Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame^] | 2 | * Copyright (c) 2019-2020 Arm Limited. |
Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +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_CORE_KERNEL_DESCRIPTORS_H |
| 25 | #define ARM_COMPUTE_CORE_KERNEL_DESCRIPTORS_H |
Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +0000 | [diff] [blame] | 26 | |
Sang-Hoon Park | c261798 | 2020-05-20 22:13:47 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/PixelValue.h" |
Gian Marco Iodice | ca1f460 | 2019-07-16 15:46:48 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/Types.h" |
| 29 | |
Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +0000 | [diff] [blame] | 30 | namespace arm_compute |
| 31 | { |
Georgios Pinitas | 8be9148 | 2019-03-26 17:23:28 +0000 | [diff] [blame] | 32 | /** Descriptor for FFT scale kernels */ |
| 33 | struct FFTScaleKernelInfo |
Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +0000 | [diff] [blame] | 34 | { |
Georgios Pinitas | 8be9148 | 2019-03-26 17:23:28 +0000 | [diff] [blame] | 35 | float scale{ 0.f }; /**< Axis to perform the kernel on. */ |
| 36 | bool conjugate{ true }; /**< Flag to conjugate the output/ */ |
| 37 | }; |
| 38 | |
| 39 | /** Descriptor for FFT digit reverse kernels */ |
| 40 | struct FFTDigitReverseKernelInfo |
| 41 | { |
| 42 | unsigned int axis{ 0 }; /**< Axis to perform the kernel on. */ |
| 43 | bool conjugate{ false }; /**< Flag to conjugate the output/ */ |
| 44 | }; |
| 45 | |
| 46 | /** Descriptor used by the FFT core kernels */ |
| 47 | struct FFTRadixStageKernelInfo |
| 48 | { |
| 49 | unsigned int axis{ 0 }; /**< Axis to run the kernel on. */ |
Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +0000 | [diff] [blame] | 50 | unsigned int radix{ 0 }; /**< Radix to use. */ |
| 51 | unsigned int Nx{ 0 }; /**< Nx coefficient. */ |
| 52 | bool is_first_stage{ false }; /**< Flags if the FFT kernels is the first stage of a decomposed FFT. */ |
| 53 | }; |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 54 | |
| 55 | /** Descriptor used by the GEMM kernels */ |
| 56 | struct GEMMKernelInfo |
| 57 | { |
morgolock | aba2f91 | 2020-05-05 16:28:19 +0100 | [diff] [blame] | 58 | GEMMKernelInfo() = default; |
| 59 | GEMMKernelInfo( |
| 60 | unsigned int im, |
| 61 | unsigned int in, |
| 62 | unsigned int ik, |
| 63 | unsigned int idepth_output_gemm3d, |
| 64 | bool ireinterpret_input_as_3d, |
| 65 | bool ibroadcast_bias, |
| 66 | bool ifp_mixed_precision, |
| 67 | ActivationLayerInfo iactivation_info, |
| 68 | int inmult_transpose1xW_width, |
| 69 | int imult_interleave4x4_height, |
| 70 | GEMMLHSMatrixInfo ilhs_info, |
| 71 | GEMMRHSMatrixInfo irhs_info, |
| 72 | int32_t ina_offset, |
| 73 | int32_t inb_offset) |
| 74 | : m(im), n(in), k(ik), depth_output_gemm3d(idepth_output_gemm3d), reinterpret_input_as_3d(ireinterpret_input_as_3d), broadcast_bias(ibroadcast_bias), fp_mixed_precision(ifp_mixed_precision), |
| 75 | activation_info(iactivation_info), mult_transpose1xW_width(inmult_transpose1xW_width), mult_interleave4x4_height(imult_interleave4x4_height), lhs_info(ilhs_info), rhs_info(irhs_info), |
| 76 | a_offset(ina_offset), b_offset(inb_offset) |
| 77 | { |
| 78 | } |
| 79 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 80 | unsigned int m{ 0 }; /**< Number of LHS rows*/ |
| 81 | unsigned int n{ 0 }; /**< Number of RHS columns*/ |
| 82 | unsigned int k{ 0 }; /**< Number of LHS columns or RHS rows */ |
| 83 | unsigned int depth_output_gemm3d{ 0 }; /**< Depth of the output tensor in case is reinterpreted as 3D */ |
| 84 | bool reinterpret_input_as_3d{ false }; /**< Flag used to reinterpret the input as 3D */ |
| 85 | bool broadcast_bias{ false }; /**< Flag used to broadcast the bias addition */ |
| 86 | bool fp_mixed_precision{ false }; /**< Flag used to indicate wider accumulators (32 bit instead of 16 for FP16). */ |
| 87 | ActivationLayerInfo activation_info{}; /**< Activation function to perform after the matrix multiplication */ |
| 88 | int mult_transpose1xW_width{ 1 }; /**< Multiplication factor for the width of the 1xW transposed block */ |
| 89 | int mult_interleave4x4_height{ 1 }; /**< Multiplication factor for the height of the 4x4 interleaved block */ |
| 90 | GEMMLHSMatrixInfo lhs_info{}; /**< LHS matrix information used to retrieve the number of rows processed by each thread */ |
| 91 | GEMMRHSMatrixInfo rhs_info{}; /**< RHS matrix information used for reshaping the RHS matrix */ |
| 92 | int32_t a_offset{ 0 }; /**< Offset to be added to each element of the matrix A */ |
| 93 | int32_t b_offset{ 0 }; /**< Offset to be added to each element of the matrix B */ |
| 94 | GEMMLowpOutputStageInfo output_stage{}; /**< GEMMLowp output stage information */ |
Gian Marco Iodice | 7026b30 | 2019-06-26 17:18:11 +0100 | [diff] [blame] | 95 | }; |
Gian Marco Iodice | 9285adb | 2019-09-05 16:10:27 +0100 | [diff] [blame] | 96 | |
| 97 | /** Descriptor used by the depthwise convolution kernels */ |
| 98 | struct DWCKernelInfo |
| 99 | { |
| 100 | ActivationLayerInfo activation_info{}; /**< Activation function to perform after the depthwise convolution */ |
| 101 | }; |
| 102 | |
| 103 | /** Descriptor used by the depthwise convolution kernels to retrieve the number of output elements processed by each thread */ |
| 104 | struct DWCWeightsKernelInfo |
| 105 | { |
| 106 | unsigned int n0{ 0 }; /**< Number of columns processed by each thread */ |
| 107 | }; |
Sang-Hoon Park | 62eeb53 | 2019-10-29 13:13:19 +0000 | [diff] [blame] | 108 | |
| 109 | /** Descriptor used by the softmax kernels */ |
| 110 | struct SoftmaxKernelInfo |
| 111 | { |
Sang-Hoon Park | 0779fec | 2019-11-13 17:08:12 +0000 | [diff] [blame] | 112 | float beta{ 1.f }; /**< A scaling factor for the exponent with default value 1.0 */ |
| 113 | bool is_log{ false }; /**< Flag used to perform Log Softmax operation */ |
| 114 | DataType input_data_type{ DataType::UNKNOWN }; /**< Input tensor data type */ |
Sang-Hoon Park | 62eeb53 | 2019-10-29 13:13:19 +0000 | [diff] [blame] | 115 | }; |
Michele Di Giorgio | 4536193 | 2019-12-19 13:53:44 +0000 | [diff] [blame] | 116 | |
| 117 | /** Descriptor used by the direct convolution layer output stage kernels */ |
| 118 | struct DirectConvolutionLayerOutputStageKernelInfo |
| 119 | { |
| 120 | int32_t result_fixedpoint_multiplier{ 0 }; /**< Result output stage multiplier used for quantizing */ |
| 121 | int32_t result_shift{ 0 }; /**< Result output stage shift used for quantizing */ |
| 122 | int32_t result_offset_after_shift{ 0 }; /**< Result offset used for quantizing */ |
| 123 | DataType output_data_type{ DataType::UNKNOWN }; /**< Output tensor data type to use if the output is not initialized */ |
| 124 | }; |
Georgios Pinitas | 55a687d | 2020-01-30 12:00:23 +0000 | [diff] [blame] | 125 | |
| 126 | struct InstanceNormalizationLayerKernelInfo |
| 127 | { |
| 128 | /** Default constructor */ |
| 129 | InstanceNormalizationLayerKernelInfo() |
| 130 | : InstanceNormalizationLayerKernelInfo(1.f, 0.f, 1e-12, true) |
| 131 | { |
| 132 | } |
| 133 | /** Constructor |
| 134 | * |
| 135 | * @param[in] gamma The scale scalar value applied to the normalized tensor. |
| 136 | * @param[in] beta The offset scalar value applied to the normalized tensor |
| 137 | * @param[in] epsilon Lower bound value for the normalization. |
| 138 | * @param[in] use_mixed_precision Use mixed precision in case of FP16 execution. |
| 139 | */ |
| 140 | InstanceNormalizationLayerKernelInfo(float gamma, float beta, float epsilon, bool use_mixed_precision) |
| 141 | : gamma(gamma), beta(beta), epsilon(epsilon), use_mixed_precision(use_mixed_precision) |
| 142 | { |
| 143 | } |
| 144 | |
| 145 | float gamma; /**< The scale scalar value applied to the normalized tensor. Defaults to 1.0 */ |
| 146 | float beta; /**< The offset scalar value applied to the normalized tensor. Defaults to 0.0 */ |
| 147 | float epsilon; /**< Lower bound value for the normalization. Defaults to 1e-12 */ |
| 148 | bool use_mixed_precision; /**< Use mixed precision in case of FP16 execution. Defaults to true */ |
| 149 | }; |
Michele Di Giorgio | a602f03 | 2020-03-12 19:34:33 +0000 | [diff] [blame] | 150 | |
| 151 | struct GEMMLowpReductionKernelInfo |
| 152 | { |
| 153 | /** Default constructor */ |
| 154 | GEMMLowpReductionKernelInfo() = default; |
| 155 | /** Constructor |
| 156 | * |
| 157 | * @param[in] k Number of matrix columns/rows. |
| 158 | * @param[in] is_reshaped True if the input tensor has been reshaped. |
| 159 | * @param[in] scalar Scalar value to multiply each reduced column/row by. |
| 160 | * @param[in] mul_by_scalar True if each column/row reduction has to be multiplied by a scalar value. |
| 161 | */ |
| 162 | GEMMLowpReductionKernelInfo(int32_t k, bool is_reshaped, int32_t scalar, bool mul_by_scalar) |
| 163 | : k(k), is_reshaped(is_reshaped), scalar(scalar), mul_by_scalar(mul_by_scalar) |
| 164 | { |
| 165 | } |
| 166 | |
| 167 | int32_t k{ 0 }; /**< Number of matrix columns/rows */ |
| 168 | bool is_reshaped{ false }; /**< True if the input tensor has been reshaped */ |
| 169 | int32_t scalar{ 0 }; /**< Scalar value to multiply each reduced column/row by */ |
| 170 | bool mul_by_scalar{ false }; /**< True if each column/row reduction has to be multiplied by a scalar value */ |
| 171 | }; |
Sang-Hoon Park | c261798 | 2020-05-20 22:13:47 +0100 | [diff] [blame] | 172 | |
| 173 | struct ScaleKernelInfo |
| 174 | { |
| 175 | /** Constructor |
| 176 | * |
| 177 | * @param[in] interpolation_policy Interpolation type to use |
| 178 | * @param[in] border_mode Border mode policy |
| 179 | * @param[in] constant_border_value (Optional) Constant value to use for borders if border_mode is set to CONSTANT and use_padding is set to false. Defaults to default @ref PixelValue |
| 180 | * @param[in] sampling_policy (Optional) Sampling policy used by the interpolation. Defaults to @ref SamplingPolicy::CENTER |
| 181 | * @param[in] use_padding (Optional) Is padding in use or not. Defaults to true. |
| 182 | * @param[in] align_corners (Optional) Align corners of input and output, only affecting bilinear policy with TOP_LEFT sampling policy. Defaults to false. |
| 183 | */ |
| 184 | ScaleKernelInfo(InterpolationPolicy interpolation_policy, |
| 185 | BorderMode border_mode, |
| 186 | PixelValue constant_border_value = PixelValue(), |
| 187 | SamplingPolicy sampling_policy = SamplingPolicy::CENTER, |
| 188 | bool use_padding = true, |
| 189 | bool align_corners = false) |
| 190 | : interpolation_policy{ interpolation_policy }, |
| 191 | border_mode{ border_mode }, |
| 192 | constant_border_value{ constant_border_value }, |
| 193 | sampling_policy{ sampling_policy }, |
| 194 | use_padding{ use_padding }, |
| 195 | align_corners{ align_corners } |
| 196 | { |
| 197 | } |
| 198 | |
| 199 | InterpolationPolicy interpolation_policy; /**< Interpolation type to use */ |
| 200 | BorderMode border_mode; /**< Border mode policy */ |
| 201 | PixelValue constant_border_value; /**< Constant value to use for constant border mode policy */ |
| 202 | SamplingPolicy sampling_policy; /**< Sampling policy used by the interpolation. */ |
| 203 | bool use_padding; /**< Indication of using padding */ |
| 204 | bool align_corners; /**< Align corners of input and output */ |
| 205 | }; |
Georgios Pinitas | 25ef721 | 2020-06-02 23:00:41 +0100 | [diff] [blame] | 206 | |
| 207 | struct ThresholdKernelInfo |
| 208 | { |
| 209 | /** Default constructor */ |
| 210 | ThresholdKernelInfo() = default; |
| 211 | /** Constructor |
| 212 | * |
| 213 | * @param[in] threshold Threshold. When the threshold type is RANGE, this is used as the lower threshold. |
| 214 | * @param[in] false_value value to set when the condition is not respected. |
| 215 | * @param[in] true_value value to set when the condition is respected. |
| 216 | * @param[in] type Thresholding type. Either RANGE or BINARY. |
| 217 | * @param[in] upper Upper threshold. Only used when the thresholding type is RANGE. |
| 218 | */ |
| 219 | ThresholdKernelInfo(uint8_t threshold, uint8_t false_value, uint8_t true_value, ThresholdType type, uint8_t upper) |
| 220 | : threshold(threshold), false_value(false_value), true_value(true_value), type(type), upper(upper) |
| 221 | { |
| 222 | } |
| 223 | |
| 224 | uint8_t threshold{ 0 }; |
| 225 | uint8_t false_value{ 0 }; |
| 226 | uint8_t true_value{ 0 }; |
| 227 | ThresholdType type{ ThresholdType::BINARY }; |
| 228 | uint8_t upper{ 0 }; |
| 229 | }; |
Georgios Pinitas | 0bc7849 | 2019-03-18 20:07:37 +0000 | [diff] [blame] | 230 | } // namespace arm_compute |
Michalis Spyrou | f464337 | 2019-11-29 16:17:13 +0000 | [diff] [blame] | 231 | #endif /* ARM_COMPUTE_CORE_KERNEL_DESCRIPTORS_H */ |