giuros01 | 15ecc9a | 2018-12-06 10:47:34 +0000 | [diff] [blame] | 1 | /* |
George Wort | 5a97b28 | 2018-12-21 16:21:04 +0000 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
giuros01 | 15ecc9a | 2018-12-06 10:47:34 +0000 | [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 |
George Wort | 5a97b28 | 2018-12-21 16:21:04 +0000 | [diff] [blame] | 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
giuros01 | 15ecc9a | 2018-12-06 10:47:34 +0000 | [diff] [blame] | 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
George Wort | 5a97b28 | 2018-12-21 16:21:04 +0000 | [diff] [blame] | 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
giuros01 | 15ecc9a | 2018-12-06 10:47:34 +0000 | [diff] [blame] | 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_NEFUSEBATCHNORMALIZATIONKERNEL_H__ |
| 25 | #define __ARM_COMPUTE_NEFUSEBATCHNORMALIZATIONKERNEL_H__ |
| 26 | |
| 27 | #include "arm_compute/core/NEON/INEKernel.h" |
| 28 | |
| 29 | namespace arm_compute |
| 30 | { |
| 31 | // Forward declarations |
| 32 | class ITensor; |
| 33 | |
| 34 | /** OpenNE kernel to fuse the batch normalization node to a preceding convolution node */ |
| 35 | class NEFuseBatchNormalizationKernel : public INEKernel |
| 36 | { |
| 37 | public: |
| 38 | const char *name() const override |
| 39 | { |
| 40 | return "NEFuseBatchNormalizationKernel"; |
| 41 | } |
| 42 | /** Default constructor */ |
| 43 | NEFuseBatchNormalizationKernel(); |
| 44 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 45 | NEFuseBatchNormalizationKernel(const NEFuseBatchNormalizationKernel &) = delete; |
| 46 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 47 | NEFuseBatchNormalizationKernel &operator=(const NEFuseBatchNormalizationKernel &) = delete; |
| 48 | /** Allow instances of this class to be moved */ |
| 49 | NEFuseBatchNormalizationKernel(NEFuseBatchNormalizationKernel &&) = default; |
| 50 | /** Allow instances of this class to be moved */ |
| 51 | NEFuseBatchNormalizationKernel &operator=(NEFuseBatchNormalizationKernel &&) = default; |
| 52 | /** Default destructor */ |
| 53 | ~NEFuseBatchNormalizationKernel() = default; |
| 54 | /** Set the source, destination of the kernel |
| 55 | * |
| 56 | * @param[in] conv_weights Convolution layer weights tensor. Data type supported: F16/F32 |
| 57 | * @param[in] bn_mean Batch normalization layer mean tensor. Same as @p conv_weights |
| 58 | * @param[in] bn_var Batch normalization layer variance tensor. Same as @p conv_weights |
| 59 | * @param[out] fused_weights Output fused weights tensor. Same as @p conv_weights |
| 60 | * @param[out] fused_bias Output fused bias tensor. Same as @p conv_weights |
| 61 | * @param[in] conv_bias (Optional) Convolution layer bias tensor. Same as @p conv_weights |
| 62 | * @param[in] bn_beta (Optional) Batch normalization layer beta tensor. Same as @p conv_weights |
| 63 | * @param[in] bn_gamma (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights |
| 64 | * @param[in] epsilon (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f. |
| 65 | */ |
| 66 | void configure(const ITensor *conv_weights, const ITensor *bn_mean, const ITensor *bn_var, ITensor *fused_weights, ITensor *fused_bias, |
| 67 | const ITensor *conv_bias = nullptr, const ITensor *bn_beta = nullptr, const ITensor *bn_gamma = nullptr, |
| 68 | float epsilon = 0.001f); |
| 69 | /** Static function to check if given info will lead to a valid configuration of @ref NEFuseBatchNormalizationKernel |
| 70 | * |
| 71 | * @param[in] conv_weights Convolution layer weights tensor. Data type supported: F16/F32 |
| 72 | * @param[in] bn_mean Batch normalization layer mean tensor. Same as @p conv_weights |
| 73 | * @param[in] bn_var Batch normalization layer variance tensor. Same as @p conv_weights |
| 74 | * @param[in] fused_weights Output fused weights tensor. Same as @p conv_weights |
| 75 | * @param[in] fused_bias Output fused bias tensor. Same as @p conv_weights |
| 76 | * @param[in] conv_bias (Optional) Convolution layer bias tensor. Same as @p conv_weights |
| 77 | * @param[in] bn_beta (Optional) Batch normalization layer beta tensor. Same as @p conv_weights |
| 78 | * @param[in] bn_gamma (Optional) Batch normalization layer gamma tensor. Same as @p conv_weights |
| 79 | * @param[in] epsilon (Optional) Batch normalization layer epsilon parameter. Defaults to 0.001f. |
| 80 | * |
| 81 | * @return a status |
| 82 | */ |
| 83 | static Status validate(const ITensorInfo *conv_weights, const ITensorInfo *bn_mean, const ITensorInfo *bn_var, |
| 84 | const ITensorInfo *fused_weights, const ITensorInfo *fused_bias, |
| 85 | const ITensorInfo *conv_bias = nullptr, const ITensorInfo *bn_beta = nullptr, const ITensorInfo *bn_gamma = nullptr, |
| 86 | float epsilon = 0.001f); |
| 87 | |
| 88 | // Inherited methods overridden: |
| 89 | void run(const Window &window, const ThreadInfo &info) override; |
| 90 | |
| 91 | private: |
| 92 | const ITensor *_conv_weights; |
| 93 | const ITensor *_conv_bias; |
| 94 | const ITensor *_bn_mean; |
| 95 | const ITensor *_bn_var; |
| 96 | const ITensor *_bn_gamma; |
| 97 | const ITensor *_bn_beta; |
| 98 | ITensor *_fused_weights; |
| 99 | ITensor *_fused_bias; |
| 100 | float _epsilon; |
| 101 | bool _run_in_place_weights; |
| 102 | bool _run_in_place_bias; |
| 103 | |
| 104 | using FuseBatchNormFunction = void(const ITensor *conv_weights, const ITensor *conv_bias, ITensor *fused_weights, ITensor *fused_bias, |
| 105 | const ITensor *bn_mean, const ITensor *bn_var, const ITensor *bn_beta, const ITensor *bn_gamma, float epsilon, const Window &window); |
| 106 | |
| 107 | FuseBatchNormFunction *_func; |
| 108 | }; |
| 109 | } // namespace arm_compute |
| 110 | #endif /*__ARM_COMPUTE_NEFUSEBATCHNORMALIZATIONKERNEL_H__ */ |