim = opRType.getDimSize(i), resultDim = resRType.getDimSize(i); if (opDim != -1 && resultDim != -1 && opDim != resultDim) return emitOpError("requires static dimensions to match"); } diff --git a/mlir/lib/SuperVectorOps/SuperVectorOps.cpp b/mlir/lib/SuperVectorOps/SuperVectorOps.cpp index 4a106b066d6..3ffff21155b 100644 --- a/mlir/lib/SuperVectorOps/SuperVectorOps.cpp +++ b/mlir/lib/SuperVectorOps/SuperVectorOps.cpp @@ -484,7 +484,7 @@ bool VectorTypeCastOp::verify() const { if (!dstVectorType) return emitOpError( "expects vector as an element of the target memref type"); - if (llvm::any_of(dstMemrefType.getShape(), [](int s) { return s == -1; })) + if (!dstMemrefType.hasStaticShape()) return emitOpError("does not support dynamic shapes"); if (!getOperand()->getType().isa<MemRefType>()) diff --git a/mlir/lib/Target/LLVMIR/ConvertToLLVMIR.cpp b/mlir/lib/Target/LLVMIR/ConvertToLLVMIR.cpp index 9dfa86c9d94..b91139afe8e 100644 --- a/mlir/lib/Target/LLVMIR/ConvertToLLVMIR.cpp +++ b/mlir/lib/Target/LLVMIR/ConvertToLLVMIR.cpp @@ -360,7 +360,7 @@ llvm::Value *ModuleLowerer::emitMemRefAlloc(ConstOpPointer<AllocOp> allocOp) { SmallVector<llvm::Value *, 4> sizes; sizes.reserve(allocOp->getNumOperands()); unsigned i = 0; - for (int s : type.getShape()) { + for (int64_t s : type.getShape()) { llvm::Value *value = (s == -1) ? valueMapping.lookup(allocOp->getOperand(i++)) : getIndexConstant(s); diff --git a/mlir/lib/Transforms/DmaGeneration.cpp b/mlir/lib/Transforms/DmaGeneration.cpp index 8b86056c8a9..e9d66ef74c3 100644 --- a/mlir/lib/Transforms/DmaGeneration.cpp +++ b/mlir/lib/Transforms/DmaGeneration.cpp @@ -114,7 +114,7 @@ struct StrideInfo { /// successively nested. // TODO(bondhugula): make this work with non-identity layout maps. static void getMultiLevelStrides(const MemRefRegion ®ion, - ArrayRef<int> bufferShape, + ArrayRef<int64_t> bufferShape, SmallVectorImpl<StrideInfo> *strideInfos) { if (bufferShape.size() <= 1) return; @@ -122,7 +122,7 @@ static void getMultiLevelStrides(const MemRefRegion ®ion, int64_t numEltPerStride = 1; int64_t stride = 1; for (int d = bufferShape.size() - 1; d >= 1; d--) { - int dimSize = region.memref->getType().cast<MemRefType>().getDimSize(d); + int64_t dimSize = region.memref->getType().cast<MemRefType>().getDimSize(d); stride *= dimSize; numEltPerStride *= bufferShape[d]; // A stride is needed only if the region has a shorter extent than the @@ -169,7 +169,7 @@ bool DmaGeneration::generateDma(const MemRefRegion ®ion, ForInst *forInst, Value *zeroIndex = top.create<ConstantIndexOp>(loc, 0); unsigned rank = memRefType.getRank(); - SmallVector<int, 4> fastBufferShape; + SmallVector<int64_t, 4> fastBufferShape; // Compute the extents of the buffer. std::vector<SmallVector<int64_t, 4>> lbs; diff --git a/mlir/lib/Transforms/LoopFusion.cpp b/mlir/lib/Transforms/LoopFusion.cpp index 94d763fcbd1..24914878656 100644 --- a/mlir/lib/Transforms/LoopFusion.cpp +++ b/mlir/lib/Transforms/LoopFusion.cpp @@ -711,7 +711,7 @@ static Value *createPrivateMemRef(ForInst *forInst, // Compute MemRefRegion for 'srcStoreOpInst' at depth 'dstLoopDepth'. MemRefRegion region; getMemRefRegion(srcStoreOpInst, dstLoopDepth, ®ion); - SmallVector<int, 4> newShape; + SmallVector<int64_t, 4> newShape; std::vector<SmallVector<int64_t, 4>> lbs; SmallVector<int64_t, 8> lbDivisors; lbs.reserve(rank); diff --git a/mlir/lib/Transforms/LowerVectorTransfers.cpp b/mlir/lib/Transforms/LowerVectorTransfers.cpp index ccda1385df4..19208d4c268 100644 --- a/mlir/lib/Transforms/LowerVectorTransfers.cpp +++ b/mlir/lib/Transforms/LowerVectorTransfers.cpp @@ -96,9 +96,9 @@ private: MLFuncGlobalLoweringState *state; MemRefType memrefType; - ArrayRef<int> memrefShape; + ArrayRef<int64_t> memrefShape; VectorType vectorType; - ArrayRef<int> vectorShape; + ArrayRef<int64_t> vectorShape; AffineMap permutationMap; /// Used for staging the transfer in a local scalar buffer. @@ -232,9 +232,9 @@ VectorTransferRewriter<VectorTransferOpTy>::makeVectorTransferAccessInfo() { } emitter .template bindZipRangeConstants<ConstantIndexOp>( - llvm::zip(lbs, SmallVector<int, 8>(ivs.size(), 0))) + llvm::zip(lbs, SmallVector<int64_t, 8>(ivs.size(), 0))) .template bindZipRangeConstants<ConstantIndexOp>( - llvm::zip(steps, SmallVector<int, 8>(ivs.size(), 1))); + llvm::zip(steps, SmallVector<int64_t, 8>(ivs.size(), 1))); return VectorTransferAccessInfo{ivs, makeExprs(lbs), diff --git a/mlir/lib/Transforms/MaterializeVectors.cpp b/mlir/lib/Transforms/MaterializeVectors.cpp index 6085edd8e8e..e82390f8db9 100644 --- a/mlir/lib/Transforms/MaterializeVectors.cpp +++ b/mlir/lib/Transforms/MaterializeVectors.cpp @@ -187,7 +187,7 @@ struct MaterializationState { MaterializationState() : hwVectorSize(clVectorSize.size(), 0) { std::copy(clVectorSize.begin(), clVectorSize.end(), hwVectorSize.begin()); } - SmallVector<int, 8> hwVectorSize; + SmallVector<int64_t, 8> hwVectorSize; VectorType superVectorType; VectorType hwVectorType; SmallVector<unsigned, 8> hwVectorInstance; @@ -458,7 +458,7 @@ static AffineMap projectedPermutationMap(VectorTransferOpTy *transfer, SmallVector<AffineExpr, 4> keep; MLIRContext *context = transfer->getInstruction()->getContext(); functional::zipApply( - [&dim, &keep, context](int shape, int ratio) { + [&dim, &keep, context](int64_t shape, int64_t ratio) { assert(shape >= ratio && "shape dim must be greater than ratio dim"); if (shape != ratio) { // HW vector is not full instantiated along this dim, keep it. diff --git a/mlir/lib/Transforms/PipelineDataTransfer.cpp b/mlir/lib/Transforms/PipelineDataTransfer.cpp index 989af0071d7..9e7c928070f 100644 --- a/mlir/lib/Transforms/PipelineDataTransfer.cpp +++ b/mlir/lib/Transforms/PipelineDataTransfer.cpp @@ -87,8 +87,8 @@ static bool doubleBuffer(Value *oldMemRef, ForInst *forInst) { // Doubles the shape with a leading dimension extent of 2. auto doubleShape = [&](MemRefType oldMemRefType) -> MemRefType { // Add the leading dimension in the shape for the double buffer. - ArrayRef<int> oldShape = oldMemRefType.getShape(); - SmallVector<int, 4> newShape(1 + oldMemRefType.getRank()); + ArrayRef<int64_t> oldShape = oldMemRefType.getShape(); + SmallVector<int64_t, 4> newShape(1 + oldMemRefType.getRank()); newShape[0] = 2; std::copy(oldShape.begin(), oldShape.end(), newShape.begin() + 1); auto newMemRefType = diff --git a/mlir/lib/Transforms/Vectorization/VectorizerTestPass.cpp b/mlir/lib/Transforms/Vectorization/VectorizerTestPass.cpp index 7e5cac0d87c..ad966e8d280 100644 --- a/mlir/lib/Transforms/Vectorization/VectorizerTestPass.cpp +++ b/mlir/lib/Transforms/Vectorization/VectorizerTestPass.cpp @@ -101,8 +101,8 @@ char VectorizerTestPass::passID = 0; void VectorizerTestPass::testVectorShapeRatio(Function *f) { using matcher::Op; - SmallVector<int, 8> shape(clTestVectorShapeRatio.begin(), - clTestVectorShapeRatio.end()); + SmallVector<int64_t, 8> shape(clTestVectorShapeRatio.begin(), + clTestVectorShapeRatio.end()); auto subVectorType = VectorType::get(shape, Type::getF32(f->getContext())); // Only filter instructions that operate on a strict super-vector and have one // return. This makes testing easier. diff --git a/mlir/lib/Transforms/Vectorize.cpp b/mlir/lib/Transforms/Vectorize.cpp index 58bb3901947..8a6d965ce0d 100644 --- a/mlir/lib/Transforms/Vectorize.cpp +++ b/mlir/lib/Transforms/Vectorize.cpp @@ -667,7 +667,7 @@ char Vectorize::passID = 0; namespace { struct VectorizationStrategy { - ArrayRef<int> vectorSizes; + SmallVector<int64_t, 8> vectorSizes; DenseMap<ForInst *, unsigned> loopToVectorDim; }; @@ -1280,7 +1280,8 @@ PassResult Vectorize::runOnFunction(Function *f) { for (auto m : matches) { VectorizationStrategy strategy; // TODO(ntv): depending on profitability, elect to reduce the vector size. - strategy.vectorSizes = clVirtualVectorSize; + strategy.vectorSizes.assign(clVirtualVectorSize.begin(), + clVirtualVectorSize.end()); auto fail = analyzeProfitability(m.second, 1, patternDepth, &strategy); if (fail) { continue; |
