Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [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/CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/AccessWindowStatic.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/Helpers.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/Validate.h" |
| 32 | #include "arm_compute/core/Window.h" |
| 33 | #include "support/ToolchainSupport.h" |
| 34 | |
| 35 | using namespace arm_compute; |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | class Coordinates; |
| 40 | } // namespace arm_compute |
| 41 | |
| 42 | CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel() |
| 43 | : _input(nullptr), _bias(nullptr), _output(nullptr) |
| 44 | { |
| 45 | } |
| 46 | |
| 47 | void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, int result_offset, int result_mult_int, int result_shift, int min, |
| 48 | int max) |
| 49 | { |
| 50 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32); |
| 51 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8); |
| 52 | ARM_COMPUTE_ERROR_ON(max > 255); |
| 53 | ARM_COMPUTE_ERROR_ON(min < 0 || min > max); |
| 54 | |
| 55 | if(bias != nullptr) |
| 56 | { |
| 57 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias); |
| 58 | ARM_COMPUTE_ERROR_ON(bias->info()->num_dimensions() > 1); |
| 59 | ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != bias->info()->dimension(0)); |
| 60 | } |
| 61 | |
| 62 | _input = input; |
| 63 | _bias = bias; |
| 64 | _output = output; |
| 65 | |
| 66 | // Set the arguments to pass at compile time |
| 67 | CLBuildOptions build_opts; |
| 68 | build_opts.add_option("-DRESULT_OFFSET=" + support::cpp11::to_string(result_offset)); |
| 69 | build_opts.add_option("-DRESULT_MULT_INT=" + support::cpp11::to_string(result_mult_int)); |
| 70 | build_opts.add_option("-DRESULT_SHIFT=" + support::cpp11::to_string(result_shift)); |
| 71 | build_opts.add_option_if((min != 0) && (min != max), "-DMIN_BOUND=" + support::cpp11::to_string(min)); |
| 72 | build_opts.add_option_if((max != 255) && (min != max), "-DMAX_BOUND=" + support::cpp11::to_string(max)); |
| 73 | build_opts.add_option_if(bias != nullptr, "-DADD_BIAS"); |
| 74 | |
| 75 | // Create kernel |
| 76 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_output_stage_quantize_down", build_opts.options())); |
| 77 | |
| 78 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 79 | |
| 80 | // Configure kernel window |
| 81 | Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); |
| 82 | |
| 83 | AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); |
| 84 | AccessWindowHorizontal output_result_access(output->info(), 0, num_elems_processed_per_iteration); |
| 85 | |
| 86 | update_window_and_padding(win, |
| 87 | input_access, |
| 88 | output_result_access); |
| 89 | |
| 90 | if(bias != nullptr) |
| 91 | { |
| 92 | AccessWindowStatic bias_access(bias->info(), 0, 0, ceil_to_multiple(bias->info()->dimension(0), num_elems_processed_per_iteration), bias->info()->tensor_shape()[1]); |
| 93 | |
| 94 | update_window_and_padding(win, |
| 95 | bias_access); |
| 96 | } |
| 97 | |
| 98 | output_result_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); |
| 99 | |
| 100 | ICLKernel::configure(win); |
| 101 | } |
| 102 | |
| 103 | void CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel::run(const Window &window, cl::CommandQueue &queue) |
| 104 | { |
| 105 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 106 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 107 | |
| 108 | Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); |
| 109 | Window slice = collapsed.first_slice_window_3D(); |
| 110 | |
| 111 | unsigned int idx1 = num_arguments_per_3D_tensor(); |
| 112 | if(_bias != nullptr) |
| 113 | { |
| 114 | Window biases_slice(slice); |
| 115 | biases_slice.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 116 | biases_slice.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 117 | add_1D_tensor_argument(idx1, _bias, biases_slice); |
| 118 | } |
| 119 | |
| 120 | do |
| 121 | { |
| 122 | unsigned int idx = 0; |
| 123 | add_3D_tensor_argument(idx, _input, slice); |
| 124 | add_3D_tensor_argument(idx1, _output, slice); |
| 125 | enqueue(queue, *this, slice); |
| 126 | } |
| 127 | while(collapsed.slide_window_slice_3D(slice)); |
| 128 | } |