Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | f72f936 | 2018-01-12 16:29:45 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 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_NEDEPTHWISECONVOLUTIONKERNEL3x3_H__ |
| 25 | #define __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONKERNEL3x3_H__ |
| 26 | |
| 27 | #include "arm_compute/core/NEON/INEKernel.h" |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 28 | #include "arm_compute/core/NEON/kernels/convolution/depthwise/depthwise.hpp" |
| 29 | |
| 30 | #include <memory> |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 31 | |
| 32 | namespace arm_compute |
| 33 | { |
| 34 | class ITensor; |
| 35 | |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 36 | /** Interface for the kernel to run a 3x3 depthwise convolution on a tensor. */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 37 | class NEDepthwiseConvolutionLayer3x3Kernel : public INEKernel |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 38 | { |
| 39 | public: |
Anthony Barbier | e8a4983 | 2018-01-18 10:04:05 +0000 | [diff] [blame] | 40 | const char *name() const override |
| 41 | { |
| 42 | return "NEDepthwiseConvolutionLayer3x3Kernel"; |
| 43 | } |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 44 | /** Default constructor */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 45 | NEDepthwiseConvolutionLayer3x3Kernel(); |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 46 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 47 | NEDepthwiseConvolutionLayer3x3Kernel(const NEDepthwiseConvolutionLayer3x3Kernel &) = delete; |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 48 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 49 | NEDepthwiseConvolutionLayer3x3Kernel &operator=(const NEDepthwiseConvolutionLayer3x3Kernel &) = delete; |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 50 | /** Default Move Constructor. */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 51 | NEDepthwiseConvolutionLayer3x3Kernel(NEDepthwiseConvolutionLayer3x3Kernel &&) = default; |
Alex Gilday | c357c47 | 2018-03-21 13:54:09 +0000 | [diff] [blame] | 52 | /** Default move assignment operator */ |
Giorgio Arena | 04a8f8c | 2017-11-23 11:45:24 +0000 | [diff] [blame] | 53 | NEDepthwiseConvolutionLayer3x3Kernel &operator=(NEDepthwiseConvolutionLayer3x3Kernel &&) = default; |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 54 | /** Initialize the function's source, destination, conv and border_size. |
| 55 | * |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 56 | * @note Supported data layouts: NCHW and NHWC |
| 57 | * |
Georgios Pinitas | 20c246a | 2018-09-12 16:45:53 +0100 | [diff] [blame] | 58 | * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32. |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 59 | * @param[in] weights Weights tensor. This is a 3D tensor with dimensions [3, 3, IFM] for NCHW or [IFM, 3, 3] if NHWC data layout. Data type supported: Same as @p input. |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 60 | * @param[out] output Destination tensor. Data type supported: Same as @p input. |
| 61 | * @param[in] conv_info Padding and stride information to use for the convolution. |
| 62 | * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. |
| 63 | * @param[in] data_layout (Optional) Data layout of the input and weights tensor |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 64 | */ |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 65 | void configure(const ITensor *input, const ITensor *weights, ITensor *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1, DataLayout data_layout = DataLayout::NCHW); |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 66 | /** Static method that checks if optimized execution is supported for the given parameters |
| 67 | * |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 68 | * @param[in] input_shape Input shape |
| 69 | * @param[in] conv_info Padding and stride information to use for the convolution. |
| 70 | * @param[in] dt Data type of the input and weights |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 71 | * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 72 | * @param[in] data_layout (Optional) Data layout of the input and weights tensor |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 73 | * |
| 74 | * @return True if the optimized kernels can be executed else false |
| 75 | */ |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 76 | static bool is_optimized_execution_possible(TensorShape input_shape, PadStrideInfo conv_info, DataType dt, unsigned int depth_multiplier = 1, DataLayout data_layout = DataLayout::NCHW); |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 77 | /** Generates the convolver object */ |
| 78 | void generate_convolver(); |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 79 | |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 80 | /** Static function to check if given info will lead to a valid configuration of @ref NEDepthwiseConvolutionLayer3x3Kernel |
| 81 | * |
| 82 | * @note Supported data layouts: NCHW and NHWC |
| 83 | * |
Georgios Pinitas | 20c246a | 2018-09-12 16:45:53 +0100 | [diff] [blame] | 84 | * @param[in] input Source tensor. DataType supported: QASYMM8/F16/F32. |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 85 | * @param[in] weights Weights tensor. This is a 3D tensor with dimensions [3, 3, IFM] for NCHW or [IFM, 3, 3] if NHWC data layout. Data type supported: Same as @p input. |
| 86 | * @param[in] output Destination tensor. Data type supported: Same as @p input. |
| 87 | * @param[in] conv_info Padding and stride information to use for the convolution. |
| 88 | * @param[in] depth_multiplier (Optional) Multiplier to apply to the input's depth in order to retrieve the output's depth. Defaults to 1. |
| 89 | * |
| 90 | * @return a status |
| 91 | */ |
| 92 | static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *output, const PadStrideInfo &conv_info, unsigned int depth_multiplier = 1); |
| 93 | |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 94 | // Inherited methods overridden: |
| 95 | void run(const Window &window, const ThreadInfo &info) override; |
| 96 | BorderSize border_size() const override; |
| 97 | |
| 98 | private: |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 99 | void configure_generic(); |
| 100 | void configure_optimized(); |
Abe Mbise | 7784c83 | 2018-05-31 16:48:41 +0100 | [diff] [blame] | 101 | |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 102 | void run_generic(const Window &window, const ThreadInfo &info); |
| 103 | void run_optimized(const Window &window, const ThreadInfo &info); |
Georgios Pinitas | be0ae93 | 2018-03-13 13:08:12 +0000 | [diff] [blame] | 104 | /** Creates an optimized backend convolver object |
| 105 | * |
| 106 | * @note Convolver of strides 1,2 and convolution size of 3 is currently supported |
| 107 | * |
| 108 | * @param[in] conv_info Padding and stride information to use for the convolution |
| 109 | * @param[in] w Weights tensor |
| 110 | * @param[in] in Input tensor |
| 111 | * @param[in] out Output tensor |
| 112 | * @param[in] setup_strides (Optional) Boolean to enable setting the strides of the tensors |
| 113 | * in the convolver in case of padding. Defaults to false |
| 114 | * |
| 115 | * @return A convolver object or nullptr if the configuration is not supported |
| 116 | */ |
| 117 | std::unique_ptr<depthwise::IDepthwiseConvolution> create_convolver_object(PadStrideInfo conv_info, |
| 118 | const ITensor *w, |
| 119 | const ITensor *in, |
| 120 | ITensor *out, |
| 121 | bool setup_strides = false); |
Georgios Pinitas | 4074c99 | 2018-01-30 18:13:46 +0000 | [diff] [blame] | 122 | |
| 123 | private: |
| 124 | BorderSize _border_size; |
| 125 | const ITensor *_input; |
| 126 | ITensor *_output; |
| 127 | const ITensor *_weights; |
| 128 | PadStrideInfo _conv_info; |
| 129 | std::unique_ptr<depthwise::IDepthwiseConvolution> _convolver; |
| 130 | unsigned int _num_elems_written_per_iteration; |
| 131 | bool _run_optimized; |
Giorgio Arena | 7657224 | 2018-04-04 17:44:26 +0100 | [diff] [blame] | 132 | unsigned int _depth_multiplier; |
Michalis Spyrou | 7362f0d | 2017-10-18 17:58:22 +0100 | [diff] [blame] | 133 | }; |
| 134 | } // namespace arm_compute |
Anthony Barbier | e8a4983 | 2018-01-18 10:04:05 +0000 | [diff] [blame] | 135 | #endif /* __ARM_COMPUTE_NEDEPTHWISECONVOLUTIONKERNEL3x3_H__ */ |