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_CLDEPTHWISECONVOLUTION_H 25 #define ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H 26 27 #include "arm_compute/core/Types.h" 28 #include "arm_compute/runtime/CL/CLTensor.h" 29 #include "arm_compute/runtime/CL/functions/CLPermute.h" 30 #include "arm_compute/runtime/IFunction.h" 31 #include "arm_compute/runtime/MemoryGroup.h" 32 33 namespace arm_compute 34 { 35 class CLCompileContext; 36 class CLDepthwiseConvolutionLayerNativeKernel; 37 class ICLTensor; 38 39 /** Function to execute a depthwise convolution 40 * 41 * -# @ref CLDepthwiseConvolutionLayerNativeKernel 42 * -# @ref CLPermute (if the data layout is NCHW) 43 * 44 */ 45 class CLDepthwiseConvolutionLayer : public IFunction 46 { 47 public: 48 /** Default constructor */ 49 CLDepthwiseConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager = nullptr); 50 /** Prevent instances of this class from being copied (As this class contains pointers) */ 51 CLDepthwiseConvolutionLayer(const CLDepthwiseConvolutionLayer &) = delete; 52 /** Default move constructor */ 53 CLDepthwiseConvolutionLayer(CLDepthwiseConvolutionLayer &&) = default; 54 /** Prevent instances of this class from being copied (As this class contains pointers) */ 55 CLDepthwiseConvolutionLayer &operator=(const CLDepthwiseConvolutionLayer &) = delete; 56 /** Default move assignment operator */ 57 CLDepthwiseConvolutionLayer &operator=(CLDepthwiseConvolutionLayer &&) = default; 58 /** Default destructor */ 59 ~CLDepthwiseConvolutionLayer(); 60 /** Initialize the function's source, destination, weights and convolution information. 61 * 62 * Valid data layouts: 63 * - NHWC 64 * - NCHW 65 * 66 * Valid data type configurations: 67 * |src0 |src1 |src2 |dst | 68 * |:--------------|:------------------|:------|:--------------| 69 * |F16 |F16 |F16 |F16 | 70 * |F32 |F32 |F32 |F32 | 71 * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | 72 * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | 73 * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | 74 * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | 75 * 76 * @param[in] compile_context The compile context to be used. 77 * @param[in, out] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP16/FP32. Data layout supported: NHWC, NCHW 78 * @param[in] weights Weights tensor. These are 3D tensors with shape [kernel_x, kernel_y, IFM]. 79 * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8. 80 * @param[in] biases Biases tensor. A 1D tensor with shape [IFM]. Must be nullptr if not needed. 81 * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. 82 * @param[out] output Destination tensor. Pass in nullptr or @p input for in-place operation. Data type supported: same as @p input. 83 * @param[in] conv_info Padding and stride information to use for the convolution. 84 * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. 85 * @param[in] act_info (Optional) Activation layer information in case of a fused activation. 86 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). 87 * 88 * @note: For in-place support, please check @ref CLDepthwiseConvolutionLayerNativeKernel 89 */ 90 void configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 91 unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); 92 93 /** Initialize the function's source, destination, weights and convolution information. 94 * 95 * Similar to @ref CLDepthwiseConvolutionLayer::configure() 96 */ 97 void configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, 98 unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); 99 100 /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayer 101 * 102 * Similar to @ref CLDepthwiseConvolutionLayer::configure() 103 * 104 * @return a status 105 */ 106 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, 107 unsigned int depth_multiplier = 1, ActivationLayerInfo act_info = ActivationLayerInfo(), const Size2D &dilation = Size2D(1U, 1U)); 108 109 // Inherited methods overriden: 110 void run() override; 111 void prepare() override; 112 set_memory_group(std::shared_ptr<IMemoryManager> memory_manager)113 void set_memory_group(std::shared_ptr<IMemoryManager> memory_manager) 114 { 115 _memory_group = MemoryGroup(std::move(memory_manager)); 116 }; 117 118 private: 119 MemoryGroup _memory_group; 120 121 std::unique_ptr<CLDepthwiseConvolutionLayerNativeKernel> _dwc_native_kernel; 122 CLPermute _permute_input_to_nhwc; 123 CLPermute _permute_weights_to_nhwc; 124 CLPermute _permute_output_to_nchw; 125 126 CLTensor _permuted_input; 127 CLTensor _permuted_weights; 128 CLTensor _permuted_output; 129 CLTensor _output_multipliers; 130 CLTensor _output_shifts; 131 const ITensor *_original_weights; 132 const ITensor *_input; 133 const ITensor *_output; 134 135 bool _needs_permute; 136 bool _is_prepared; 137 bool _is_quantized; 138 }; 139 } // namespace arm_compute 140 #endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTION_H */ 141