Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2016, 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/CLHistogramKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 27 | #include "arm_compute/core/CL/ICLDistribution1D.h" |
| 28 | #include "arm_compute/core/CL/ICLTensor.h" |
| 29 | #include "arm_compute/core/CL/OpenCL.h" |
| 30 | #include "arm_compute/core/Error.h" |
| 31 | #include "arm_compute/core/Helpers.h" |
| 32 | #include "arm_compute/core/TensorInfo.h" |
| 33 | #include "arm_compute/core/Types.h" |
| 34 | #include "arm_compute/core/Validate.h" |
| 35 | #include "arm_compute/core/Window.h" |
| 36 | |
| 37 | #include <cstring> |
| 38 | #include <string> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | // each thread handle 16 pixels |
| 43 | constexpr signed int pixels_per_item = 16; |
| 44 | |
| 45 | // local work group size in X dimension |
| 46 | constexpr unsigned int local_x_size = 16; |
| 47 | |
| 48 | CLHistogramKernel::CLHistogramKernel() |
| 49 | : _input(nullptr), _output(nullptr) |
| 50 | { |
| 51 | } |
| 52 | |
| 53 | void CLHistogramKernel::configure(const ICLImage *input, ICLDistribution1D *output) |
| 54 | { |
| 55 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); |
| 56 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 57 | |
| 58 | // Check input size |
| 59 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 60 | |
| 61 | // Check offset |
| 62 | ARM_COMPUTE_ERROR_ON_MSG(0 > output->offset() || output->offset() > 256, "Offset is larger than the image value range."); |
| 63 | |
| 64 | // Check range |
| 65 | ARM_COMPUTE_ERROR_ON_MSG(output->range() > 256 /* max range */, "Range larger than the image value range."); |
| 66 | |
| 67 | _input = input; |
| 68 | _output = output; |
| 69 | |
| 70 | if(_input->info()->dimension(0) < pixels_per_item) |
| 71 | { |
| 72 | return; |
| 73 | } |
| 74 | |
| 75 | unsigned int num_bins = _output->num_bins(); |
| 76 | unsigned int window_size = _output->window(); |
| 77 | unsigned int offset = _output->offset(); |
| 78 | unsigned int range = _output->range(); |
| 79 | unsigned int offrange = offset + range; |
| 80 | unsigned int bin_size = _output->size(); |
| 81 | unsigned int buffer_size = bin_size + 1; // We need one extra place for pixels that don't meet the conditions |
| 82 | |
| 83 | // Create kernel |
| 84 | bool is_fixed_size = (256 == num_bins) && (1 == window_size) && (0 == offset) && (256 == offrange); |
| 85 | std::string kernel_name = is_fixed_size ? "hist_local_kernel_fixed" : "hist_local_kernel"; |
| 86 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name)); |
| 87 | |
| 88 | // Set static kernel arguments |
| 89 | unsigned int idx = num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 90 | _kernel.setArg(idx++, buffer_size, nullptr); |
| 91 | _kernel.setArg(idx++, _output->cl_buffer()); |
| 92 | if(!is_fixed_size) |
| 93 | { |
| 94 | _kernel.setArg<cl_uint>(idx++, num_bins); |
| 95 | _kernel.setArg<cl_uint>(idx++, offset); |
| 96 | _kernel.setArg<cl_uint>(idx++, range); |
| 97 | _kernel.setArg<cl_uint>(idx++, offrange); |
| 98 | } |
| 99 | |
| 100 | // We only run histogram on Image, therefore only 2 dimensions here |
| 101 | unsigned int end_position = (_input->info()->dimension(0) / pixels_per_item) * pixels_per_item; |
| 102 | |
| 103 | // Configure kernel window |
| 104 | Window win; |
| 105 | win.set(0, Window::Dimension(0, end_position, pixels_per_item)); |
| 106 | win.set(1, Window::Dimension(0, _input->info()->dimension(1))); |
| 107 | |
| 108 | update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, pixels_per_item)); |
| 109 | |
| 110 | ICLKernel::configure(win); |
| 111 | } |
| 112 | |
| 113 | void CLHistogramKernel::run(const Window &window, cl::CommandQueue &queue) |
| 114 | { |
| 115 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 116 | ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window); |
| 117 | |
Isabella Gottardi | 02aabcc | 2017-10-12 17:28:51 +0100 | [diff] [blame] | 118 | // TODO (COMPMID-679): Add CLMemFill |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 119 | _output->map(queue, true); |
| 120 | ARM_COMPUTE_ERROR_ON(_output->buffer() == nullptr); |
| 121 | memset(_output->buffer(), 0, _output->size()); |
| 122 | _output->unmap(queue); |
| 123 | |
Isabella Gottardi | 02aabcc | 2017-10-12 17:28:51 +0100 | [diff] [blame] | 124 | if(_input->info()->dimension(0) < pixels_per_item) |
| 125 | { |
| 126 | return; |
| 127 | } |
| 128 | |
| 129 | Window slice = window.first_slice_window_2D(); |
| 130 | const unsigned int gws_x = (window.x().end() - window.x().start()) / window.x().step(); |
| 131 | cl::NDRange lws = (local_x_size < gws_x) ? cl::NDRange(local_x_size, 1) : cl::NDRange(1, 1); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 132 | |
| 133 | do |
| 134 | { |
| 135 | /* Run the core part which has width can be divided by 16 */ |
| 136 | unsigned int idx = 0; |
| 137 | add_2D_tensor_argument(idx, _input, slice); |
| 138 | |
| 139 | enqueue(queue, *this, slice, lws); |
| 140 | } |
| 141 | while(window.slide_window_slice_2D(slice)); |
| 142 | } |
| 143 | |
| 144 | CLHistogramBorderKernel::CLHistogramBorderKernel() |
| 145 | : _input(nullptr), _output(nullptr) |
| 146 | { |
| 147 | } |
| 148 | |
| 149 | void CLHistogramBorderKernel::configure(const ICLImage *input, ICLDistribution1D *output) |
| 150 | { |
| 151 | ARM_COMPUTE_ERROR_ON_TENSOR_NOT_2D(input); |
| 152 | ARM_COMPUTE_ERROR_ON(nullptr == output); |
| 153 | |
| 154 | // Check input size |
| 155 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 156 | |
| 157 | // Check offset |
| 158 | ARM_COMPUTE_ERROR_ON_MSG(0 > output->offset() || output->offset() > 256, "Offset is larger than the image value range."); |
| 159 | |
| 160 | // Check range |
| 161 | ARM_COMPUTE_ERROR_ON_MSG(output->range() > 256 /* max range */, "Range larger than the image value range."); |
| 162 | |
| 163 | // We only run histogram on Image, therefore only 2 dimensions here |
| 164 | unsigned int start_position = (input->info()->dimension(0) / pixels_per_item) * pixels_per_item; |
| 165 | |
| 166 | if(start_position >= input->info()->dimension(0)) |
| 167 | { |
| 168 | return; // no need to run histogram border kernel |
| 169 | } |
| 170 | |
| 171 | _input = input; |
| 172 | _output = output; |
| 173 | |
| 174 | unsigned int num_bins = _output->num_bins(); |
| 175 | unsigned int window_size = _output->window(); |
| 176 | unsigned int offset = _output->offset(); |
| 177 | unsigned int range = _output->range(); |
| 178 | unsigned int offrange = offset + range; |
| 179 | |
| 180 | // Create kernel |
| 181 | bool is_fixed_size = (256 == num_bins) && (1 == window_size) && (0 == offset) && (256 == offrange); |
| 182 | std::string kernel_name = is_fixed_size ? "hist_border_kernel_fixed" : "hist_border_kernel"; |
| 183 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel(kernel_name)); |
| 184 | |
| 185 | // Set static kernel arguments |
| 186 | unsigned int idx = num_arguments_per_2D_tensor(); //Skip the input and output parameters |
| 187 | _kernel.setArg(idx++, _output->cl_buffer()); |
| 188 | if(!is_fixed_size) |
| 189 | { |
| 190 | _kernel.setArg<cl_uint>(idx++, num_bins); |
| 191 | _kernel.setArg<cl_uint>(idx++, offset); |
| 192 | _kernel.setArg<cl_uint>(idx++, range); |
| 193 | _kernel.setArg<cl_uint>(idx++, offrange); |
| 194 | } |
| 195 | |
| 196 | // Configure kernel window |
| 197 | Window win; |
| 198 | win.set(0, Window::Dimension(start_position, _input->info()->dimension(0))); |
| 199 | win.set(1, Window::Dimension(0, _input->info()->dimension(1))); |
| 200 | update_window_and_padding(win, AccessWindowHorizontal(input->info(), 0, 1)); |
| 201 | ICLKernel::configure(win); |
| 202 | } |
| 203 | |
| 204 | void CLHistogramBorderKernel::run(const Window &window, cl::CommandQueue &queue) |
| 205 | { |
| 206 | if(window.x().start() >= window.x().end()) |
| 207 | { |
| 208 | return; |
| 209 | } |
| 210 | |
| 211 | ARM_COMPUTE_ERROR_ON_MISMATCHING_WINDOWS(ICLKernel::window(), window); |
| 212 | |
| 213 | cl::NDRange lws = cl::NDRange(1, 1); |
| 214 | |
| 215 | Window slice = window.first_slice_window_2D(); |
| 216 | |
| 217 | do |
| 218 | { |
| 219 | /* Run the border part which has width cannot be divided by 16 */ |
| 220 | unsigned int idx = 0; |
| 221 | add_2D_tensor_argument(idx, _input, slice); |
| 222 | |
| 223 | enqueue(queue, *this, slice, lws); |
| 224 | } |
| 225 | while(window.slide_window_slice_2D(slice)); |
| 226 | } |