Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 1 | /* |
Matthew Bentham | f1aeab9 | 2023-05-30 13:35:34 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2021, 2023 Arm Limited. |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +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 | */ |
SiCong Li | 9129549 | 2023-07-21 18:16:13 +0100 | [diff] [blame] | 24 | #include "arm_compute/function_info/ActivationLayerInfo.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 25 | |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 26 | #include "src/core/common/Macros.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 27 | #include "src/cpu/ICpuOperator.h" |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 28 | |
| 29 | namespace arm_compute |
| 30 | { |
| 31 | namespace cpu |
| 32 | { |
| 33 | /** Basic function to simulate a convolution layer. This function calls one of the following functions: |
| 34 | * -# @ref CpuGemm (executed only in case GEMM is required for the operation) |
| 35 | * -# @ref CpuWinogradConv2d (executed only in case Winograd is required for the operation) |
| 36 | * -# @ref CpuDirectConv2d (executed only in case Direct Convolution is required for the operation) |
| 37 | * |
| 38 | * |
| 39 | * The function selects one of the algorithms mentioned above based on: |
| 40 | * - The size of the kernel |
| 41 | * - Number of input/output feature maps |
| 42 | * - Amount of memory needed |
| 43 | * |
| 44 | * Generally GEMM-based convolution is executed when neither Winograd nor FFT nor Direct convolution can be performed. |
| 45 | * |
| 46 | * FP32 Algorithm| Filter Size | Input/Output feature maps | |
| 47 | * --------------|----------------------------------------------------|-------------------------------------------| |
| 48 | * Winograd | 3x3 1x3 3x1 5x1 1x5 5x5(fast maths) 7x1 1x7 | Input channels is greater than 3 | |
| 49 | * FFT | Squared kernels and greater than 9x9 | Input feature maps > Output feature maps | |
| 50 | * DirectConv | 9x9 | | |
| 51 | * GEMM | Any size | | |
| 52 | * |
| 53 | * Winograd 5x5 requires fast maths enabled. |
| 54 | * |
| 55 | * FP16 Algorithm| Filter Size | |
| 56 | * --------------|------------------| |
| 57 | * Winograd | Not supported | |
| 58 | * FFT | Not supported | |
| 59 | * DirectConv | 9x9 | |
| 60 | * GEMM | Any size | |
| 61 | * |
| 62 | * |
| 63 | */ |
| 64 | class CpuConv2d : public ICpuOperator |
| 65 | { |
| 66 | public: |
| 67 | /** Constructor */ |
| 68 | CpuConv2d(); |
| 69 | ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuConv2d); |
| 70 | /** Default destructor */ |
| 71 | ~CpuConv2d(); |
| 72 | /** Set the input and output tensors. |
| 73 | * |
| 74 | * Valid data layouts: |
| 75 | * - NHWC |
| 76 | * - NCHW |
| 77 | * |
| 78 | * Valid data type configurations: |
| 79 | * |src0 |src1 |src2 |dst | |
| 80 | * |:--------------|:------------------|:------|:--------------| |
| 81 | * |F16 |F16 |F16 |F16 | |
| 82 | * |F32 |F32 |F32 |F32 | |
| 83 | * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | |
| 84 | * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | |
| 85 | * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | |
| 86 | * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | |
| 87 | * |
| 88 | * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], |
| 89 | * while every optional dimension from 4 and above represent a batch of inputs. |
| 90 | * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. |
| 91 | * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
| 92 | * Data type supported: Same as @p src, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED. |
| 93 | * @param[in] biases Biases tensor info. Shared biases supported. Biases are 1D tensor with dimensions [OFM]. |
| 94 | * Data type supported: Same as @p src, except for input of QASYMM8/QASYMM8_SIGNED type where biases should be of S32 type. |
| 95 | * @param[out] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 96 | * Data types supported: Same as @p src. |
| 97 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 98 | * @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 |
| 99 | * tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input. |
| 100 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
| 101 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. Only RELU, BOUNDED_RELU and LU_BOUNDED_RELU supported. |
| 102 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 103 | * available which may introduce a drop of accuracy as well. Default is false |
| 104 | * @param[in] num_groups (Optional) Number of groups when performing a grouped convolution. num_groups != 1 is not supported |
| 105 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 106 | void configure(ITensorInfo *src, |
| 107 | ITensorInfo *weights, |
| 108 | const ITensorInfo *biases, |
| 109 | ITensorInfo *dst, |
| 110 | const PadStrideInfo &conv_info, |
| 111 | const WeightsInfo &weights_info = WeightsInfo(), |
| 112 | const Size2D &dilation = Size2D(1U, 1U), |
| 113 | const ActivationLayerInfo &act_info = ActivationLayerInfo(), |
| 114 | bool enable_fast_math = false, |
| 115 | unsigned int num_groups = 1); |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 116 | /** Static function to check if given info will lead to a valid configuration of @ref CpuConv2d |
| 117 | * |
| 118 | * Similar to CpuConv2d::configure() |
| 119 | * |
| 120 | * @return a status |
| 121 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 122 | static Status validate(const ITensorInfo *src, |
| 123 | const ITensorInfo *weights, |
| 124 | const ITensorInfo *biases, |
| 125 | const ITensorInfo *output, |
| 126 | const PadStrideInfo &conv_info, |
| 127 | const WeightsInfo &weights_info = WeightsInfo(), |
| 128 | const Size2D &dilation = Size2D(1U, 1U), |
| 129 | const ActivationLayerInfo &act_info = ActivationLayerInfo(), |
| 130 | bool enable_fast_math = false, |
| 131 | unsigned int num_groups = 1); |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 132 | /** Static function to check if given info will return the convolution called by @ref CpuConv2d |
| 133 | * |
| 134 | * @param[in] src Source tensor info. 3 lower dimensions represent a single input [width, height, IFM], |
| 135 | * while every optional dimension from 4 and above represent a batch of inputs. |
| 136 | * Data types supported: QASYMM8/QASYMM8_SIGNED/F16/F32. |
| 137 | * @param[in] weights Weights tensor info. Weights are 4D tensor with dimensions [kernel_x, kernel_y, IFM, OFM]. |
| 138 | * Data type supported:Same as @p src, also could be QSYMM8_PER_CHANNEL if input is QASYMM8/QASYMM8_SIGNED. |
| 139 | * @param[in] dst Destination tensor info. 3 lower dimensions represent a single output [width, height, OFM], while the rest represent batch of outputs. |
| 140 | * Data types supported: Same as @p src. |
| 141 | * @param[in] conv_info Contains padding and stride information described in @ref PadStrideInfo. |
| 142 | * @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 |
| 143 | * tensor has also been transposed with cpu::kernels::CpuGemmTranspose1xWKernel. Data type supported: Same as @p input. |
| 144 | * @param[in] dilation (Optional) Dilation, in elements, across x and y. Defaults to (1, 1). |
| 145 | * @param[in] act_info (Optional) Activation layer information in case of a fused activation. |
| 146 | * @param[in] enable_fast_math (Optional) Enable fast math computation. In case this flag were set, the function could dispatch the fastest implementation |
| 147 | * available which may introduce a drop of accuracy as well. Default is false |
| 148 | * |
| 149 | * @return the Convolution Method Hint |
| 150 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 151 | static ConvolutionMethod get_convolution_method(const ITensorInfo *src, |
| 152 | const ITensorInfo *weights, |
| 153 | const ITensorInfo *dst, |
| 154 | const PadStrideInfo &conv_info, |
| 155 | const WeightsInfo &weights_info = WeightsInfo(), |
| 156 | const Size2D &dilation = Size2D(1U, 1U), |
| 157 | const ActivationLayerInfo &act_info = ActivationLayerInfo(), |
| 158 | bool enable_fast_math = false); |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 159 | // Inherited methods overridden: |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 160 | void run(ITensorPack &tensors) override; |
| 161 | void prepare(ITensorPack &constants) override; |
Michalis Spyrou | b55f8e8 | 2021-07-22 11:23:11 +0100 | [diff] [blame] | 162 | experimental::MemoryRequirements workspace() const override; |
| 163 | |
| 164 | private: |
| 165 | std::unique_ptr<ICpuOperator> _function; |
| 166 | experimental::MemoryRequirements _aux_mem{}; |
| 167 | }; |
| 168 | } // namespace cpu |
| 169 | } // namespace arm_compute |