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/CLGEMMLowpReductionKernel.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/TensorInfo.h" |
| 31 | #include "arm_compute/core/Types.h" |
| 32 | #include "arm_compute/core/Utils.h" |
| 33 | #include "arm_compute/core/Validate.h" |
| 34 | #include "arm_compute/core/Window.h" |
| 35 | #include "support/ToolchainSupport.h" |
| 36 | |
| 37 | #include <cstddef> |
| 38 | #include <cstdint> |
| 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
| 42 | namespace arm_compute |
| 43 | { |
| 44 | class Coordinates; |
| 45 | } // namespace arm_compute |
| 46 | |
| 47 | ICLGEMMLowpReductionKernel::ICLGEMMLowpReductionKernel() |
| 48 | : _input(), _output() |
| 49 | { |
| 50 | } |
| 51 | |
| 52 | void CLGEMMLowpMatrixAReductionKernel::configure(const ICLTensor *mtx_a, ICLTensor *vector_sum_row) |
| 53 | { |
| 54 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_a, 1, DataType::QASYMM8); |
| 55 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_row, 1, DataType::S32); |
| 56 | |
| 57 | _input = mtx_a; |
| 58 | _output = vector_sum_row; |
| 59 | |
| 60 | // Set the arguments to pass at compile time |
| 61 | CLBuildOptions build_opts; |
| 62 | build_opts.add_option("-DCOLS_A=" + support::cpp11::to_string(mtx_a->info()->dimension(0))); |
| 63 | |
| 64 | // Create kernel |
| 65 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_matrix_a_reduction", build_opts.options())); |
| 66 | |
| 67 | const unsigned int num_elems_processed_per_iteration = 1; |
| 68 | |
| 69 | // Configure kernel window |
| 70 | Window win = calculate_max_window(*_output->info(), Steps(num_elems_processed_per_iteration)); |
| 71 | |
| 72 | AccessWindowStatic input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1)); |
| 73 | AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); |
| 74 | |
| 75 | update_window_and_padding(win, |
| 76 | input_access, |
| 77 | output_access); |
| 78 | |
| 79 | output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), _output->info()->tensor_shape())); |
| 80 | |
| 81 | ICLKernel::configure(win); |
| 82 | } |
| 83 | |
| 84 | void CLGEMMLowpMatrixAReductionKernel::run(const Window &window, cl::CommandQueue &queue) |
| 85 | { |
| 86 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 87 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 88 | |
| 89 | Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimY); |
| 90 | Window slice_in = collapsed.first_slice_window_2D(); |
| 91 | Window slice_out = collapsed.first_slice_window_2D(); |
| 92 | |
| 93 | // Setup input slice. Its dimensions are increased in the cl kernel. |
| 94 | slice_in.set(Window::DimX, Window::Dimension(0, 0, 0)); |
| 95 | slice_in.set(Window::DimY, Window::Dimension(0, 0, 0)); |
| 96 | slice_in.set(Window::DimZ, Window::Dimension(0, 0, 0)); |
| 97 | |
| 98 | do |
| 99 | { |
| 100 | unsigned int idx = 0; |
| 101 | add_3D_tensor_argument(idx, _input, slice_in); |
| 102 | add_2D_tensor_argument(idx, _output, slice_out); |
| 103 | enqueue(queue, *this, slice_out); |
| 104 | } |
| 105 | while(collapsed.slide_window_slice_2D(slice_out)); |
| 106 | } |
| 107 | |
| 108 | void CLGEMMLowpMatrixBReductionKernel::configure(const ICLTensor *mtx_b, ICLTensor *vector_sum_col) |
| 109 | { |
| 110 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(mtx_b, 1, DataType::QASYMM8); |
| 111 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(vector_sum_col, 1, DataType::S32); |
| 112 | |
| 113 | _input = mtx_b; |
| 114 | _output = vector_sum_col; |
| 115 | |
| 116 | // Set the arguments to pass at compile time |
| 117 | CLBuildOptions build_opts; |
| 118 | build_opts.add_option("-DCOLS_B=" + support::cpp11::to_string(mtx_b->info()->dimension(0))); |
| 119 | build_opts.add_option("-DROWS_B=" + support::cpp11::to_string(mtx_b->info()->dimension(1))); |
| 120 | |
| 121 | // Create kernel |
| 122 | _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("gemmlowp_matrix_b_reduction", build_opts.options())); |
| 123 | |
| 124 | constexpr unsigned int num_elems_processed_per_iteration = 16; |
| 125 | |
| 126 | // Configure kernel window |
| 127 | Window win = calculate_max_window(*vector_sum_col->info(), Steps(num_elems_processed_per_iteration)); |
| 128 | |
| 129 | AccessWindowStatic input_access(_input->info(), 0, 0, ceil_to_multiple(_input->info()->dimension(0), 16), _input->info()->dimension(1)); |
| 130 | AccessWindowHorizontal output_access(_output->info(), 0, num_elems_processed_per_iteration); |
| 131 | |
| 132 | update_window_and_padding(win, |
| 133 | input_access, |
| 134 | output_access); |
| 135 | |
| 136 | output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), _output->info()->tensor_shape())); |
| 137 | |
| 138 | ICLKernel::configure(win); |
| 139 | } |
| 140 | |
| 141 | void CLGEMMLowpMatrixBReductionKernel::run(const Window &window, cl::CommandQueue &queue) |
| 142 | { |
| 143 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 144 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); |
| 145 | |
| 146 | Window collapsed = window.collapse_if_possible(IKernel::window(), Window::DimY); |
| 147 | |
| 148 | Window slice_out = collapsed.first_slice_window_2D(); |
| 149 | Window slice_in = slice_out; |
| 150 | |
| 151 | slice_in.set(Window::DimY, Window::Dimension(0, 1, 1)); |
| 152 | slice_in.set(Window::DimZ, Window::Dimension(0, 1, 1)); |
| 153 | |
| 154 | do |
| 155 | { |
| 156 | unsigned int idx = 0; |
| 157 | add_3D_tensor_argument(idx, _input, slice_in); |
| 158 | add_2D_tensor_argument(idx, _output, slice_out); |
| 159 | enqueue(queue, *this, slice_out); |
| 160 | } |
| 161 | while(collapsed.slide_window_slice_2D(slice_out)); |
| 162 | } |