Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 1 | /* |
Matthew Bentham | f1aeab9 | 2023-05-30 13:35:34 +0000 | [diff] [blame^] | 2 | * Copyright (c) 2017-2023 Arm Limited. |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +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 |
| 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 | */ |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 24 | #ifndef ARM_COMPUTE_NEGEMMCONVOLUTIONLAYER_H |
| 25 | #define ARM_COMPUTE_NEGEMMCONVOLUTIONLAYER_H |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/runtime/IFunction.h" |
| 28 | |
Matthew Bentham | f1aeab9 | 2023-05-30 13:35:34 +0000 | [diff] [blame^] | 29 | #include "arm_compute/core/ActivationLayerInfo.h" |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 30 | #include "arm_compute/core/Types.h" |
Manuel Bottini | 29599d0 | 2021-07-06 15:01:35 +0100 | [diff] [blame] | 31 | #include "arm_compute/runtime/IFunction.h" |
| 32 | #include "arm_compute/runtime/IMemoryManager.h" |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 33 | #include "arm_compute/runtime/IWeightsManager.h" |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 34 | #include "arm_compute/runtime/MemoryGroup.h" |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 35 | |
| 36 | #include <memory> |
| 37 | |
| 38 | namespace arm_compute |
| 39 | { |
| 40 | class ITensor; |
Manuel Bottini | 29599d0 | 2021-07-06 15:01:35 +0100 | [diff] [blame] | 41 | class ITensorInfo; |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 42 | |
Michele Di Giorgio | 33f41fa | 2021-03-09 14:09:08 +0000 | [diff] [blame] | 43 | /** Basic function to compute the convolution layer. This function calls the following kernels/functions: |
Gian Marco Iodice | 597a856 | 2018-08-01 15:06:06 +0100 | [diff] [blame] | 44 | * |
Georgios Pinitas | 1988463 | 2021-08-16 12:38:54 +0100 | [diff] [blame] | 45 | * -# @ref cpu::CpuGemmConv2d |
Gian Marco Iodice | 597a856 | 2018-08-01 15:06:06 +0100 | [diff] [blame] | 46 | * |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 47 | */ |
| 48 | class NEGEMMConvolutionLayer : public IFunction |
| 49 | { |
| 50 | public: |
| 51 | /** Constructor */ |
Michalis Spyrou | 1a569a3 | 2019-09-10 17:20:34 +0100 | [diff] [blame] | 52 | NEGEMMConvolutionLayer(const std::shared_ptr<IMemoryManager> &memory_manager = nullptr, IWeightsManager *weights_manager = nullptr); |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 53 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 54 | NEGEMMConvolutionLayer(const NEGEMMConvolutionLayer &) = delete; |
Michalis Spyrou | 770dfeb | 2020-11-04 18:55:34 +0000 | [diff] [blame] | 55 | /** Prevent instances of this class from being moved (As this class contains non movable objects) */ |
| 56 | NEGEMMConvolutionLayer(NEGEMMConvolutionLayer &&) = delete; |
Georgios Pinitas | 1562be3 | 2018-03-08 19:09:19 +0000 | [diff] [blame] | 57 | /** Prevent instances of this class from being copied (As this class contains pointers) */ |
| 58 | NEGEMMConvolutionLayer &operator=(const NEGEMMConvolutionLayer &) = delete; |
Michalis Spyrou | 770dfeb | 2020-11-04 18:55:34 +0000 | [diff] [blame] | 59 | /** Prevent instances of this class from being moved (As this class contains non movable objects) */ |
| 60 | NEGEMMConvolutionLayer &operator=(NEGEMMConvolutionLayer &&) = delete; |
Michalis Spyrou | ebcebf1 | 2020-10-21 00:04:14 +0100 | [diff] [blame] | 61 | /** Default destructor */ |
| 62 | ~NEGEMMConvolutionLayer(); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 63 | /** Set the input and output tensors. |
| 64 | * |
Teresa Charlin | 6268742 | 2021-04-28 10:58:49 +0100 | [diff] [blame] | 65 | * Valid data layouts: |
| 66 | * - NHWC |
| 67 | * - NCHW |
| 68 | * |
| 69 | * Valid data type configurations: |
| 70 | * |src0 |src1 |src2 |dst | |
| 71 | * |:--------------|:------------------|:--------|:--------------| |
| 72 | * |F16 |F16 |F16 |F16 | |
| 73 | * |F32 |F32 |F32 |F32 | |
| 74 | * |BFLOAT16 |BFLOAT16 |BFLOAT16 |BFLOAT16 | |
| 75 | * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | |
| 76 | * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | |
| 77 | * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | |
| 78 | * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | |
| 79 | * |
Georgios Pinitas | 69a9ac4 | 2021-07-22 13:30:13 +0100 | [diff] [blame] | 80 | * @param[in] input Source tensor. 3 lower dimensions represent a single input [width, height, IFM], |
| 81 | * while every optional dimension from 4 and above represent a batch of inputs. |
| 82 | * Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. |
| 83 | * @param[in] weights Weights tensor. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
| 84 | * Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. |
| 85 | * @param[in] biases Biases tensor. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. |
| 86 | * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. |
| 87 | * @param[out] output Destination tensor. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 88 | * Data types supported: Same as @p input. |
| 89 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 90 | * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights |
| 91 | * tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input. |
| 92 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
| 93 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. |
| 94 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 95 | * available which may introduce a drop of accuracy as well. Default is false |
| 96 | * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 97 | */ |
Alex Gilday | 7da29b6 | 2018-03-23 14:16:00 +0000 | [diff] [blame] | 98 | void configure(const ITensor *input, const ITensor *weights, const ITensor *biases, ITensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info = WeightsInfo(), |
Georgios Pinitas | 69a9ac4 | 2021-07-22 13:30:13 +0100 | [diff] [blame] | 99 | const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), bool enable_fast_math = false, unsigned int num_groups = 1); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 100 | /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMConvolutionLayer |
| 101 | * |
Georgios Pinitas | 69a9ac4 | 2021-07-22 13:30:13 +0100 | [diff] [blame] | 102 | * @param[in] input Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], |
| 103 | * while every optional dimension from 4 and above represent a batch of inputs. |
| 104 | * Data types supported: QASYMM8/QASYMM8_SIGNED/BFLOAT16/F16/F32. |
| 105 | * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
| 106 | * Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8_PER_CHANNEL/BFLOAT16/F16/F32. |
| 107 | * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. |
| 108 | * Data type supported: Should match @p input data type, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. |
| 109 | * @param[in] output Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 110 | * Data types supported: Same as @p input. |
| 111 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 112 | * @param[in] weights_info Specifies if the weights tensor has been reshaped with NEWeightsReshapeKernel. If this is not part of the fully connected layer the weights |
| 113 | * tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input. |
| 114 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
| 115 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. |
| 116 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 117 | * available which may introduce a drop of accuracy as well. Default is false |
| 118 | * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 119 | * |
| 120 | * @return a status |
| 121 | */ |
| 122 | static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
Georgios Pinitas | 69a9ac4 | 2021-07-22 13:30:13 +0100 | [diff] [blame] | 123 | const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), |
| 124 | bool enable_fast_math = false, unsigned int num_groups = 1); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 125 | |
Francesco Petrogalli | 553f695 | 2022-06-30 10:22:01 +0000 | [diff] [blame] | 126 | /** Static function to check if there is an optimized version of |
| 127 | * GEMM available for the input parameters. |
| 128 | * |
| 129 | * The method is intended to be used to find out the optimal |
| 130 | * memory layout to be used for the weights tensor when running |
| 131 | * variable weights execution. |
| 132 | * |
| 133 | * The user can query the database of optimised kernels in |
| 134 | * arm_gemm by specifying one of the enumerations of |
Ramy Elgammal | 9178002 | 2022-07-20 14:57:37 +0100 | [diff] [blame] | 135 | * arm_compute::WeightFormat in the weight_format field of the input |
Francesco Petrogalli | 553f695 | 2022-06-30 10:22:01 +0000 | [diff] [blame] | 136 | * parameter weights_info. In case of success, the method |
| 137 | * writes the expected format in the output parameter |
| 138 | * expected_weight_format. The expected_weight_format can than be |
| 139 | * used in the configure method of the class for retrieving the |
| 140 | * best optimal kernel. |
| 141 | * |
| 142 | * Use case one - query for a specific format: |
| 143 | * |
Ramy Elgammal | 9178002 | 2022-07-20 14:57:37 +0100 | [diff] [blame] | 144 | * WeightInfo weights_info(..., arm_compute::WeightFormat::OHWIo4, ...); // Set the value of the input query. |
Francesco Petrogalli | 553f695 | 2022-06-30 10:22:01 +0000 | [diff] [blame] | 145 | * if (NEGEMMConvolutionlayer::has_opt_impl(WeightFormat(), ...., weights_info, ...)) |
| 146 | * { |
| 147 | * auto conv = std::unique_ptr<NEGEMMConvolutionlayer>(); |
| 148 | * conv->configure(..., weights_info, ...); // uses the same WeightFormat the user wanted originally, OHWYo4. |
| 149 | * conv->run(...); |
| 150 | * } |
| 151 | * |
| 152 | * Use case two - query for any format that would be optimal for the GEMM to execute: |
| 153 | * |
Ramy Elgammal | 9178002 | 2022-07-20 14:57:37 +0100 | [diff] [blame] | 154 | * WeightInfo weights_info(..., arm_compute::WeightFormat::ANY, ...); // Set the value of the input query. |
| 155 | * arm_compute::WeightFormat expected_wf; |
Francesco Petrogalli | 553f695 | 2022-06-30 10:22:01 +0000 | [diff] [blame] | 156 | * if (NEGEMMConvolutionlayer::has_opt_impl(expected_wf, ...., weights_info, ...)) |
| 157 | * { |
| 158 | * auto conv = std::unique_ptr<NEGEMMConvolutionlayer>(); |
| 159 | * // ... code to convert the layout of the weights tensor to the layout returned by has_opt_impl |
| 160 | * WeightInfo new_weights_info(..., expected_wf, ...); // Set the value of the WeightFormat returned by has_opt_impl. |
| 161 | * conv->configure(..., new_weights_info, ...); |
| 162 | * conv->run(...); |
| 163 | * } |
| 164 | * |
| 165 | * Notice that a GEMM configured with a WeightFormat other than |
| 166 | * UNSPECIFIED will run GEMM with variable weights mode. |
| 167 | * |
| 168 | * @param[out] expected_weight_format The arm_compute::WeightFormat expected by the kernel. |
| 169 | * @param[in] src Source tensor info. |
| 170 | * @param[in] weights Weights tensor info. |
| 171 | * @param[in] biases Biases tensor info. Shared biases supported. |
| 172 | * @param[in] dst Destination tensor info. |
| 173 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 174 | * @param[in] weights_info (optional) Specifies additional configuration parameters for the weights of the GEMM computation. |
| 175 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
| 176 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. And no activation (i.e. Linear) which is the default value. |
| 177 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 178 | * |
| 179 | * @return a Status |
| 180 | */ |
Ramy Elgammal | 9178002 | 2022-07-20 14:57:37 +0100 | [diff] [blame] | 181 | static Status has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, |
Francesco Petrogalli | 553f695 | 2022-06-30 10:22:01 +0000 | [diff] [blame] | 182 | const PadStrideInfo &conv_info, |
| 183 | const WeightsInfo &weights_info = WeightsInfo(), const Size2D &dilation = Size2D(1U, 1U), const ActivationLayerInfo &act_info = ActivationLayerInfo(), |
| 184 | bool enable_fast_math = false); |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 185 | // Inherited methods overridden: |
| 186 | void run() override; |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 187 | void prepare() override; |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 188 | |
| 189 | private: |
Manuel Bottini | 29599d0 | 2021-07-06 15:01:35 +0100 | [diff] [blame] | 190 | struct Impl; |
| 191 | std::unique_ptr<Impl> _impl; |
Isabella Gottardi | 6acc6ad | 2018-02-02 17:19:18 +0000 | [diff] [blame] | 192 | }; |
Georgios Pinitas | 041f36d | 2018-09-18 18:38:37 +0100 | [diff] [blame] | 193 | } // namespace arm_compute |
Manuel Bottini | 29599d0 | 2021-07-06 15:01:35 +0100 | [diff] [blame] | 194 | #endif /* ARM_COMPUTE_NEGEMMCONVOLUTIONLAYER_H */ |