Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 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 | #include "arm_compute/core/CL/kernels/CLSoftmaxLayerKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/CL/CLHelpers.h" |
| 28 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 29 | #include "arm_compute/core/CL/ICLTensor.h" |
| 30 | #include "arm_compute/core/CL/OpenCL.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Utils.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
| 36 | |
| 37 | #include <set> |
| 38 | #include <string> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | void CLLogits1DMaxKernel::configure(const ICLTensor *input, ICLTensor *output) |
| 43 | { |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 44 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 45 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 46 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 47 | |
| 48 | _input = input; |
| 49 | _output = output; |
| 50 | |
| 51 | // The kernel loops over all elements in steps of 16 |
| 52 | const unsigned int num_elems_processed_per_iteration = ceil_to_multiple(input->info()->dimension(0), 16); |
| 53 | |
| 54 | // Set build options |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 55 | std::set<std::string> build_opts; |
| 56 | build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); |
| 57 | if(is_data_type_fixed_point(input->info()->data_type())) |
| 58 | { |
| 59 | build_opts.emplace(("-DFIXED_POINT_POSITION=" + val_to_string(input->info()->fixed_point_position()))); |
| 60 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 61 | |
| 62 | // Tell the kernel that the width is not a multiple of 16 |
| 63 | if((input->info()->dimension(0) % max_cl_vector_width) != 0) |
| 64 | { |
| 65 | build_opts.emplace("-DNON_MULTIPLE_OF_16"); |
| 66 | } |
| 67 | |
| 68 | // Create kernel |
| 69 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("softmax_layer_max", build_opts)); |
| 70 | |
| 71 | // Set fixed arguments |
| 72 | unsigned int idx = 2 * num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 73 | _kernel.setArg<cl_uint>(idx++, input->info()->dimension(0)); |
| 74 | |
| 75 | // Configure kernel window |
| 76 | constexpr unsigned int num_elems_written_per_iteration = 1; |
| 77 | |
| 78 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 79 | AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); |
| 80 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); |
| 81 | |
| 82 | update_window_and_padding(win, input_access, output_access); |
| 83 | |
| 84 | output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); |
| 85 | |
| 86 | ICLKernel::configure(win); |
| 87 | } |
| 88 | |
| 89 | CLLogits1DShiftExpSumKernel::CLLogits1DShiftExpSumKernel() |
| 90 | : _input(nullptr), _max(nullptr), _output(nullptr), _sum(nullptr) |
| 91 | { |
| 92 | } |
| 93 | |
| 94 | void CLLogits1DShiftExpSumKernel::configure(const ICLTensor *input, const ICLTensor *max, ICLTensor *output, ICLTensor *sum) |
| 95 | { |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 96 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 97 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, max, sum); |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 98 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 99 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output, max, sum); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 100 | |
| 101 | _input = input; |
| 102 | _max = max; |
| 103 | _output = output; |
| 104 | _sum = sum; |
| 105 | |
| 106 | // The kernel loops over all elements in steps of 16 |
| 107 | const unsigned int num_elems_processed_per_iteration = ceil_to_multiple(input->info()->dimension(0), 16); |
| 108 | |
| 109 | // Set build options |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 110 | std::set<std::string> build_opts; |
| 111 | build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); |
| 112 | if(is_data_type_fixed_point(input->info()->data_type())) |
| 113 | { |
| 114 | build_opts.emplace(("-DFIXED_POINT_POSITION=" + val_to_string(input->info()->fixed_point_position()))); |
| 115 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 116 | |
| 117 | // Tell the kernel that the width is not a multiple of 16 |
| 118 | if((input->info()->dimension(0) % max_cl_vector_width) != 0) |
| 119 | { |
| 120 | build_opts.emplace("-DNON_MULTIPLE_OF_16"); |
| 121 | } |
| 122 | |
| 123 | // Create kernel |
| 124 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("softmax_layer_shift_exp_sum", build_opts)); |
| 125 | |
| 126 | // Set fixed arguments |
| 127 | unsigned int idx = 4 * num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 128 | _kernel.setArg<cl_uint>(idx++, input->info()->dimension(0)); |
| 129 | |
| 130 | // Configure window |
| 131 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 132 | |
| 133 | AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); |
| 134 | AccessWindowHorizontal max_access(max->info(), 0, 1); |
| 135 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); |
| 136 | AccessWindowHorizontal sum_access(sum->info(), 0, 1); |
| 137 | |
| 138 | update_window_and_padding(win, input_access, max_access, output_access, sum_access); |
| 139 | |
| 140 | output_access.set_valid_region(win, input->info()->valid_region()); |
| 141 | sum_access.set_valid_region(win, ValidRegion(Coordinates(), sum->info()->tensor_shape())); |
| 142 | |
| 143 | ICLKernel::configure(win); |
| 144 | } |
| 145 | |
| 146 | void CLLogits1DShiftExpSumKernel::run(const Window &window, cl::CommandQueue &queue) |
| 147 | { |
| 148 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 149 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 150 | |
| 151 | Window slice = window.first_slice_window_2D(); |
| 152 | |
| 153 | do |
| 154 | { |
| 155 | unsigned int idx = 0; |
| 156 | // Set inputs |
| 157 | add_2D_tensor_argument(idx, _input, slice); |
| 158 | add_2D_tensor_argument(idx, _max, slice); |
| 159 | add_2D_tensor_argument(idx, _output, slice); |
| 160 | add_2D_tensor_argument(idx, _sum, slice); |
| 161 | enqueue(queue, *this, slice); |
| 162 | } |
| 163 | while(window.slide_window_slice_2D(slice)); |
| 164 | } |
| 165 | |
| 166 | CLLogits1DNormKernel::CLLogits1DNormKernel() |
| 167 | : _input(nullptr), _sum(nullptr), _output(nullptr) |
| 168 | { |
| 169 | } |
| 170 | |
| 171 | void CLLogits1DNormKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output) |
| 172 | { |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 173 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QS8, DataType::F16, DataType::F32); |
| 174 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum, output); |
| 175 | ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); |
| 176 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, sum, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 177 | |
| 178 | _input = input; |
| 179 | _sum = sum; |
| 180 | _output = output; |
| 181 | |
| 182 | // Set build options |
| 183 | std::set<std::string> build_opts; |
Georgios Pinitas | e5f8fd6 | 2017-06-23 18:03:44 +0100 | [diff] [blame] | 184 | build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); |
| 185 | if(is_data_type_fixed_point(input->info()->data_type())) |
| 186 | { |
| 187 | build_opts.emplace(("-DFIXED_POINT_POSITION=" + val_to_string(input->info()->fixed_point_position()))); |
| 188 | } |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 189 | |
| 190 | // Create kernel |
| 191 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("softmax_layer_norm", build_opts)); |
| 192 | |
| 193 | // Configure window |
| 194 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 195 | |
| 196 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); |
| 197 | |
| 198 | AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); |
| 199 | AccessWindowStatic sum_access(sum->info(), 0, 0, 1, sum->info()->dimension(1)); |
| 200 | AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); |
| 201 | |
| 202 | update_window_and_padding(win, input_access, sum_access, output_access); |
| 203 | |
| 204 | output_access.set_valid_region(win, input->info()->valid_region()); |
| 205 | |
| 206 | ICLKernel::configure(win); |
| 207 | } |
| 208 | |
| 209 | void CLLogits1DNormKernel::run(const Window &window, cl::CommandQueue &queue) |
| 210 | { |
| 211 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 212 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); |
| 213 | |
| 214 | Window slice = window.first_slice_window_2D(); |
| 215 | |
| 216 | do |
| 217 | { |
| 218 | Window sum_slice = slice; |
| 219 | sum_slice.set(Window::DimX, Window::Dimension(0, 1, 1)); |
| 220 | |
| 221 | unsigned int idx = 0; |
| 222 | // Set inputs |
| 223 | add_2D_tensor_argument(idx, _input, slice); |
| 224 | add_2D_tensor_argument(idx, _sum, sum_slice); |
| 225 | add_2D_tensor_argument(idx, _output, slice); |
| 226 | enqueue(queue, *this, slice); |
| 227 | } |
| 228 | while(window.slide_window_slice_2D(slice)); |
| 229 | } |