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" |
| 27 | #include "arm_compute/core/Utils.h" |
| 28 | #include "arm_compute/core/Validate.h" |
| 29 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 30 | |
| 31 | #include <cmath> |
| 32 | #include <tuple> |
| 33 | |
| 34 | using namespace arm_compute; |
| 35 | |
| 36 | CLConvolutionLayerReshapeWeights::CLConvolutionLayerReshapeWeights() |
| 37 | : _weights_reshape_kernel(), _weights_transposed_kernel(), _weights_reshaped(), _transpose1xW(false) |
| 38 | { |
| 39 | } |
| 40 | |
| 41 | void CLConvolutionLayerReshapeWeights::configure(const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, bool transpose1xW) |
| 42 | { |
| 43 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32); |
| 44 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F32); |
| 45 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F32); |
| 46 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases, output); |
| 47 | ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT(weights, biases, output); |
| 48 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 49 | |
| 50 | if(biases != nullptr) |
| 51 | { |
| 52 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F32); |
| 53 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(weights, biases); |
| 54 | ARM_COMPUTE_ERROR_ON(biases->info()->dimension(0) != weights->info()->dimension(3)); |
| 55 | ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); |
| 56 | } |
| 57 | |
| 58 | const bool _has_bias = (biases != nullptr); |
| 59 | |
| 60 | _transpose1xW = transpose1xW; |
| 61 | |
| 62 | if(transpose1xW) |
| 63 | { |
| 64 | // Create tensor to store the reshaped weights |
| 65 | const unsigned int mat_weights_cols = weights->info()->dimension(3); |
| 66 | const unsigned int mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + (_has_bias ? 1 : 0); |
| 67 | TensorShape shape_wr(mat_weights_cols, mat_weights_rows); |
| 68 | const DataType dt = weights->info()->data_type(); |
| 69 | TensorInfo info_wr(shape_wr, 1, dt); |
| 70 | |
| 71 | _weights_reshaped.allocator()->init(info_wr); |
| 72 | _weights_reshape_kernel.configure(weights, biases, &_weights_reshaped); |
| 73 | _weights_transposed_kernel.configure(&_weights_reshaped, output); |
| 74 | _weights_reshaped.allocator()->allocate(); |
| 75 | } |
| 76 | else |
| 77 | { |
| 78 | _weights_reshape_kernel.configure(weights, biases, output); |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | void CLConvolutionLayerReshapeWeights::run() |
| 83 | { |
| 84 | cl::CommandQueue q = CLScheduler::get().queue(); |
| 85 | CLScheduler::get().enqueue(_weights_reshape_kernel); |
| 86 | if(_transpose1xW) |
| 87 | { |
| 88 | CLScheduler::get().enqueue(_weights_transposed_kernel); |
| 89 | } |
| 90 | } |
| 91 | |
| 92 | CLConvolutionLayer::CLConvolutionLayer() |
| 93 | : _reshape_weights(), _input_im2col_kernel(), _input_interleave_kernel(), _mm_kernel(), _output_col2im_kernel(), _input_im2col_reshaped(), _input_interleaved_reshaped(), _weights_reshaped(), |
| 94 | _weights_transposed(), _gemm_output(), _has_bias(false), _is_fully_connected_convolution(false), _are_weights_reshaped(false) |
| 95 | { |
| 96 | } |
| 97 | |
| 98 | void CLConvolutionLayer::configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const WeightsInfo &weights_info) |
| 99 | { |
| 100 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); |
| 101 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(weights, 1, DataType::F16, DataType::F32); |
| 102 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::F16, DataType::F32); |
| 103 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, weights, output); |
| 104 | ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && weights->info()->dimension(2) != input->info()->dimension(2)); |
| 105 | ARM_COMPUTE_ERROR_ON(weights->info()->num_dimensions() > 4); |
| 106 | |
| 107 | if(biases != nullptr) |
| 108 | { |
| 109 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(biases, 1, DataType::F16, DataType::F32); |
| 110 | ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, biases); |
| 111 | ARM_COMPUTE_ERROR_ON(!weights_info.are_reshaped() && biases->info()->dimension(0) != weights->info()->dimension(3)); |
| 112 | ARM_COMPUTE_ERROR_ON(biases->info()->num_dimensions() > 1); |
| 113 | } |
| 114 | |
| 115 | _has_bias = (biases != nullptr); |
| 116 | _are_weights_reshaped = weights_info.are_reshaped(); |
| 117 | |
| 118 | // Get parameters for conv_info |
| 119 | unsigned int stride_x = 0; |
| 120 | unsigned int stride_y = 0; |
| 121 | unsigned int pad_x = 0; |
| 122 | unsigned int pad_y = 0; |
| 123 | std::tie(stride_x, stride_y) = conv_info.stride(); |
| 124 | std::tie(pad_x, pad_y) = conv_info.pad(); |
| 125 | |
| 126 | // Get convolved dimensions |
| 127 | unsigned int conv_w = 0; |
| 128 | unsigned int conv_h = 0; |
| 129 | |
Gian Marco Iodice | 4e28869 | 2017-06-27 11:41:59 +0100 | [diff] [blame] | 130 | const unsigned int kernel_width = _are_weights_reshaped ? weights_info.kernel_size().first : weights->info()->dimension(0); |
| 131 | const unsigned int kernel_height = _are_weights_reshaped ? weights_info.kernel_size().second : weights->info()->dimension(1); |
| 132 | std::tie(conv_w, conv_h) = scaled_dimensions(input->info()->dimension(0), input->info()->dimension(1), kernel_width, kernel_height, |
| 133 | conv_info); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 134 | 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"); |
| 135 | |
| 136 | // Check if its a "fully connected" convolution |
| 137 | _is_fully_connected_convolution = ((conv_w == 1) && (conv_h == 1)); |
| 138 | |
| 139 | // Create tensor to store the reshaped weights |
| 140 | size_t mat_weights_cols = weights->info()->dimension(3); |
| 141 | size_t mat_weights_rows = weights->info()->dimension(0) * weights->info()->dimension(1) * weights->info()->dimension(2) + ((_has_bias) ? 1 : 0); |
| 142 | if(_are_weights_reshaped) |
| 143 | { |
| 144 | mat_weights_cols = output->info()->dimension(2); |
| 145 | const unsigned int quarter_reshaped_cols = weights->info()->dimension(0) / 4; |
| 146 | mat_weights_rows = (_has_bias ? 1 + quarter_reshaped_cols : quarter_reshaped_cols); |
| 147 | } |
| 148 | else |
| 149 | { |
| 150 | if(_is_fully_connected_convolution) |
| 151 | { |
| 152 | // Create tensor to store the reshaped weights |
| 153 | TensorShape shape_wr(mat_weights_cols, mat_weights_rows); |
| 154 | TensorInfo info_wr(shape_wr, 1, weights->info()->data_type()); |
| 155 | _weights_reshaped.allocator()->init(info_wr); |
| 156 | _reshape_weights.configure(weights, biases, &_weights_reshaped, false); |
| 157 | weights = &_weights_reshaped; |
| 158 | } |
| 159 | else |
| 160 | { |
| 161 | // Create tensor to store transposed weights |
| 162 | TensorShape shape_wt(mat_weights_rows * 4, static_cast<size_t>(std::ceil(mat_weights_cols / 4.f))); |
| 163 | TensorInfo info_wt(shape_wt, 1, weights->info()->data_type()); |
| 164 | _weights_transposed.allocator()->init(info_wt); |
| 165 | _reshape_weights.configure(weights, biases, &_weights_transposed, true); |
| 166 | weights = &_weights_transposed; |
| 167 | } |
| 168 | } |
| 169 | // Create tensor to store im2col reshaped inputs |
| 170 | const size_t mat_input_cols = mat_weights_rows; |
| 171 | const size_t mat_input_rows = conv_w * conv_h; |
| 172 | TensorShape shape_im2col = input->info()->tensor_shape(); |
| 173 | shape_im2col.set(0, mat_input_cols); |
| 174 | shape_im2col.set(1, mat_input_rows); |
| 175 | shape_im2col.set(2, 1); |
| 176 | _input_im2col_reshaped.allocator()->init(TensorInfo(shape_im2col, 1, input->info()->data_type())); |
| 177 | |
| 178 | // Create tensor (interleave) to prepare input tensor for GEMM |
| 179 | if(!_is_fully_connected_convolution) |
| 180 | { |
| 181 | TensorShape shape_interleaved = shape_im2col; |
| 182 | shape_interleaved.set(0, shape_interleaved.x() * 4); |
| 183 | shape_interleaved.set(1, std::ceil(static_cast<float>(shape_interleaved.y()) / 4.f)); |
| 184 | _input_interleaved_reshaped.allocator()->init(TensorInfo(shape_interleaved, 1, input->info()->data_type())); |
| 185 | } |
| 186 | |
| 187 | // Create GEMM output tensor |
| 188 | TensorShape shape_gemm = _input_im2col_reshaped.info()->tensor_shape(); |
| 189 | shape_gemm.set(0, mat_weights_cols); |
| 190 | shape_gemm.set(1, mat_input_rows); |
| 191 | _gemm_output.allocator()->init(TensorInfo(shape_gemm, 1, input->info()->data_type())); |
| 192 | |
| 193 | // Configure kernels |
| 194 | _input_im2col_kernel.configure(input, &_input_im2col_reshaped, std::make_pair(conv_w, conv_h), conv_info, _has_bias); |
| 195 | _output_col2im_kernel.configure(&_gemm_output, output, std::make_pair(conv_w, conv_h)); |
| 196 | |
| 197 | if(_is_fully_connected_convolution) |
| 198 | { |
| 199 | _mm_kernel.configure(&_input_im2col_reshaped, weights, &_gemm_output, 1.0f); |
| 200 | } |
| 201 | else |
| 202 | { |
| 203 | _input_interleave_kernel.configure(&_input_im2col_reshaped, &_input_interleaved_reshaped); |
| 204 | _mm_kernel.configure(&_input_interleaved_reshaped, weights, &_gemm_output, 1.0f); |
| 205 | } |
| 206 | |
| 207 | if(!_are_weights_reshaped) |
| 208 | { |
| 209 | if(!_is_fully_connected_convolution) |
| 210 | { |
| 211 | _weights_transposed.allocator()->allocate(); |
| 212 | } |
| 213 | else |
| 214 | { |
| 215 | _weights_reshaped.allocator()->allocate(); |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | _input_im2col_reshaped.allocator()->allocate(); |
| 220 | if(!_is_fully_connected_convolution) |
| 221 | { |
| 222 | _input_interleaved_reshaped.allocator()->allocate(); |
| 223 | } |
| 224 | _gemm_output.allocator()->allocate(); |
| 225 | } |
| 226 | |
| 227 | void CLConvolutionLayer::run() |
| 228 | { |
| 229 | // Run weights reshaping (Runs once for every configure) |
| 230 | if(!_are_weights_reshaped) |
| 231 | { |
| 232 | _are_weights_reshaped = true; |
| 233 | _reshape_weights.run(); |
| 234 | } |
| 235 | |
| 236 | // Run input reshaping |
| 237 | CLScheduler::get().enqueue(_input_im2col_kernel); |
| 238 | if(!_is_fully_connected_convolution) |
| 239 | { |
| 240 | CLScheduler::get().enqueue(_input_interleave_kernel); |
| 241 | } |
| 242 | |
| 243 | // Runs matrix multiply on reshaped matrices |
| 244 | CLScheduler::get().enqueue(_mm_kernel); |
| 245 | |
| 246 | // Reshape output matrix |
| 247 | CLScheduler::get().enqueue(_output_col2im_kernel, false); |
| 248 | } |