| 1 | //===- GPUBase.h - GPU types, attributes, and enums -------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file declares GPU types, attributes, and enums independently of GPU |
| 10 | // operations. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #ifndef MLIR_DIALECT_GPU_IR_GPUBASE_H |
| 15 | #define MLIR_DIALECT_GPU_IR_GPUBASE_H |
| 16 | |
| 17 | // These representations can be inspected without linking GPU operations. |
| 18 | // Constructing or parsing GPU types and attributes still requires loading |
| 19 | // GPUDialect, which owns and registers them as before. |
| 20 | // Attribute methods that build IR are provided by MLIRGPUDialect. |
| 21 | #include "mlir/Dialect/GPU/IR/CompilationInterfaces.h" |
| 22 | #include "mlir/Dialect/GPU/IR/GPUDialectDecl.h" |
| 23 | #include "mlir/Dialect/SCF/IR/DeviceMappingInterface.h" |
| 24 | #include "mlir/IR/BuiltinTypes.h" |
| 25 | #include "mlir/Interfaces/FunctionInterfaces.h" |
| 26 | |
| 27 | namespace mlir::gpu { |
| 28 | namespace detail { |
| 29 | /// Register storage whose definitions are private to MLIRGPUBase. Attribute |
| 30 | /// metadata and interfaces are registered by GPUDialect::initialize(). |
| 31 | void registerAttributeStorage(MLIRContext *context); |
| 32 | } // namespace detail |
| 33 | |
| 34 | class AsyncTokenType |
| 35 | : public Type::TypeBase<AsyncTokenType, Type, TypeStorage> { |
| 36 | public: |
| 37 | // Used for generic hooks in TypeBase. |
| 38 | using Base::Base; |
| 39 | |
| 40 | static constexpr StringLiteral name = "gpu.async_token" ; |
| 41 | }; |
| 42 | |
| 43 | class NamedBarrierType |
| 44 | : public Type::TypeBase<NamedBarrierType, Type, TypeStorage> { |
| 45 | public: |
| 46 | using Base::Base; |
| 47 | |
| 48 | static constexpr StringLiteral name = "gpu.named_barrier" ; |
| 49 | }; |
| 50 | |
| 51 | /// MMAMatrixType storage and uniquing. Array is uniqued based on its shape |
| 52 | /// and type. |
| 53 | struct MMAMatrixStorageType : public TypeStorage { |
| 54 | MMAMatrixStorageType(unsigned numDims, const int64_t *dimShapes, |
| 55 | Type elementType, StringRef operand) |
| 56 | : dimShapes(dimShapes), numDims(numDims), elementType(elementType), |
| 57 | operand(operand) {} |
| 58 | |
| 59 | /// The hash key for uniquing. |
| 60 | using KeyTy = std::tuple<ArrayRef<int64_t>, Type, StringRef>; |
| 61 | bool operator==(const KeyTy &key) const { |
| 62 | return key == KeyTy(getShape(), elementType, operand); |
| 63 | } |
| 64 | |
| 65 | /// Construction. |
| 66 | static MMAMatrixStorageType *construct(TypeStorageAllocator &allocator, |
| 67 | const KeyTy &key) { |
| 68 | ArrayRef<int64_t> shape = allocator.copyInto(elements: std::get<0>(t: key)); |
| 69 | StringRef operand = allocator.copyInto(str: std::get<2>(t: key)); |
| 70 | |
| 71 | return new (allocator.allocate<MMAMatrixStorageType>()) |
| 72 | MMAMatrixStorageType(shape.size(), shape.data(), std::get<1>(t: key), |
| 73 | operand); |
| 74 | } |
| 75 | |
| 76 | ArrayRef<int64_t> getShape() const { |
| 77 | return ArrayRef<int64_t>(dimShapes, numDims); |
| 78 | } |
| 79 | |
| 80 | StringRef getOperand() const { return operand; } |
| 81 | |
| 82 | /// Reference to the shape of the MMA matrix. |
| 83 | const int64_t *dimShapes; |
| 84 | |
| 85 | /// Number of dimensions in the MMA matrix. |
| 86 | unsigned numDims; |
| 87 | |
| 88 | /// Element type of elements held in the MMA matrix. |
| 89 | Type elementType; |
| 90 | |
| 91 | /// MMA operand that this MMAMatrix holds. The general form of operation this |
| 92 | /// type supports is given by the equation C += A*B. This field specifies |
| 93 | /// which operand in the given equation is held by this type. The valid values |
| 94 | /// are "AOp", "BOp" and "COp". |
| 95 | StringRef operand; |
| 96 | }; |
| 97 | |
| 98 | /// MMAMatrix represents a matrix held by a subgroup for matrix-matrix multiply |
| 99 | /// accumulate operations. MMAMatrices are taken as direct operands by these |
| 100 | /// operations and are also produced as results. These matrices are meant to |
| 101 | /// reside in the registers. A limited number of pointwise operations can be |
| 102 | /// performed on these matrices, i.e., operations which operate uniformly on |
| 103 | /// all the elements in the matrix and do not change the order of matrix |
| 104 | /// elements. The above conditions exist because the layout of matrix elements |
| 105 | /// inside the matrix is opaque i.e., the elements may be present in the |
| 106 | /// matrix in any order. The general usage of this type is shown as follows:- |
| 107 | /// |
| 108 | /// %0 = gpu.subgroup_mma_load_matrix %arg0[%c0, %c0] {leadDimension = 16 : |
| 109 | /// index} : memref<16x16xf16> -> !gpu.mma_matrix<16x16xf16, "AOp"> |
| 110 | /// |
| 111 | /// The MMAMatrixType describes the shape of the matrix being loaded and the |
| 112 | /// operand being loaded too. The operand needs to be specified to aid the |
| 113 | /// lowering of this type to dialects such as NVVM where each workitem may |
| 114 | /// hold different amount of elements depending on the elementType of the |
| 115 | /// matrix. For e.g., Each workitem holds 4 vector<2xf16>s for f16 data type |
| 116 | /// and 8 f32s for f32 data type of MMAMatrix. Some other instances of usage |
| 117 | /// are:- |
| 118 | /// |
| 119 | /// %3 = gpu.subgroup_mma_compute %0, %1, %2 : |
| 120 | /// !gpu.mma_matrix<16x16xf16, "AOp">, !gpu.mma_matrix<16x16xf16, "BOp"> |
| 121 | /// -> !gpu.mma_matrix<16x16xf32, "COp"> |
| 122 | /// |
| 123 | /// |
| 124 | /// gpu.subgroup_mma_store_matrix %3, %arg22[%c0, %c0] {leadDimension = 16 |
| 125 | /// : index}: !gpu.mma_matrix<16x16xf32, "COp">, memref<16x16xf32> |
| 126 | /// |
| 127 | /// MMA matrices may be memref elements on targets that support storing them, |
| 128 | /// for example SPIRV can represent arrays of MMA matrices. |
| 129 | // TODO: consider moving this to ODS. |
| 130 | class MMAMatrixType |
| 131 | : public Type::TypeBase<MMAMatrixType, Type, MMAMatrixStorageType, |
| 132 | MemRefElementTypeInterface::Trait> { |
| 133 | public: |
| 134 | using Base::Base; |
| 135 | |
| 136 | static constexpr StringLiteral name = "gpu.mma_matrix" ; |
| 137 | |
| 138 | /// Get MMAMatrixType and verify construction Invariants. |
| 139 | static MMAMatrixType get(ArrayRef<int64_t> shape, Type elementType, |
| 140 | StringRef operand); |
| 141 | |
| 142 | /// Get MMAMatrixType at a particular location and verify construction |
| 143 | /// Invariants. |
| 144 | static MMAMatrixType getChecked(function_ref<InFlightDiagnostic()> emitError, |
| 145 | ArrayRef<int64_t> shape, Type elementType, |
| 146 | StringRef operand); |
| 147 | |
| 148 | /// Check if a type is valid a MMAMatrixType elementType. |
| 149 | static bool isValidElementType(Type elementType); |
| 150 | |
| 151 | /// Verify that shape and elementType are actually allowed for the |
| 152 | /// MMAMatrixType. |
| 153 | static LogicalResult |
| 154 | verifyInvariants(function_ref<InFlightDiagnostic()> emitError, |
| 155 | ArrayRef<int64_t> shape, Type elementType, |
| 156 | StringRef operand); |
| 157 | |
| 158 | /// Get number of dims. |
| 159 | unsigned getNumDims() const; |
| 160 | |
| 161 | /// Get shape of the matrix. |
| 162 | ArrayRef<int64_t> getShape() const; |
| 163 | |
| 164 | /// Get elementType of a single element. |
| 165 | Type getElementType() const; |
| 166 | |
| 167 | /// The general form of operation this type supports is given by the equation |
| 168 | /// C += A*B. This function returns which operand in the given equation is |
| 169 | /// held by this type. String returned can be one of"AOp", "BOp" and "COp". |
| 170 | StringRef getOperand() const; |
| 171 | }; |
| 172 | |
| 173 | // Handle types for sparse. |
| 174 | enum class SparseHandleKind { SpMat, DnTensor, SpGEMMOp }; |
| 175 | |
| 176 | class SparseDnTensorHandleType |
| 177 | : public Type::TypeBase<SparseDnTensorHandleType, Type, TypeStorage> { |
| 178 | public: |
| 179 | using Base = |
| 180 | Type::TypeBase<SparseDnTensorHandleType, Type, TypeStorage>::Base; |
| 181 | using Base::Base; |
| 182 | |
| 183 | static constexpr StringLiteral name = "gpu.sparse.dntensor_handle" ; |
| 184 | }; |
| 185 | |
| 186 | class SparseSpMatHandleType |
| 187 | : public Type::TypeBase<SparseSpMatHandleType, Type, TypeStorage> { |
| 188 | public: |
| 189 | using Base = Type::TypeBase<SparseSpMatHandleType, Type, TypeStorage>::Base; |
| 190 | using Base::Base; |
| 191 | |
| 192 | static constexpr StringLiteral name = "gpu.sparse.spmat_handle" ; |
| 193 | }; |
| 194 | |
| 195 | class SparseSpGEMMOpHandleType |
| 196 | : public Type::TypeBase<SparseSpGEMMOpHandleType, Type, TypeStorage> { |
| 197 | public: |
| 198 | using Base = |
| 199 | Type::TypeBase<SparseSpGEMMOpHandleType, Type, TypeStorage>::Base; |
| 200 | using Base::Base; |
| 201 | |
| 202 | static constexpr StringLiteral name = "gpu.sparse.spgemmop_handle" ; |
| 203 | }; |
| 204 | } // namespace mlir::gpu |
| 205 | |
| 206 | #define GET_ATTRDEF_CLASSES |
| 207 | #include "mlir/Dialect/GPU/IR/GPUOpsAttributes.h.inc" |
| 208 | |
| 209 | #endif // MLIR_DIALECT_GPU_IR_GPUBASE_H |
| 210 | |