1 /* 2 * Copyright (c) 2017-2021 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 ARM_COMPUTE_CL_FULLY_CONNECTED_H 25 #define ARM_COMPUTE_CL_FULLY_CONNECTED_H 26 27 #include "arm_compute/core/TensorInfo.h" 28 29 #include "src/gpu/cl/ClCompileContext.h" 30 #include "src/gpu/cl/IClOperator.h" 31 32 #include <memory> 33 34 namespace arm_compute 35 { 36 namespace opencl 37 { 38 // Forward declarations 39 class ClConvertFullyConnectedWeights; 40 class ClFlatten; 41 class ClGemm; 42 class ClGemmLowpMatrixMultiplyCore; 43 class ClTranspose; 44 45 /** Basic function to compute a Fully Connected layer on OpenCL. This function calls the following OpenCL kernels: 46 * 47 * -# @ref opencl::kernels::ClIm2ColKernel (called when the input comes from a convolutional layer) 48 * -# @ref CLTranspose (if @p are_weights_reshaped is set to false and transpose_weights is set to true ) (called once) 49 * -# @ref opencl::ClGemm or @ref CLGEMMLowpMatrixMultiplyCore (if quantized asymmetric) 50 * 51 * @note The fully connected layer accepts "weights" tensors only with 2 dimensions. 52 */ 53 class ClFullyConnected : public IClOperator 54 { 55 public: 56 ClFullyConnected(); 57 ~ClFullyConnected(); 58 /** Set the input and output tensors. 59 * 60 * Valid data layouts: 61 * - NHWC 62 * - NCHW 63 * 64 * Valid data type configurations: 65 * |src0 |src1 |src2 |dst | 66 * |:--------------|:------------------|:------|:--------------| 67 * |F16 |F16 |F16 |F16 | 68 * |F32 |F32 |F32 |F32 | 69 * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | 70 * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | 71 * 72 * @param[in] compile_context The compile context to be used. 73 * @param[in] src Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/F16/F32. 74 * @param[in] weights Weights tensor. The weights must be 2 dimensional. 75 * If this function is called after a Convolution Layer, the (transposed) weights will have as many rows as the product of the first 3 input's dimensions. 76 * If it is called after another FullyConnected Layer, the (transposed) weights will have as many rows as the input's first dimension. 77 * Data type supported: Same as @p src. 78 * @param[in] biases Bias tensor. Can be nullptr. Data type supported:Same as @p src. 79 * @param[out] dst Destination tensor. Its shape should be equal to the output of a matrix multiplication between: 80 * - The output of im2col on the input and the (transposed) 2D weights, if the function is called after a Convolution Layer 81 * - The input tensor and the (transposed) 2D weights, if the function is called after another FullyConnected Layer. 82 * Data type supported: Same as @p src. 83 * @param[in] fc_info (Optional) Fully connected layer additional info 84 */ 85 void configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, 86 FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo()); 87 /** Static function to check if given info will lead to a valid configuration 88 * 89 * Similar to ClFullyConnected::configure() 90 * 91 * @return a status 92 */ 93 static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, 94 FullyConnectedLayerInfo fc_info = FullyConnectedLayerInfo()); 95 96 // Inherited methods overriden 97 void run(ITensorPack &tensors) override; 98 void prepare(ITensorPack &tensors) override; 99 experimental::MemoryRequirements workspace() const override; 100 101 private: 102 void configure_fc_fc(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst, const FullyConnectedLayerInfo &fc_info); 103 void configure_conv_fc(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst, const FullyConnectedLayerInfo &fc_info); 104 void configure_mm(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *bias, ITensorInfo *dst, const FullyConnectedLayerInfo &fc_info); 105 106 private: 107 enum AuxTensorIdx 108 { 109 TransposedWeights = 10, 110 ConvertedWeights = 11, 111 FlattenedSrc = 12, 112 Count = 13 113 }; 114 115 std::unique_ptr<ClConvertFullyConnectedWeights> _convert_weights; 116 std::unique_ptr<ClFlatten> _flatten; 117 std::unique_ptr<ClTranspose> _reshape_weights; 118 std::unique_ptr<ClGemm> _mm_gemm; 119 std::unique_ptr<ClGemmLowpMatrixMultiplyCore> _mm_gemmlowp; 120 121 experimental::MemoryRequirements _aux_mem{}; 122 123 TensorInfo _flattened_src{}; 124 TensorInfo _converted_weights{}; 125 TensorInfo _reshaped_weights{}; 126 127 TensorInfo _weights_to_use{}; 128 int _weights_to_use_idx{ ACL_SRC_1 }; 129 130 bool _are_weights_converted{ true }; 131 bool _are_weights_reshaped{ true }; 132 bool _is_fc_after_conv{ true }; 133 bool _is_quantized{ false }; 134 bool _is_prepared{ false }; 135 }; 136 } // namespace opencl 137 } // namespace arm_compute 138 #endif /* ARM_COMPUTE_CL_FULLY_CONNECTED_H */ 139