1 /* 2 * Copyright (c) 2019-2020 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_CLDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H 25 #define ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H 26 27 #include "src/core/CL/ICLKernel.h" 28 29 #include "arm_compute/core/KernelDescriptors.h" 30 31 namespace arm_compute 32 { 33 class ICLTensor; 34 35 /** Interface for the kernel to run a MxN depthwise convolution. M and N are respectively the rows and columns of the filter 36 This kernel assumes that tensor for the weights is NOT reshaped (Native version) */ 37 class CLDepthwiseConvolutionLayerNativeKernel : public ICLKernel 38 { 39 public: 40 /** Default Constructor */ 41 CLDepthwiseConvolutionLayerNativeKernel(); 42 /** Prevent instances of this class from being copied (As this class contains pointers) */ 43 CLDepthwiseConvolutionLayerNativeKernel(const CLDepthwiseConvolutionLayerNativeKernel &) = delete; 44 /** Prevent instances of this class from being copied (As this class contains pointers) */ 45 CLDepthwiseConvolutionLayerNativeKernel &operator=(const CLDepthwiseConvolutionLayerNativeKernel &) = delete; 46 /** Allow instances of this class to be moved */ 47 CLDepthwiseConvolutionLayerNativeKernel(CLDepthwiseConvolutionLayerNativeKernel &&) = default; 48 /** Allow instances of this class to be moved */ 49 CLDepthwiseConvolutionLayerNativeKernel &operator=(CLDepthwiseConvolutionLayerNativeKernel &&) = default; 50 /** Initialize the function's source, destination and parameters 51 * 52 * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC 53 * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, N, M]. 54 * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8. 55 * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. 56 * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. 57 * @param[out] output Destination tensor. Data type supported: Same as @p input. 58 * @param[in] dwc_weights_info Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread 59 * @param[in] dwc_info Depthwise convolution layer info 60 * @param[in] conv_info Padding and stride information to use for the convolution. 61 * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. 62 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). 63 * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization, 64 * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 65 * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization, 66 * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 67 */ 68 void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info, 69 const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U), 70 const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr); 71 /** Initialize the function's source, destination and parameters 72 * 73 * @param[in] compile_context The compile context to be used. 74 * @param[in] input Source tensor. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC 75 * @param[in] weights Weights tensor. A 3D tensor with dimensions [IFM, N, M]. 76 * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8. 77 * @param[in] biases Biases tensor. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. 78 * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. 79 * @param[out] output Destination tensor. Data type supported: Same as @p input. 80 * @param[in] dwc_weights_info Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread 81 * @param[in] dwc_info Depthwise convolution layer info 82 * @param[in] conv_info Padding and stride information to use for the convolution. 83 * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. 84 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). 85 * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization, 86 * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 87 * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization, 88 * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 89 */ 90 void configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const DWCWeightsKernelInfo &dwc_weights_info, 91 const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U), 92 const ICLTensor *output_multipliers = nullptr, const ICLTensor *output_shifts = nullptr); 93 /** Static function to check if given info will lead to a valid configuration of @ref CLDepthwiseConvolutionLayerNativeKernel 94 * 95 * @param[in] input Source tensor info. Data type supported: QASYMM8/QASYMM8_SIGNED/FP32/FP16. Data layout supported: NHWC 96 * @param[in] weights Weights tensor info. A 3D tensor with dimensions [IFM, N, M]. 97 * Data type supported: Same as @p input or QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL when @p input is QASYMM8. 98 * @param[in] biases Biases tensor info. A 1D tensor with dimensions [IFM]. Must be nullptr if not needed. 99 * Data type supported: Same as @p input, S32 when input is QASYMM8/QASYMM8_SIGNED. 100 * @param[in] output Destination tensor info. Data type supported: Same as @p input. 101 * @param[in] dwc_weights_info Depthwise convolution layer weights info to retrieve the number of output elements processed by each thread 102 * @param[in] dwc_info Depthwise convolution layer info 103 * @param[in] conv_info Padding and stride information to use for the convolution. 104 * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. 105 * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). 106 * @param[in] output_multipliers (Optional) Output multipliers tensor for quantized computations. In case of per-channel quantization, 107 * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 108 * @param[in] output_shifts (Optional) Output shifts tensor for quantized computations. In case of per-channel quantization, 109 * the number of multipliers must be equal to the number of filters (IFM). Supported data types: S32 110 * 111 * @return a status 112 */ 113 static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const DWCWeightsKernelInfo &dwc_weights_info, 114 const DWCKernelInfo &dwc_info, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, const Size2D &dilation = Size2D(1U, 1U), 115 const ITensorInfo *output_multipliers = nullptr, const ITensorInfo *output_shifts = nullptr); 116 117 // Inherited methods overridden: 118 void run(const Window &window, cl::CommandQueue &queue) override; 119 120 private: 121 const ICLTensor *_input; 122 const ICLTensor *_weights; 123 const ICLTensor *_biases; 124 ICLTensor *_output; 125 unsigned int _depth_multiplier; 126 const ICLTensor *_output_multipliers; 127 const ICLTensor *_output_shifts; 128 bool _is_quantized; 129 }; 130 } // namespace arm_compute 131 #endif /*ARM_COMPUTE_CLDEPTHWISECONVOLUTIONLAYERNATIVEKERNEL_H */ 132