Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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/runtime/CL/functions/CLConvolutionLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/PixelValue.h" |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 27 | #include "arm_compute/core/Size2D.h" |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 28 | #include "arm_compute/core/Utils.h" |
| 29 | #include "arm_compute/core/Validate.h" |
| 30 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 31 | |
| 32 | #include <cmath> |
| 33 | #include <tuple> |
| 34 | |
| 35 | using namespace arm_compute; |
| 36 | |
| 37 | CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights() |
| 38 | : _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) |
| 39 | { |
| 40 | } |
| 41 | |
| 42 | void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW) |
| 43 | { |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 44 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F16, DataType::F32); |
| 45 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, output); |
| 46 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, output); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 47 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 48 | |
| 49 | if(biases != nullptr) |
| 50 | { |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 51 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 52 | ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); |
| 53 | ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); |
| 54 | } |
| 55 | |
| 56 | const bool _has_bias = (biases != nullptr); |
| 57 | |
| 58 | _transpose1xW = transpose1xW; |
| 59 | |
| 60 | if(transpose1xW) |
| 61 | { |
| 62 | // Create tensor to store the reshaped weights |
| 63 | const unsigned int mat_weights_cols = weights->info()->dimension(3); |
| 64 | const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0); |
| 65 | TensorShape shape_wr(mat_weights_cols, mat_weights_rows); |
| 66 | const DataType dt = weights->info()->data_type(); |
| 67 | TensorInfo info_wr(shape_wr, 1, dt); |
| 68 | |
| 69 | _weights_reshaped.allocator()->init(info_wr); |
| 70 | _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped); |
| 71 | _weights_transposed_kernel.configure(&_weights_reshaped, output); |
| 72 | _weights_reshaped.allocator()->allocate(); |
| 73 | } |
| 74 | else |
| 75 | { |
| 76 | _weights_reshape_kernel.configure(weights, biases, output); |
| 77 | } |
| 78 | } |
| 79 | |
| 80 | void CLConvolutionLayerReshapeWeights::run() |
| 81 | { |
| 82 | cl::CommandQueue q = CLScheduler::get().queue(); |
| 83 | CLScheduler::get().enqueue(_weights_reshape_kernel); |
| 84 | if(_transpose1xW) |
| 85 | { |
| 86 | CLScheduler::get().enqueue(_weights_transposed_kernel); |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | CLConvolutionLayer::CLConvolutionLayer() |
| 91 | : _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), |
| 92 | _weights_transposed(), _gemm_output(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) |
| 93 | { |
| 94 | } |
| 95 | |
| 96 | void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) |
| 97 | { |
| 98 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 99 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 100 | ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2)); |
| 101 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 102 | |
| 103 | if(biases != nullptr) |
| 104 | { |
| 105 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F16, DataType::F32); |
| 106 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 107 | ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3)); |
| 108 | ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); |
| 109 | } |
| 110 | |
| 111 | _has_bias = (biases != nullptr); |
| 112 | _are_weights_reshaped = weights_info.are_reshaped(); |
| 113 | |
| 114 | // Get parameters for conv_info |
| 115 | unsigned int stride_x = 0; |
| 116 | unsigned int stride_y = 0; |
| 117 | unsigned int pad_x = 0; |
| 118 | unsigned int pad_y = 0; |
| 119 | std::tie(stride_x, stride_y) = conv_info.stride(); |
| 120 | std::tie(pad_x, pad_y) = conv_info.pad(); |
| 121 | |
| 122 | // Get convolved dimensions |
| 123 | unsigned int conv_w = 0; |
| 124 | unsigned int conv_h = 0; |
| 125 | |
Gian Marco Iodice | 4e28869 | 2017-06-27 11:41:59 +0100 | [diff] [blame] | 126 | const unsigned int kernel_width = _are_weights_reshaped ? weights_info.kernel_size().first : weights->info()->dimension(0); |
| 127 | const unsigned int kernel_height = _are_weights_reshaped ? weights_info.kernel_size().second : weights->info()->dimension(1); |
| 128 | std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height, |
| 129 | conv_info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 130 | ARM_COMPUTE_ERROR_ON_MSG((output->info()->dimension(0) != conv_w) || (output->info()->dimension(1) != conv_h), "Output shape does not match the expected one"); |
| 131 | |
| 132 | // Check if its a "fully connected" convolution |
| 133 | _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); |
| 134 | |
| 135 | // Create tensor to store the reshaped weights |
| 136 | size_t mat_weights_cols = weights->info()->dimension(3); |
| 137 | size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0); |
| 138 | if(_are_weights_reshaped) |
| 139 | { |
| 140 | mat_weights_cols = output->info()->dimension(2); |
| 141 | const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4; |
| 142 | mat_weights_rows = (_has_bias ? 1 + quarter_reshaped_cols : quarter_reshaped_cols); |
| 143 | } |
| 144 | else |
| 145 | { |
| 146 | if(_is_fully_connected_convolution) |
| 147 | { |
| 148 | // Create tensor to store the reshaped weights |
| 149 | TensorShape shape_wr(mat_weights_cols, mat_weights_rows); |
| 150 | TensorInfo info_wr(shape_wr, 1, weights->info()->data_type()); |
| 151 | _weights_reshaped.allocator()->init(info_wr); |
| 152 | _reshape_weights.configure(weights, biases, &_weights_reshaped, false); |
| 153 | weights = &_weights_reshaped; |
| 154 | } |
| 155 | else |
| 156 | { |
| 157 | // Create tensor to store transposed weights |
| 158 | TensorShape shape_wt(mat_weights_rows * 4, static_cast<size_t>(std::ceil(mat_weights_cols / 4.f))); |
| 159 | TensorInfo info_wt(shape_wt, 1, weights->info()->data_type()); |
| 160 | _weights_transposed.allocator()->init(info_wt); |
| 161 | _reshape_weights.configure(weights, biases, &_weights_transposed, true); |
| 162 | weights = &_weights_transposed; |
| 163 | } |
| 164 | } |
| 165 | // Create tensor to store im2col reshaped inputs |
| 166 | const size_t mat_input_cols = mat_weights_rows; |
| 167 | const size_t mat_input_rows = conv_w * conv_h; |
| 168 | TensorShape shape_im2col = input->info()->tensor_shape(); |
| 169 | shape_im2col.set(0, mat_input_cols); |
| 170 | shape_im2col.set(1, mat_input_rows); |
| 171 | shape_im2col.set(2, 1); |
| 172 | _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type())); |
| 173 | |
| 174 | // Create tensor (interleave) to prepare input tensor for GEMM |
| 175 | if(!_is_fully_connected_convolution) |
| 176 | { |
| 177 | TensorShape shape_interleaved = shape_im2col; |
| 178 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 179 | shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4.f)); |
| 180 | _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type())); |
| 181 | } |
| 182 | |
| 183 | // Create GEMM output tensor |
| 184 | TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape(); |
| 185 | shape_gemm.set(0, mat_weights_cols); |
| 186 | shape_gemm.set(1, mat_input_rows); |
| 187 | _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type())); |
| 188 | |
| 189 | // Configure kernels |
Gian Marco Iodice | 13edbff | 2017-06-26 17:20:16 +0100 | [diff] [blame] | 190 | _input_im2col_kernel.configure(input, &_input_im2col_reshaped, Size2D(kernel_width, kernel_height), conv_info, _has_bias); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 191 | _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h)); |
| 192 | |
| 193 | if(_is_fully_connected_convolution) |
| 194 | { |
| 195 | _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f); |
| 196 | } |
| 197 | else |
| 198 | { |
| 199 | _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); |
| 200 | _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f); |
| 201 | } |
| 202 | |
| 203 | if(!_are_weights_reshaped) |
| 204 | { |
| 205 | if(!_is_fully_connected_convolution) |
| 206 | { |
| 207 | _weights_transposed.allocator()->allocate(); |
| 208 | } |
| 209 | else |
| 210 | { |
| 211 | _weights_reshaped.allocator()->allocate(); |
| 212 | } |
| 213 | } |
| 214 | |
| 215 | _input_im2col_reshaped.allocator()->allocate(); |
| 216 | if(!_is_fully_connected_convolution) |
| 217 | { |
| 218 | _input_interleaved_reshaped.allocator()->allocate(); |
| 219 | } |
| 220 | _gemm_output.allocator()->allocate(); |
| 221 | } |
| 222 | |
| 223 | void CLConvolutionLayer::run() |
| 224 | { |
| 225 | // Run weights reshaping (Runs once for every configure) |
| 226 | if(!_are_weights_reshaped) |
| 227 | { |
| 228 | _are_weights_reshaped = true; |
| 229 | _reshape_weights.run(); |
| 230 | } |
| 231 | |
| 232 | // Run input reshaping |
| 233 | CLScheduler::get().enqueue(_input_im2col_kernel); |
| 234 | if(!_is_fully_connected_convolution) |
| 235 | { |
| 236 | CLScheduler::get().enqueue(_input_interleave_kernel); |
| 237 | } |
| 238 | |
| 239 | // Runs matrix multiply on reshaped matrices |
| 240 | CLScheduler::get().enqueue(_mm_kernel); |
| 241 | |
| 242 | // Reshape output matrix |
| 243 | CLScheduler::get().enqueue(_output_col2im_kernel, false); |
| 244 | } |