| 1 | //===- OpenACCParMapping.h - OpenACC Parallelism Mapping -------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the MLIR 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 defines the interface for mapping OpenACC parallelism levels |
| 10 | // (gang, worker, vector) to target-specific parallel dimension attributes. |
| 11 | // |
| 12 | // Users can provide custom implementations of ACCParMappingPolicy to |
| 13 | // support different mapping strategies and target attributes. |
| 14 | // |
| 15 | //===----------------------------------------------------------------------===// |
| 16 | |
| 17 | #ifndef MLIR_DIALECT_OPENACC_OPENACCPARMAPPING_H_ |
| 18 | #define MLIR_DIALECT_OPENACC_OPENACCPARMAPPING_H_ |
| 19 | |
| 20 | #include "mlir/Dialect/OpenACC/OpenACC.h" |
| 21 | #include "mlir/IR/MLIRContext.h" |
| 22 | #include "llvm/Support/ErrorHandling.h" |
| 23 | |
| 24 | namespace mlir { |
| 25 | namespace acc { |
| 26 | |
| 27 | //===----------------------------------------------------------------------===// |
| 28 | // Utility functions |
| 29 | //===----------------------------------------------------------------------===// |
| 30 | |
| 31 | /// Convert a gang dimension value (1, 2, or 3) to the corresponding ParLevel. |
| 32 | /// Asserts if the value is not a valid gang dimension. |
| 33 | inline ParLevel getGangParLevel(int64_t gangDimValue) { |
| 34 | assert((gangDimValue >= 1 && gangDimValue <= 3) && |
| 35 | "gang dimension must be 1, 2, or 3" ); |
| 36 | switch (gangDimValue) { |
| 37 | case 1: |
| 38 | return ParLevel::gang_dim1; |
| 39 | case 2: |
| 40 | return ParLevel::gang_dim2; |
| 41 | case 3: |
| 42 | return ParLevel::gang_dim3; |
| 43 | } |
| 44 | llvm_unreachable("validated gang dimension" ); |
| 45 | } |
| 46 | |
| 47 | //===----------------------------------------------------------------------===// |
| 48 | // ACCParMappingPolicy |
| 49 | //===----------------------------------------------------------------------===// |
| 50 | |
| 51 | /// Policy class that defines how OpenACC parallelism levels map to |
| 52 | /// target-specific parallel dimension attributes. Implementations provide the |
| 53 | /// actual mapping. |
| 54 | /// |
| 55 | /// Template parameter ParDimAttrT specifies the attribute type returned by |
| 56 | /// the mapping functions (e.g., mlir::acc::GPUParallelDimAttr for GPU targets). |
| 57 | /// |
| 58 | /// This policy allows different mapping strategies: |
| 59 | /// - Standard GPU mapping (gang->block, worker->threadY, vector->threadX) |
| 60 | /// - Custom mappings for specific targets or optimization strategies |
| 61 | /// |
| 62 | /// Pass an implementation to functions that need to perform the mapping. |
| 63 | template <typename ParDimAttrT> |
| 64 | class ACCParMappingPolicy { |
| 65 | public: |
| 66 | virtual ~ACCParMappingPolicy() = default; |
| 67 | |
| 68 | /// Map an OpenACC parallelism level to target dimension. |
| 69 | /// @param ctx The MLIR context |
| 70 | /// @param level The OpenACC parallelism level (gang_dim1, gang_dim2, |
| 71 | /// gang_dim3, worker, vector, or seq) |
| 72 | /// @return The corresponding parallel dimension attribute |
| 73 | virtual ParDimAttrT map(MLIRContext *ctx, ParLevel level) const = 0; |
| 74 | |
| 75 | /// Convenience methods for specific parallelism levels. |
| 76 | ParDimAttrT gangDim(MLIRContext *ctx, ParLevel level) const { |
| 77 | assert((level == ParLevel::gang_dim1 || level == ParLevel::gang_dim2 || |
| 78 | level == ParLevel::gang_dim3) && |
| 79 | "gangDim requires a gang parallelism level" ); |
| 80 | return map(ctx, level); |
| 81 | } |
| 82 | ParDimAttrT workerDim(MLIRContext *ctx) const { |
| 83 | return map(ctx, level: ParLevel::worker); |
| 84 | } |
| 85 | ParDimAttrT vectorDim(MLIRContext *ctx) const { |
| 86 | return map(ctx, level: ParLevel::vector); |
| 87 | } |
| 88 | ParDimAttrT seqDim(MLIRContext *ctx) const { return map(ctx, level: ParLevel::seq); } |
| 89 | |
| 90 | //===--------------------------------------------------------------------===// |
| 91 | // Predicate methods - check if an attribute matches a parallelism level |
| 92 | //===--------------------------------------------------------------------===// |
| 93 | |
| 94 | /// Check if the attribute represents vector parallelism. |
| 95 | virtual bool isVector(ParDimAttrT attr) const = 0; |
| 96 | |
| 97 | /// Check if the attribute represents worker parallelism. |
| 98 | virtual bool isWorker(ParDimAttrT attr) const = 0; |
| 99 | |
| 100 | /// Check if the attribute represents gang parallelism (any gang dimension). |
| 101 | virtual bool isGang(ParDimAttrT attr) const = 0; |
| 102 | |
| 103 | /// Check if the attribute represents sequential execution. |
| 104 | virtual bool isSeq(ParDimAttrT attr) const = 0; |
| 105 | }; |
| 106 | |
| 107 | //===----------------------------------------------------------------------===// |
| 108 | // DefaultACCToGPUMappingPolicy |
| 109 | //===----------------------------------------------------------------------===// |
| 110 | |
| 111 | /// Default policy that provides the standard GPU mapping: |
| 112 | /// gang(dim:1) -> BlockX (gridDim.x / blockIdx.x) |
| 113 | /// gang(dim:2) -> BlockY (gridDim.y / blockIdx.y) |
| 114 | /// gang(dim:3) -> BlockZ (gridDim.z / blockIdx.z) |
| 115 | /// worker -> ThreadY (blockDim.y / threadIdx.y) |
| 116 | /// vector -> ThreadX (blockDim.x / threadIdx.x) |
| 117 | /// seq -> Sequential |
| 118 | class DefaultACCToGPUMappingPolicy |
| 119 | : public ACCParMappingPolicy<mlir::acc::GPUParallelDimAttr> { |
| 120 | public: |
| 121 | mlir::acc::GPUParallelDimAttr map(MLIRContext *ctx, |
| 122 | ParLevel level) const override { |
| 123 | switch (level) { |
| 124 | case ParLevel::gang_dim1: |
| 125 | return mlir::acc::GPUParallelDimAttr::blockXDim(ctx); |
| 126 | case ParLevel::gang_dim2: |
| 127 | return mlir::acc::GPUParallelDimAttr::blockYDim(ctx); |
| 128 | case ParLevel::gang_dim3: |
| 129 | return mlir::acc::GPUParallelDimAttr::blockZDim(ctx); |
| 130 | case ParLevel::worker: |
| 131 | return mlir::acc::GPUParallelDimAttr::threadYDim(ctx); |
| 132 | case ParLevel::vector: |
| 133 | return mlir::acc::GPUParallelDimAttr::threadXDim(ctx); |
| 134 | case ParLevel::seq: |
| 135 | return mlir::acc::GPUParallelDimAttr::seqDim(ctx); |
| 136 | } |
| 137 | llvm_unreachable("Unknown ParLevel" ); |
| 138 | } |
| 139 | |
| 140 | bool isVector(mlir::acc::GPUParallelDimAttr attr) const override { |
| 141 | return attr.isThreadX(); |
| 142 | } |
| 143 | |
| 144 | bool isWorker(mlir::acc::GPUParallelDimAttr attr) const override { |
| 145 | return attr.isThreadY(); |
| 146 | } |
| 147 | |
| 148 | bool isGang(mlir::acc::GPUParallelDimAttr attr) const override { |
| 149 | return attr.isAnyBlock(); |
| 150 | } |
| 151 | |
| 152 | bool isSeq(mlir::acc::GPUParallelDimAttr attr) const override { |
| 153 | return attr.isSeq(); |
| 154 | } |
| 155 | }; |
| 156 | |
| 157 | /// Type alias for the GPU-specific mapping policy |
| 158 | using ACCToGPUMappingPolicy = |
| 159 | ACCParMappingPolicy<mlir::acc::GPUParallelDimAttr>; |
| 160 | |
| 161 | } // namespace acc |
| 162 | } // namespace mlir |
| 163 | |
| 164 | #endif // MLIR_DIALECT_OPENACC_OPENACCPARMAPPING_H_ |
| 165 | |