giuros01 | 3175fcf | 2018-11-21 09:59:17 +0000 | [diff] [blame] | 1 | /* |
Manuel Bottini | 8481d83 | 2019-12-10 15:28:40 +0000 | [diff] [blame^] | 2 | * Copyright (c) 2018-2020 ARM Limited. |
giuros01 | 3175fcf | 2018-11-21 09:59:17 +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/CLTileKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 28 | #include "arm_compute/core/CL/CLValidate.h" |
| 29 | #include "arm_compute/core/CL/ICLTensor.h" |
| 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/IAccessWindow.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 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 37 | |
| 38 | namespace arm_compute |
| 39 | { |
| 40 | namespace |
| 41 | { |
| 42 | Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const Multiples &multiples) |
| 43 | { |
| 44 | ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); |
Manuel Bottini | 8481d83 | 2019-12-10 15:28:40 +0000 | [diff] [blame^] | 45 | ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); |
giuros01 | 3175fcf | 2018-11-21 09:59:17 +0000 | [diff] [blame] | 46 | ARM_COMPUTE_RETURN_ERROR_ON(multiples.size() > 4); |
| 47 | ARM_COMPUTE_RETURN_ERROR_ON(multiples.empty()); |
| 48 | ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(multiples.begin(), multiples.end(), [](uint32_t e) |
| 49 | { |
| 50 | return e == 0; |
| 51 | })); |
| 52 | |
| 53 | // Validate output if initialized |
| 54 | if(output->total_size() != 0) |
| 55 | { |
| 56 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_tiled_shape(input->tensor_shape(), multiples), output->tensor_shape()); |
| 57 | ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 58 | } |
| 59 | |
| 60 | return Status{}; |
| 61 | } |
| 62 | } // namespace |
| 63 | |
| 64 | CLTileKernel::CLTileKernel() |
| 65 | : _input(nullptr), _output(nullptr) |
| 66 | { |
| 67 | } |
| 68 | |
| 69 | void CLTileKernel::configure(const ICLTensor *input, ICLTensor *output, const Multiples &multiples) |
| 70 | { |
| 71 | ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); |
| 72 | |
| 73 | // Auto initialize output |
| 74 | TensorShape tiled_shape = misc::shape_calculator::compute_tiled_shape(input->info()->tensor_shape(), multiples); |
| 75 | auto_init_if_empty(*output->info(), tiled_shape, 1, input->info()->data_type()); |
| 76 | |
| 77 | // Validate |
| 78 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), multiples)); |
| 79 | |
| 80 | _input = input; |
| 81 | _output = output; |
| 82 | |
| 83 | const DataType data_type = input->info()->data_type(); |
| 84 | const int vec_size_x = 16 / input->info()->element_size(); |
| 85 | const int input_width_x = input->info()->tensor_shape().x(); |
| 86 | const unsigned int offset = ceil_to_multiple(input_width_x, vec_size_x) - input_width_x; |
| 87 | const bool multi_access_x = (input_width_x / vec_size_x > 0); |
| 88 | |
| 89 | // Create kernel |
| 90 | CLBuildOptions build_opts; |
| 91 | build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); |
| 92 | build_opts.add_option("-DSRC_WIDTH=" + support::cpp11::to_string(input_width_x)); |
| 93 | build_opts.add_option("-DSRC_HEIGHT=" + support::cpp11::to_string(input->info()->dimension(1))); |
| 94 | build_opts.add_option("-DSRC_DEPTH=" + support::cpp11::to_string(input->info()->dimension(2))); |
| 95 | build_opts.add_option("-DSRC_BATCHES=" + support::cpp11::to_string(input->info()->dimension(3))); |
| 96 | build_opts.add_option("-DDST_DEPTH=" + support::cpp11::to_string(output->info()->dimension(2))); |
| 97 | build_opts.add_option_if(multi_access_x, "-DOFFSET=" + support::cpp11::to_string(offset)); |
| 98 | build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); |
| 99 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("tile", build_opts.options())); |
| 100 | |
| 101 | // Configure window without padding |
| 102 | Window win = calculate_max_window(*output->info()); |
| 103 | |
| 104 | if(multi_access_x) |
| 105 | { |
| 106 | // If multi-access is enabled, no thread should cross the tile boundaries. This means we need |
| 107 | // as many threads as those to cover a single tile times multiples[0]. Note that if threads |
| 108 | // do not cross the boundaries of the tiles, they won't cross the boundaries of the last tile, and |
| 109 | // we don't need to pad the output |
| 110 | const unsigned int size_win_x = ceil_to_multiple(input->info()->dimension(0), vec_size_x) * multiples[0]; |
| 111 | win.set(Window::DimX, |
| 112 | Window::Dimension(win.x().start(), size_win_x, vec_size_x)); |
| 113 | } |
| 114 | |
| 115 | ICLKernel::configure_internal(win); |
| 116 | |
| 117 | // Set config_id for enabling LWS tuning |
| 118 | _config_id = "tile"; |
| 119 | _config_id += "_"; |
| 120 | _config_id += lower_string(string_from_data_type(input->info()->data_type())); |
| 121 | for(unsigned int i = 0; i < multiples.size(); ++i) |
| 122 | { |
| 123 | _config_id += "_"; |
| 124 | _config_id += support::cpp11::to_string(input->info()->dimension(i)); |
| 125 | _config_id += "_"; |
| 126 | _config_id += support::cpp11::to_string(multiples[i]); |
| 127 | } |
| 128 | } |
| 129 | |
| 130 | Status CLTileKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const Multiples &multiples) |
| 131 | { |
| 132 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, multiples)); |
| 133 | return Status{}; |
| 134 | } |
| 135 | |
| 136 | void CLTileKernel::run(const Window &window, cl::CommandQueue &queue) |
| 137 | { |
| 138 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 139 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 140 | |
| 141 | Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); |
| 142 | Window slice = collapsed.first_slice_window_4D(); |
| 143 | |
| 144 | do |
| 145 | { |
| 146 | unsigned int idx = 0; |
| 147 | add_4D_tensor_argument(idx, _input, slice); |
| 148 | add_4D_tensor_argument(idx, _output, slice); |
Georgios Pinitas | 275f99c | 2019-08-23 12:44:11 +0100 | [diff] [blame] | 149 | enqueue(queue, *this, slice, lws_hint()); |
giuros01 | 3175fcf | 2018-11-21 09:59:17 +0000 | [diff] [blame] | 150 | } |
| 151 | while(collapsed.slide_window_slice_4D(slice)); |
| 152 | } |
| 153 | } // namespace arm_compute |