1 /* 2 * Copyright (c) 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_CLCONV2D_H 25 #define ARM_COMPUTE_CLCONV2D_H 26 27 #include "arm_compute/core/Types.h" 28 #include "arm_compute/runtime/FunctionDescriptors.h" 29 #include "src/gpu/cl/ClCompileContext.h" 30 #include "src/gpu/cl/IClKernel.h" 31 #include "src/gpu/cl/IClOperator.h" 32 33 namespace arm_compute 34 { 35 namespace opencl 36 { 37 /** Basic function to compute the convolution layer. This function calls the following OpenCL kernels/functions: 38 * 39 * -# @ref opencl::ClGemmConv2d 40 * -# @ref opencl::ClWinogradConv2d 41 * -# @ref opencl::ClIndirectConv2d 42 * -# @ref opencl::ClDirectConv2d 43 * -# @ref CLFFTConvolutionLayer 44 * 45 * The function selects one of the algorithms mentioned above based on: 46 * - The size of the kernel 47 * - Number of src/dst feature maps 48 * - Amount of memory needed 49 * 50 * Generally GEMM-based convolution is executed when neither Winograd nor FFT nor Direct convolution can be performed. 51 * 52 * FP32 Algorithm| Filter Size | Input/Output feature maps | 53 * --------------|-------------------------------------------------------------|-------------------------------------------| 54 * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5(fast maths) 7x1 1x7 | Input channels is greater than 3 | 55 * FFT | Squared kernels and greater than 9x9 | Input feature maps > Output feature maps | 56 * DirectConv | 9x9 | | 57 * GEMM | Any size | | 58 * 59 * Winograd 5x5 requires fast maths enabled. 60 * 61 * FP16 Algorithm| Filter Size | Input/Output feature maps | 62 * --------------|----------------------------|-------------------------------------------| 63 * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5 | Input channels is greater than 3 | 64 * FFT | Not supported | | 65 * DirectConv | 9x9 | | 66 * GEMM | Any size | | 67 * 68 * Winograd FP16 requires fast maths enabled. 69 * 70 */ 71 class ClConv2d : public IClOperator 72 { 73 public: 74 /** Default constructor */ 75 ClConv2d(); 76 /** Default Destructor */ 77 ~ClConv2d(); 78 /** Prevent instances of this class from being copied (As this class contains pointers) */ 79 ClConv2d(const ClConv2d &) = delete; 80 /** Default move constructor */ 81 ClConv2d(ClConv2d &&) = default; 82 /** Prevent instances of this class from being copied (As this class contains pointers) */ 83 ClConv2d &operator=(const ClConv2d &) = delete; 84 /** Default move assignment operator */ 85 ClConv2d &operator=(ClConv2d &&) = default; 86 /** Set the src and dst tensors. 87 * 88 * Valid data layouts: 89 * - NHWC 90 * - NCHW 91 * 92 * Valid data type configurations: 93 * |src0 |src1 |src2 |dst | 94 * |:--------------|:------------------|:------|:--------------| 95 * |F16 |F16 |F16 |F16 | 96 * |F32 |F32 |F32 |F32 | 97 * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | 98 * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | 99 * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | 100 * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | 101 * 102 * @param[in] compile_context The compile context to be used. 103 * @param[in] src Source tensor info. 3 lower dimensions represent a single src [width, height, IFM], 104 * while every optional dimension from 4 and above represent a batch of srcs. 105 * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. 106 * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. 107 * Data type supported: Same as @p src, also could be QSYMM8_PER_CHANNEL if src is QASYMM8/QASYMM8_SIGNED. 108 * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. 109 * Data type supported: Same as @p src, except for src of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. 110 * @param[out] dst Destination tensor info. 3 lower dimensions represent a single dst [width, height, OFM], while the rest represent batch of dsts. 111 * Data types supported: Same as @p src. 112 * @param[in] conv2d_info Contains convolution 2d info described in @ref Conv2dInfo. 113 * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. Data type supported: Same as @p src. 114 */ 115 void configure(const CLCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, const Conv2dInfo &conv2d_info, 116 const WeightsInfo &weights_info = WeightsInfo()); 117 /** Static function to check if given info will lead to a valid configuration of @ref ClConv2d 118 * 119 * Similar to ClConv2d::configure() 120 * 121 * @return a status 122 */ 123 static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const Conv2dInfo &conv2d_info, 124 const WeightsInfo &weights_info = WeightsInfo()); 125 /** Static function to check if given info will return the convolution called by @ref ClConv2d 126 * 127 * @param[in] src Source tensor. 3 lower dimensions represent a single src [width, height, IFM], 128 * while every optional dimension from 4 and above represent a batch of srcs. 129 * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. 130 * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. 131 * Data type supported: Same as @p src, also could be QSYMM8_PER_CHANNEL if src is QASYMM8/QASYMM8_SIGNED. 132 * @param[in] dst Destination tensor. 3 lower dimensions represent a single dst [width, height, OFM], while the rest represent batch of dsts. 133 * Data types supported: Same as @p src. 134 * @param[in] conv2d_info Contains convolution 2d info described in @ref Conv2dInfo. 135 * @param[in] weights_info Specifies if the weights tensor has been reshaped with CLWeightsReshapeKernel. 136 * @param[in] gpu_target Specifies the @p GPUTarget. 137 * 138 * @return the Convolution Method Hint 139 */ 140 static ConvolutionMethod get_convolution_method(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const Conv2dInfo &conv2d_info, 141 const WeightsInfo &weights_info, const GPUTarget gpu_target); 142 // Inherited methods overridden: 143 void run(ITensorPack &tensors) override; 144 void prepare(ITensorPack &tensors) override; 145 experimental::MemoryRequirements workspace() const override; 146 147 private: 148 std::unique_ptr<IClOperator> _operator; 149 experimental::MemoryRequirements _aux_mem{}; 150 }; 151 } // namespace opencl 152 } // namespace arm_compute 153 #endif /* ARM_COMPUTE_CLCONV2D_H */ 154