Michele Di Giorgio | 27400b9 | 2018-11-01 13:44:05 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 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/CLWidthConcatenate2TensorsKernel.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" |
| 38 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 39 | |
| 40 | #include "support/ToolchainSupport.h" |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | namespace |
| 45 | { |
| 46 | constexpr unsigned int num_elems_processed_per_iteration = 8; |
| 47 | |
| 48 | std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input1, ITensorInfo *input2, ITensorInfo *output) |
| 49 | { |
| 50 | // The window needs to be based on the output |
| 51 | Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); |
| 52 | AccessWindowStatic input1_access(input1, 0, 0, ceil_to_multiple(input1->dimension(0), num_elems_processed_per_iteration) + num_elems_processed_per_iteration, input1->dimension(1)); |
| 53 | AccessWindowStatic input2_access(input2, -num_elems_processed_per_iteration, 0, ceil_to_multiple(input2->dimension(0), num_elems_processed_per_iteration) + num_elems_processed_per_iteration, |
| 54 | input2->dimension(1)); |
| 55 | AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); |
| 56 | bool window_changed = update_window_and_padding(win, input1_access, input2_access, output_access); |
| 57 | |
| 58 | Window win_collapsed = win.collapse(win, Window::DimZ); |
| 59 | |
| 60 | Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; |
| 61 | return std::make_pair(err, win_collapsed); |
| 62 | } |
| 63 | Status validate_arguments(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) |
| 64 | { |
| 65 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input1, input2, output); |
| 66 | ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input1); |
| 67 | 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, |
| 68 | DataType::F32); |
| 69 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input1, input2, output); |
| 70 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(0) + input2->dimension(0) > output->dimension(0)); |
| 71 | |
| 72 | for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) |
| 73 | { |
| 74 | ARM_COMPUTE_RETURN_ERROR_ON(input1->dimension(i) != output->dimension(i)); |
| 75 | ARM_COMPUTE_RETURN_ERROR_ON(input2->dimension(i) != output->dimension(i)); |
| 76 | } |
| 77 | ARM_COMPUTE_RETURN_ERROR_ON(input1->num_dimensions() > 4); |
| 78 | |
| 79 | return Status{}; |
| 80 | } |
| 81 | } // namespace |
| 82 | |
| 83 | CLWidthConcatenate2TensorsKernel::CLWidthConcatenate2TensorsKernel() |
| 84 | : _input1(nullptr), _input2(nullptr), _output(nullptr) |
| 85 | { |
| 86 | } |
| 87 | |
| 88 | Status CLWidthConcatenate2TensorsKernel::validate(const ITensorInfo *input1, const ITensorInfo *input2, const ITensorInfo *output) |
| 89 | { |
| 90 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input1, input2, output)); |
| 91 | ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input1->clone().get(), input2->clone().get(), output->clone().get()).first); |
| 92 | return Status{}; |
| 93 | } |
| 94 | |
| 95 | void CLWidthConcatenate2TensorsKernel::configure(const ICLTensor *input1, const ICLTensor *input2, ICLTensor *output) |
| 96 | { |
| 97 | ARM_COMPUTE_ERROR_ON_NULLPTR(input1, input2, output); |
| 98 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input1->info(), input2->info(), output->info())); |
| 99 | |
| 100 | _input1 = input1; |
| 101 | _input2 = input2; |
| 102 | _output = output; |
| 103 | |
| 104 | // Add build options |
| 105 | CLBuildOptions build_opts; |
| 106 | build_opts.add_option("-DDATA_TYPE=" + get_underlying_cl_type_from_data_type(input1->info()->data_type())); |
| 107 | build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); |
| 108 | build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input1->info()->dimension(2))); |
| 109 | build_opts.add_option("-DINPUT1_WIDTH=" + support::cpp11::to_string(input1->info()->dimension(0))); |
| 110 | build_opts.add_option("-DELEMENT_SIZE=" + support::cpp11::to_string(input1->info()->element_size())); |
| 111 | |
| 112 | // Create kernel |
| 113 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_width_x2", build_opts.options())); |
| 114 | |
| 115 | // Configure kernel window |
| 116 | auto win_config = validate_and_configure_window(input1->info(), input2->info(), output->info()); |
| 117 | ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); |
| 118 | |
| 119 | ICLKernel::configure_internal(std::get<1>(win_config)); |
| 120 | |
| 121 | // Set config_id for enabling LWS tuning |
| 122 | _config_id = "concatenate_width_x2_"; |
| 123 | _config_id += lower_string(string_from_data_type(input1->info()->data_type())); |
| 124 | _config_id += "_"; |
| 125 | _config_id += support::cpp11::to_string(input1->info()->dimension(0)); |
| 126 | _config_id += "_"; |
| 127 | _config_id += support::cpp11::to_string(input1->info()->dimension(1)); |
| 128 | _config_id += "_"; |
| 129 | _config_id += support::cpp11::to_string(input2->info()->dimension(0)); |
| 130 | _config_id += "_"; |
| 131 | _config_id += support::cpp11::to_string(input2->info()->dimension(1)); |
| 132 | } |
| 133 | |
| 134 | void CLWidthConcatenate2TensorsKernel::run(const Window &window, cl::CommandQueue &queue) |
| 135 | { |
| 136 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 137 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 138 | |
| 139 | Window slice = window.first_slice_window_4D(); |
| 140 | |
| 141 | do |
| 142 | { |
| 143 | unsigned int idx = 0; |
| 144 | add_4D_tensor_argument(idx, _input1, slice); |
| 145 | add_4D_tensor_argument(idx, _input2, slice); |
| 146 | add_4D_tensor_argument(idx, _output, slice); |
| 147 | enqueue(queue, *this, window, lws_hint()); |
| 148 | } |
| 149 | while(window.slide_window_slice_4D(slice)); |
| 150 | } |
| 151 | } // namespace arm_compute |