Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [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/NEON/kernels/NEGEMMInterleaveBlockedKernel.h" |
| 25 | |
| 26 | #include "arm_compute/core/Error.h" |
| 27 | #include "arm_compute/core/Helpers.h" |
| 28 | #include "arm_compute/core/ITensor.h" |
| 29 | #include "arm_compute/core/NEON/INEKernel.h" |
| 30 | #include "arm_compute/core/Types.h" |
| 31 | #include "arm_compute/core/Validate.h" |
| 32 | #include "arm_compute/core/Window.h" |
| 33 | |
| 34 | #include <arm_neon.h> |
| 35 | #include <cstddef> |
| 36 | #include <cstdint> |
| 37 | #include <tuple> |
| 38 | |
| 39 | using namespace arm_compute; |
| 40 | |
| 41 | namespace |
| 42 | { |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 43 | inline void gemm_interleave_blocked_transposed_8bit(const ITensor *input, ITensor *output, const Window &window, unsigned int block_width, unsigned int block_height) |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 44 | { |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 45 | const size_t in_stride = input->info()->strides_in_bytes()[1]; |
| 46 | |
| 47 | const unsigned int in_height = input->info()->dimension(1); |
| 48 | const unsigned int in_width = input->info()->dimension(0); |
| 49 | |
| 50 | const float scale_y_factor = 1.f / float(block_height); |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 51 | |
| 52 | // Set window for output tensor |
| 53 | Window win_out(window); |
| 54 | win_out.scale(Window::DimY, scale_y_factor); |
| 55 | Iterator in(input, window); |
| 56 | |
| 57 | win_out.set_dimension_step(Window::DimX, block_width * block_height); |
| 58 | Iterator out(output, win_out); |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 59 | |
| 60 | execute_window_loop(window, [&](const Coordinates &) |
| 61 | { |
| 62 | std::fill_n(out.ptr(), block_width * block_height, 0); |
| 63 | }, |
| 64 | out); |
| 65 | |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 66 | execute_window_loop(window, [&](const Coordinates & id) |
| 67 | { |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 68 | for(unsigned int z = id.y(); (z < in_width) && z < (id.y() + block_height); ++z) |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 69 | { |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 70 | int j = (z - id.y()) * block_width; |
| 71 | for(unsigned int b = id.x(); (b < in_height) && (b < (id.x() + block_width)); ++b) |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 72 | { |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 73 | *(out.ptr() + j++) = *(input->buffer() + b * in_stride + z); |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 74 | } |
| 75 | } |
| 76 | }, |
| 77 | in, out); |
| 78 | } |
| 79 | |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 80 | inline void gemm_interleave_blocked_8bit(const ITensor *input, ITensor *output, const Window &window, unsigned int block_width, unsigned int block_height) |
| 81 | { |
| 82 | const size_t in_stride = input->info()->strides_in_bytes()[1]; |
| 83 | |
| 84 | const unsigned int in_height = input->info()->dimension(1); |
| 85 | const unsigned int in_width = input->info()->dimension(0); |
| 86 | |
| 87 | const float scale_y_factor = 1.f / float(block_height); |
| 88 | |
| 89 | // Set window for output tensor |
| 90 | Window win_out(window); |
| 91 | win_out.scale(Window::DimY, scale_y_factor); |
| 92 | Iterator in(input, window); |
| 93 | |
| 94 | win_out.set_dimension_step(Window::DimX, block_width * block_height); |
| 95 | Iterator out(output, win_out); |
| 96 | |
| 97 | execute_window_loop(window, [&](const Coordinates &) |
| 98 | { |
| 99 | std::fill_n(out.ptr(), block_width * block_height, 0); |
| 100 | }, |
| 101 | out); |
| 102 | |
| 103 | execute_window_loop(window, [&](const Coordinates & id) |
| 104 | { |
| 105 | for(unsigned int z = id.y(); (z < in_height) && z < (id.y() + block_height); ++z) |
| 106 | { |
| 107 | int j = (z - id.y()) * block_width; |
| 108 | for(unsigned int b = id.x(); (b < in_width) && (b < (id.x() + block_width)); ++b) |
| 109 | { |
| 110 | *(out.ptr() + j++) = *(input->buffer() + z * in_stride + b); |
| 111 | } |
| 112 | } |
| 113 | }, |
| 114 | in, out); |
| 115 | } |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 116 | } // namespace |
| 117 | |
| 118 | NEGEMMInterleaveBlockedKernel::NEGEMMInterleaveBlockedKernel() |
| 119 | : _block_height(0), _block_width(0), _transpose(false) |
| 120 | { |
| 121 | } |
| 122 | |
| 123 | void NEGEMMInterleaveBlockedKernel::configure(const ITensor *input, ITensor *output, unsigned int block_height, unsigned int block_width, bool transpose) |
| 124 | { |
| 125 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); |
| 126 | ARM_COMPUTE_ERROR_ON_NULLPTR(output); |
| 127 | ARM_COMPUTE_ERROR_ON_MSG(block_height < 1, "Block height must be greater than 0"); |
| 128 | ARM_COMPUTE_ERROR_ON_MSG(block_width < 1, "Block window must be greater than 0"); |
| 129 | |
| 130 | TensorShape output_shape = input->info()->tensor_shape(); |
| 131 | const float interleave_by_f32 = block_height; |
| 132 | output_shape.set(0, input->info()->dimension(0) * interleave_by_f32); |
| 133 | output_shape.set(1, std::ceil(static_cast<float>(input->info()->dimension(1)) / interleave_by_f32)); |
| 134 | // Output auto inizialitation if not yet initialized |
| 135 | auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); |
| 136 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); |
| 137 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); |
| 138 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(input, output); |
| 139 | |
| 140 | _input = input; |
| 141 | _output = output; |
| 142 | _block_height = block_height; |
| 143 | _block_width = block_width; |
| 144 | _transpose = transpose; |
| 145 | |
| 146 | const unsigned int num_elems_processed_per_iteration_x = block_width; |
| 147 | const unsigned int num_elems_processed_per_iteration_y = block_height; |
| 148 | |
| 149 | // Configure kernel window |
| 150 | Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); |
| 151 | const float scaley_factor = 1.f / interleave_by_f32; |
| 152 | |
| 153 | AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x * num_elems_processed_per_iteration_y, 1, num_elems_processed_per_iteration_y, scaley_factor); |
| 154 | AccessWindowRectangle input_access(input->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); |
| 155 | update_window_and_padding(win, output_access, input_access); |
| 156 | |
| 157 | output_access.set_valid_region(win, input->info()->valid_region()); |
| 158 | |
| 159 | INEKernel::configure(win); |
| 160 | } |
| 161 | |
| 162 | void NEGEMMInterleaveBlockedKernel::run(const Window &window, const ThreadInfo &info) |
| 163 | { |
| 164 | ARM_COMPUTE_UNUSED(info); |
| 165 | ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); |
| 166 | ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window); |
Pablo Tello | eae4ce0 | 2017-10-13 15:24:49 +0100 | [diff] [blame] | 167 | if(_transpose) |
| 168 | { |
| 169 | gemm_interleave_blocked_transposed_8bit(_input, _output, window, _block_width, _block_height); |
| 170 | } |
| 171 | else |
| 172 | { |
| 173 | gemm_interleave_blocked_8bit(_input, _output, window, _block_width, _block_height); |
| 174 | } |
Pablo Tello | bf2fb95 | 2017-09-29 16:43:25 +0100 | [diff] [blame] | 175 | } |