Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | 8e150a1 | 2018-12-21 15:20:56 +0000 | [diff] [blame] | 2 | * Copyright (c) 2018-2019 ARM Limited. |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +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 | */ |
| 24 | #include "arm_compute/core/CL/kernels/CLWidthConcatenate4TensorsKernel.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/CLValidate.h" |
| 30 | #include "arm_compute/core/CL/ICLTensor.h" |
| 31 | #include "arm_compute/core/CL/OpenCL.h" |
| 32 | #include "arm_compute/core/Error.h" |
| 33 | #include "arm_compute/core/Helpers.h" |
| 34 | #include "arm_compute/core/IAccessWindow.h" |
| 35 | #include "arm_compute/core/TensorInfo.h" |
| 36 | #include "arm_compute/core/Utils.h" |
| 37 | #include "arm_compute/core/Window.h" |
Georgios Pinitas | 6631ac2 | 2019-04-17 12:12:56 +0100 | [diff] [blame] | 38 | #include "arm_compute/core/utils/helpers/tensor_info.h" |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 39 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 40 | |
| 41 | #include "support/ToolchainSupport.h" |
| 42 | |
| 43 | namespace arm_compute |
| 44 | { |
| 45 | namespace |
| 46 | { |
| 47 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 48 | |
| 49 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *input3, ITensorInfo *input4, ITensorInfo *output) |
| 50 | { |
Michele Di Giorgio | 8e150a1 | 2018-12-21 15:20:56 +0000 | [diff] [blame] | 51 | const unsigned int input1_width = input1->dimension(0); |
| 52 | const unsigned int input2_width = input2->dimension(0); |
| 53 | const unsigned int input3_width = input3->dimension(0); |
| 54 | const unsigned int input4_width = input4->dimension(0); |
| 55 | |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 56 | // The window needs to be based on the output |
| 57 | Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); |
Michele Di Giorgio | 8e150a1 | 2018-12-21 15:20:56 +0000 | [diff] [blame] | 58 | AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1_width, num_elems_processed_per_iteration), input1->dimension(1)); |
| 59 | |
| 60 | const unsigned int input2_left_padding = input1_width % num_elems_processed_per_iteration; |
| 61 | const unsigned int input2_right_padding = ((input1_width + input2_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width + num_elems_processed_per_iteration - |
| 62 | input2_width; |
| 63 | AccessWindowStatic input2_access(input2, -input2_left_padding, 0, input2_width + input2_right_padding, input2->dimension(1)); |
| 64 | |
| 65 | const unsigned int input3_left_padding = (input1_width + input2_width) % num_elems_processed_per_iteration; |
| 66 | const unsigned int input3_right_padding = ((input1_width + input2_width + input3_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width - input2_width + |
| 67 | num_elems_processed_per_iteration - input3_width; |
| 68 | AccessWindowStatic input3_access(input3, -input3_left_padding, 0, input3_width + input3_right_padding, input3->dimension(1)); |
| 69 | |
| 70 | const unsigned int input4_left_padding = (input1_width + input2_width + input3_width) % num_elems_processed_per_iteration; |
| 71 | const unsigned int input4_right_padding = (output->dimension(0) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration + num_elems_processed_per_iteration - output->dimension(0); |
| 72 | AccessWindowStatic input4_access(input4, -input4_left_padding, 0, input4_width + input4_right_padding, input4->dimension(1)); |
| 73 | |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 74 | AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); |
| 75 | bool window_changed = update_window_and_padding(win, input1_access, input2_access, input3_access, input4_access, output_access); |
| 76 | |
| 77 | Window win_collapsed = win.collapse(win, Window::DimZ); |
| 78 | |
| 79 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 80 | return std::make_pair(err, win_collapsed); |
| 81 | } |
| 82 | Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *input3, const ITensorInfo *input4, const ITensorInfo *output) |
| 83 | { |
| 84 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, input3, input4, output); |
| 85 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); |
| 86 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input1, 1, DataType::U8, DataType::S8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::F16, DataType::U32, |
Manuel Bottini | 10c53f1 | 2019-07-17 16:11:53 +0100 | [diff] [blame] | 87 | DataType::S32, DataType::F32); |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 88 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2, input3, input4, output); |
| 89 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) + input2->dimension(0) + input3->dimension(0) + input4->dimension(0) > output->dimension(0)); |
| 90 | |
| 91 | for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) |
| 92 | { |
| 93 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(i) != output->dimension(i)); |
| 94 | ARM_COMPUTE_RETURN_ERROR_ON(input2->dimension(i) != output->dimension(i)); |
| 95 | ARM_COMPUTE_RETURN_ERROR_ON(input3->dimension(i) != output->dimension(i)); |
| 96 | ARM_COMPUTE_RETURN_ERROR_ON(input4->dimension(i) != output->dimension(i)); |
| 97 | } |
| 98 | ARM_COMPUTE_RETURN_ERROR_ON(input1->num_dimensions() > 4); |
| 99 | |
| 100 | return Status{}; |
| 101 | } |
| 102 | } // namespace |
| 103 | |
| 104 | CLWidthConcatenate4TensorsKernel::CLWidthConcatenate4TensorsKernel() |
| 105 | : _input1(nullptr), _input2(nullptr), _input3(nullptr), _input4(nullptr), _output(nullptr) |
| 106 | { |
| 107 | } |
| 108 | |
| 109 | Status CLWidthConcatenate4TensorsKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *input3, const ITensorInfo *input4, const ITensorInfo *output) |
| 110 | { |
| 111 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, input3, input4, output)); |
| 112 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), input3->clone().get(), input4->clone().get(), output->clone().get()).first); |
| 113 | return Status{}; |
| 114 | } |
| 115 | |
| 116 | void CLWidthConcatenate4TensorsKernel::configure(const ICLTensor *input1, const ICLTensor *input2, const ICLTensor *input3, const ICLTensor *input4, ICLTensor *output) |
| 117 | { |
| 118 | ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, input3, input4, output); |
| 119 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), input3->info(), input4->info(), output->info())); |
| 120 | |
| 121 | _input1 = input1; |
| 122 | _input2 = input2; |
| 123 | _input3 = input3; |
| 124 | _input4 = input4; |
| 125 | _output = output; |
| 126 | |
| 127 | // Add build options |
| 128 | CLBuildOptions build_opts; |
| 129 | build_opts.add_option("-DDATA_TYPE=" + get_underlying_cl_type_from_data_type(input1->info()->data_type())); |
| 130 | build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); |
| 131 | build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 132 | build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->info()->dimension(0))); |
| 133 | build_opts.add_option("-DINPUT2_WIDTH=" + support::cpp11::to_string(input2->info()->dimension(0))); |
| 134 | build_opts.add_option("-DINPUT3_WIDTH=" + support::cpp11::to_string(input3->info()->dimension(0))); |
| 135 | build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->info()->element_size())); |
| 136 | |
Georgios Pinitas | 6631ac2 | 2019-04-17 12:12:56 +0100 | [diff] [blame] | 137 | // If input have different quantization info set quantization parameters needed for the re-quantization process |
| 138 | const bool have_different_qinfo = helpers::tensor_info::tensors_have_different_quantization_info(output->info(), input1->info(), input2->info(), input3->info(), input4->info()); |
| 139 | if(is_data_type_quantized_asymmetric(input1->info()->data_type()) && have_different_qinfo) |
Pablo Tello | eb6c88a | 2019-02-07 15:53:19 +0000 | [diff] [blame] | 140 | { |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 141 | const UniformQuantizationInfo iq1_info = input1->info()->quantization_info().uniform(); |
| 142 | const UniformQuantizationInfo iq2_info = input2->info()->quantization_info().uniform(); |
| 143 | const UniformQuantizationInfo iq3_info = input3->info()->quantization_info().uniform(); |
| 144 | const UniformQuantizationInfo iq4_info = input4->info()->quantization_info().uniform(); |
| 145 | const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform(); |
| 146 | |
| 147 | build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq1_info.offset)); |
| 148 | build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq1_info.scale)); |
| 149 | build_opts.add_option("-DOFFSET_IN2=" + float_to_string_with_full_precision(iq2_info.offset)); |
| 150 | build_opts.add_option("-DSCALE_IN2=" + float_to_string_with_full_precision(iq2_info.scale)); |
| 151 | build_opts.add_option("-DOFFSET_IN3=" + float_to_string_with_full_precision(iq3_info.offset)); |
| 152 | build_opts.add_option("-DSCALE_IN3=" + float_to_string_with_full_precision(iq3_info.scale)); |
| 153 | build_opts.add_option("-DOFFSET_IN4=" + float_to_string_with_full_precision(iq4_info.offset)); |
| 154 | build_opts.add_option("-DSCALE_IN4=" + float_to_string_with_full_precision(iq4_info.scale)); |
| 155 | build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset)); |
| 156 | build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale)); |
Pablo Tello | eb6c88a | 2019-02-07 15:53:19 +0000 | [diff] [blame] | 157 | } |
| 158 | |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 159 | // Create kernel |
| 160 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_width_x4", build_opts.options())); |
| 161 | |
| 162 | // Configure kernel window |
| 163 | auto win_config = validate_and_configure_window(input1->info(), input2->info(), input3->info(), input4->info(), output->info()); |
| 164 | ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); |
| 165 | |
| 166 | ICLKernel::configure_internal(std::get<1>(win_config)); |
| 167 | |
Michele Di Giorgio | 8e150a1 | 2018-12-21 15:20:56 +0000 | [diff] [blame] | 168 | // Pass paddings as arguments to the kernel |
| 169 | const unsigned int input1_width = input1->info()->dimension(0); |
| 170 | const unsigned int input2_width = input2->info()->dimension(0); |
| 171 | const unsigned int input3_width = input3->info()->dimension(0); |
| 172 | |
| 173 | const unsigned int input1_right_padding = ceil_to_multiple(input1_width, num_elems_processed_per_iteration) - input1_width; |
| 174 | const unsigned int input2_left_padding = input1_width % num_elems_processed_per_iteration; |
| 175 | const unsigned int input2_right_padding = ((input1_width + input2_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width + num_elems_processed_per_iteration - |
| 176 | input2_width; |
| 177 | const unsigned int input3_left_padding = (input1_width + input2_width) % num_elems_processed_per_iteration; |
| 178 | const unsigned int input3_right_padding = ((input1_width + input2_width + input3_width) / num_elems_processed_per_iteration) * num_elems_processed_per_iteration - input1_width - input2_width + |
| 179 | num_elems_processed_per_iteration - input3_width; |
| 180 | const unsigned int input4_left_padding = (input1_width + input2_width + input3_width) % num_elems_processed_per_iteration; |
| 181 | unsigned int idx0 = 5 * num_arguments_per_4D_tensor(); |
| 182 | _kernel.setArg<cl_uint>(idx0++, input1_right_padding); |
| 183 | _kernel.setArg<cl_uint>(idx0++, input2_left_padding); |
| 184 | _kernel.setArg<cl_uint>(idx0++, input2_right_padding); |
| 185 | _kernel.setArg<cl_uint>(idx0++, input3_left_padding); |
| 186 | _kernel.setArg<cl_uint>(idx0++, input3_right_padding); |
| 187 | _kernel.setArg<cl_uint>(idx0++, input4_left_padding); |
| 188 | |
Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 189 | // Set config_id for enabling LWS tuning |
| 190 | _config_id = "concatenate_width_x4_"; |
| 191 | _config_id += lower_string(string_from_data_type(input1->info()->data_type())); |
| 192 | _config_id += "_"; |
| 193 | _config_id += support::cpp11::to_string(input1->info()->dimension(0)); |
| 194 | _config_id += "_"; |
| 195 | _config_id += support::cpp11::to_string(input1->info()->dimension(1)); |
| 196 | _config_id += "_"; |
| 197 | _config_id += support::cpp11::to_string(input2->info()->dimension(0)); |
| 198 | _config_id += "_"; |
| 199 | _config_id += support::cpp11::to_string(input2->info()->dimension(1)); |
| 200 | _config_id += "_"; |
| 201 | _config_id += support::cpp11::to_string(input3->info()->dimension(0)); |
| 202 | _config_id += "_"; |
| 203 | _config_id += support::cpp11::to_string(input3->info()->dimension(1)); |
| 204 | _config_id += "_"; |
| 205 | _config_id += support::cpp11::to_string(input4->info()->dimension(0)); |
| 206 | _config_id += "_"; |
| 207 | _config_id += support::cpp11::to_string(input4->info()->dimension(1)); |
| 208 | } |
| 209 | |
| 210 | void CLWidthConcatenate4TensorsKernel::run(const Window &window, cl::CommandQueue &queue) |
| 211 | { |
| 212 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 213 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 214 | |
| 215 | Window slice = window.first_slice_window_4D(); |
| 216 | |
| 217 | do |
| 218 | { |
| 219 | unsigned int idx = 0; |
| 220 | add_4D_tensor_argument(idx, _input1, slice); |
| 221 | add_4D_tensor_argument(idx, _input2, slice); |
| 222 | add_4D_tensor_argument(idx, _input3, slice); |
| 223 | add_4D_tensor_argument(idx, _input4, slice); |
| 224 | add_4D_tensor_argument(idx, _output, slice); |
| 225 | enqueue(queue, *this, window, lws_hint()); |
| 226 | } |
| 227 | while(window.slide_window_slice_4D(slice)); |
| 228 | } |
| 229 | } // namespace arm_compute |