SiCong Li | b63b119 | 2022-01-28 18:24:39 +0000 | [diff] [blame^] | 1 | /* |
| 2 | * Copyright (c) 2022 Arm Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #ifndef ENABLE_EXPERIMENTAL_DYNAMIC_FUSION |
| 25 | #error "This experimental feature must be enabled with -DENABLE_EXPERIMENTAL_DYNAMIC_FUSION" |
| 26 | #endif /* ENABLE_EXPERIMENTAL_DYNAMIC_FUSION */ |
| 27 | #ifndef ARM_COMPUTE_EXPERIMENTAL_DYNAMIC_FUSION_CLCOMPOSITEOPERATOR_H |
| 28 | #define ARM_COMPUTE_EXPERIMENTAL_DYNAMIC_FUSION_CLCOMPOSITEOPERATOR_H |
| 29 | |
| 30 | #include "arm_compute/core/CL/CLCompileContext.h" |
| 31 | #include "arm_compute/runtime/CL/CLTensor.h" |
| 32 | #include "arm_compute/runtime/IOperator.h" |
| 33 | |
| 34 | #include "arm_compute/core/experimental/ClWorkload.h" |
| 35 | |
| 36 | #include <memory> |
| 37 | |
| 38 | namespace arm_compute |
| 39 | { |
| 40 | namespace experimental |
| 41 | { |
| 42 | namespace dynamic_fusion |
| 43 | { |
| 44 | /** Map OpTensor handles to their corresponding ITensor memory |
| 45 | */ |
| 46 | using OpTensorBinding = std::map<OpTensor, ITensor *>; |
| 47 | |
| 48 | /** Map a kernel (as identified by its unit workload id) to its corresponding tensor pack |
| 49 | * |
| 50 | * @note External user should not use the add_tensor_pack method to alter this tensor pack map, and should only use the map returned by @ref bind_tensors |
| 51 | */ |
| 52 | class TensorPackMap |
| 53 | { |
| 54 | public: |
| 55 | /** Find a tensor pack associated with the unit workload Id @p uwk_id |
| 56 | * |
| 57 | * @param[in] uwk_id unit workload Id associated with the tensor pack |
| 58 | * |
| 59 | * @return ITensorPack* |
| 60 | */ |
| 61 | ITensorPack *find_tensor_pack(UnitWorkload::Id uwk_id); |
| 62 | /** Get a tensor pack associated with @p uwk_id. Throws a exception if it cannot be found. |
| 63 | * |
| 64 | * @param[in] uwk_id unit workload Id associated with the tensor pack |
| 65 | * |
| 66 | * @return ITensorPack* |
| 67 | */ |
| 68 | ITensorPack &get_tensor_pack(UnitWorkload::Id uwk_id); |
| 69 | /** Add a tensor pack and associate it with unit workload Id @p uwk_id |
| 70 | * @note Should not be used by external user |
| 71 | * |
| 72 | * @param[in] uwk_id unit workload Id associated with the tensor pack |
| 73 | * @param[in] tensor_pack Tensor Pack to be added |
| 74 | */ |
| 75 | void add_tensor_pack(UnitWorkload::Id uwk_id, const ITensorPack &tensor_pack); |
| 76 | |
| 77 | private: |
| 78 | std::map<UnitWorkload::Id, ITensorPack> _tensor_packs{}; |
| 79 | }; |
| 80 | |
| 81 | /** Holder of any auxiliary CLTensors required by a ClWorkload. |
| 82 | * |
| 83 | * @note The tensors are not allocated by default, and require the user to explicitly allocate them using the TensorInfo and AuxMemoryInfo |
| 84 | * |
| 85 | * @note This data holder must remain valid until the ClCompositeOperator that it's passed to is out of scope |
| 86 | * |
| 87 | * @note External user should not use the add_aux_tensor method, and should only use the data returned by @ref bind_tensors |
| 88 | */ |
| 89 | class ClAuxTensorData |
| 90 | { |
| 91 | public: |
| 92 | /** A view of a single auxiliary data and the associated TensorInfo and AuxMemoryInfo |
| 93 | */ |
| 94 | struct DataView |
| 95 | { |
| 96 | DataView() = default; |
| 97 | DataView(CLTensor *tensor, const TensorInfo &tensor_info, const AuxMemoryInfo &memory_info) |
| 98 | : tensor{ tensor }, tensor_info{ tensor_info }, memory_info{ memory_info } |
| 99 | { |
| 100 | } |
| 101 | ~DataView() = default; |
| 102 | DataView(const DataView &other) = default; |
| 103 | DataView &operator=(const DataView &other) = default; |
| 104 | DataView(DataView &&other) = default; |
| 105 | DataView &operator=(DataView &&other) = default; |
| 106 | CLTensor *tensor{}; /**< Pointer to the auxiliary tensor */ |
| 107 | TensorInfo tensor_info{}; /**< Associated TensorInfo */ |
| 108 | AuxMemoryInfo memory_info{}; /**< Memory requirement */ |
| 109 | }; |
| 110 | |
| 111 | /** Add auxiliary tensor. |
| 112 | * |
| 113 | * @note Should not be used by external user |
| 114 | * |
| 115 | * @param[in] tensor_id Any Id that can uniquely identify an auxiliary tensor. Usually ClWorkloadTensor Id |
| 116 | * @param[in] tensor_info TensorInfo associated with the tensor |
| 117 | * @param[in] memory_info Memory requirements |
| 118 | * |
| 119 | * @return CLTensor* if successfully added, otherwise nullptr |
| 120 | */ |
| 121 | CLTensor *add_aux_tensor(int tensor_id, const ITensorInfo &tensor_info, const AuxMemoryInfo &memory_info); |
| 122 | |
| 123 | /** Get views of all auxiliary tensors. This is mainly used for allocating the auxiliary tensors. |
| 124 | * |
| 125 | * @return std::vector<DataView>& |
| 126 | */ |
| 127 | std::vector<DataView> &get_tensors(); |
| 128 | |
| 129 | private: |
| 130 | std::map<int, std::unique_ptr<CLTensor>> _owned_tensors{}; |
| 131 | std::vector<DataView> _tensors{}; |
| 132 | }; |
| 133 | |
| 134 | /** Bind tensor memory to packs used by prepare and run methods. Create auxiliary tensor objects and their memory requirements if needed |
| 135 | * |
| 136 | * @note This is the only method for external user to create ClAuxTensorData, and the prepare and run TensorPackMaps |
| 137 | * |
| 138 | * @param[out] aux_tensor_data Auxiliary Tensors required by the workload |
| 139 | * @param[out] prepare_pack_map TensorPackMap used by the prepare method |
| 140 | * @param[out] run_pack_map TensorPackMap used by the run method |
| 141 | * @param[in] workload ClWorkload to bind the tensors to |
| 142 | * @param[in] op_tensors CLTensor memory objects mapped from Core OpTensors |
| 143 | * |
| 144 | * @return Status |
| 145 | */ |
| 146 | Status bind_tensors(ClAuxTensorData &aux_tensor_data, TensorPackMap &prepare_pack_map, TensorPackMap &run_pack_map, const ClWorkload &workload, const OpTensorBinding &op_tensors); |
| 147 | |
| 148 | /** Operator runtime to run a @ref ClWorkload |
| 149 | * |
| 150 | * @note User must explicitly call prepare before run otherwise run will fail. |
| 151 | * |
| 152 | */ |
| 153 | class ClCompositeOperator |
| 154 | { |
| 155 | public: |
| 156 | ClCompositeOperator(); |
| 157 | ~ClCompositeOperator(); |
| 158 | /** Configures a @ref ClCompositeOperator with a @ref ClWorkload |
| 159 | * This includes the compilation of Cl kernels inside the @ref ClWorkload |
| 160 | * |
| 161 | * @param[in] ctx CLCompileContext |
| 162 | * @param[in] workload ClWorkload to configure with |
| 163 | */ |
| 164 | void configure(const CLCompileContext &ctx, const ClWorkload &workload); |
| 165 | /** Validate ClWorkload @p workload |
| 166 | * |
| 167 | * @param[in] workload ClWorkload to be validated |
| 168 | * |
| 169 | * @return Status |
| 170 | */ |
| 171 | static Status validate(const ClWorkload &workload); |
| 172 | /** Enqueue prepare workloads |
| 173 | * |
| 174 | * @param tensor_pack_map Tensors required by the prepare workloads |
| 175 | */ |
| 176 | void prepare(TensorPackMap &tensor_pack_map); |
| 177 | /** Enqueue run workloads |
| 178 | * |
| 179 | * @param tensor_pack_map Tensors required by the run workloads |
| 180 | */ |
| 181 | void run(TensorPackMap &tensor_pack_map); |
| 182 | |
| 183 | private: |
| 184 | struct Implementation; |
| 185 | std::unique_ptr<Implementation> _impl; |
| 186 | }; |
| 187 | |
| 188 | } // namespace dynamic_fusion |
| 189 | } // namespace experimental |
| 190 | } // namespace arm_compute |
| 191 | #endif //ARM_COMPUTE_EXPERIMENTAL_DYNAMIC_FUSION_CLCOMPOSITEOPERATOR_H |