Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018 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/CLWinogradConvolutionLayer.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
| 27 | #include "arm_compute/core/Utils.h" |
| 28 | #include "arm_compute/core/Validate.h" |
| 29 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 30 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 31 | |
| 32 | using namespace arm_compute; |
| 33 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 34 | namespace |
| 35 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 36 | Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataLayout data_layout) |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 37 | { |
| 38 | Size2D output_tile = Size2D{}; |
| 39 | |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 40 | const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height); |
| 41 | |
| 42 | // Check if the input spatial dimensions are smaller than 4 |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 43 | const bool is_input_lt4_nchw = (input_dims.width <= 4 && input_dims.height <= 4) && (data_layout == DataLayout::NCHW); |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 44 | |
| 45 | if(kernel_max_dim == 3U) |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 46 | { |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 47 | if(kernel_dims == Size2D(3U, 3U)) |
| 48 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 49 | output_tile = is_input_lt4_nchw ? Size2D(2U, 2U) : Size2D(4U, 4U); |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 50 | } |
| 51 | else if(kernel_dims == Size2D(3U, 1U)) |
| 52 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 53 | output_tile = is_input_lt4_nchw ? Size2D(2U, 1U) : Size2D(4U, 1U); |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 54 | } |
| 55 | else |
| 56 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 57 | output_tile = is_input_lt4_nchw ? Size2D(1U, 2U) : Size2D(1U, 4U); |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 58 | } |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 59 | } |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 60 | else if(kernel_max_dim == 5U) |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 61 | { |
Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 62 | output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U, |
| 63 | kernel_dims.height == 1 ? 1U : 4U); |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 64 | } |
| 65 | |
| 66 | return output_tile; |
| 67 | } |
| 68 | |
| 69 | bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size) |
| 70 | { |
| 71 | // Check if we want to configure a Winograd configuration which requires fast math |
| 72 | using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; |
| 73 | |
| 74 | std::vector<WinogradConfiguration> fast_math_winograd = |
| 75 | { |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 76 | WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)) |
| 77 | }; |
| 78 | |
| 79 | auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), |
| 80 | std::pair<int, int>(kernel_size.width, kernel_size.height)); |
| 81 | |
| 82 | return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end(); |
| 83 | } |
| 84 | } // namespace |
| 85 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 86 | CLWinogradConvolutionLayer::CLWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 87 | : _memory_group(memory_manager), _batched_mm(memory_manager), _input_transform(), _filter_transform(), _output_transform(), _activationlayer_function(), _input0(), _input1(), _batched_mm_output(), |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 88 | _original_weights(nullptr), _is_prepared(false), _is_activationlayer_enabled(false) |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 89 | { |
| 90 | } |
| 91 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 92 | void CLWinogradConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, |
| 93 | bool enable_fast_math) |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 94 | { |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 95 | // Get indices for the width and height |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 96 | const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH); |
| 97 | const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT); |
| 98 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 99 | // Input shape, kernel size and output tile |
| 100 | const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]); |
| 101 | const Size2D kernel_size = Size2D(weights->info()->tensor_shape()[idx_width], weights->info()->tensor_shape()[idx_height]); |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 102 | const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->info()->data_layout()); |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 103 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 104 | // Check if the Winograd configuration requires fast math |
| 105 | if(!enable_fast_math) |
| 106 | { |
| 107 | ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); |
| 108 | } |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 109 | |
| 110 | const WinogradInfo winograd_info = WinogradInfo(output_tile, |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 111 | kernel_size, |
| 112 | input_dims, |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 113 | conv_info, |
| 114 | input->info()->data_layout()); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 115 | |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 116 | _is_prepared = false; |
| 117 | _original_weights = weights; |
| 118 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 119 | // Manage intermediate tensors |
| 120 | _memory_group.manage(&_input0); |
| 121 | _memory_group.manage(&_batched_mm_output); |
| 122 | |
| 123 | // Do not manage _input1 as it contains the weights |
| 124 | |
| 125 | // Configure input transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 126 | _input_transform.configure(input, &_input0, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 127 | |
| 128 | // Configure filter transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 129 | _filter_transform.configure(weights, &_input1, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 130 | |
| 131 | // Configure batched matrix multiply |
| 132 | _batched_mm.configure(&_input0, &_input1, nullptr, &_batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/)); |
| 133 | |
| 134 | // Configure output transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 135 | _output_transform.configure(&_batched_mm_output, biases, output, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 136 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 137 | // Configure activation layer |
| 138 | _is_activationlayer_enabled = act_info.enabled(); |
| 139 | if(_is_activationlayer_enabled) |
| 140 | { |
| 141 | _activationlayer_function.configure(output, nullptr, act_info); |
| 142 | } |
| 143 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 144 | // Allocate temporary tensors |
| 145 | _input0.allocator()->allocate(); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 146 | _batched_mm_output.allocator()->allocate(); |
| 147 | } |
| 148 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 149 | Status CLWinogradConvolutionLayer::validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, const PadStrideInfo &conv_info, |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 150 | const ActivationLayerInfo &act_info, bool enable_fast_math) |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 151 | { |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 152 | // Get indeces for the width and height |
| 153 | const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); |
| 154 | const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); |
| 155 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 156 | // Input shape, kernel size and output tile |
| 157 | const Size2D input_dims = Size2D(input->tensor_shape()[idx_width], input->tensor_shape()[idx_height]); |
| 158 | const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 159 | const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, input->data_layout()); |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 160 | |
Vidhya Sudhan Loganathan | d307f06 | 2018-10-04 09:21:15 +0100 | [diff] [blame] | 161 | //FP16 implementation of winograd is slower than direct convolution. |
| 162 | //The following check needs to be removed when fp16 winograd is faster than direct convolution (COMPMID-1266) |
| 163 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); |
| 164 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 165 | // Check if the Winograd configuration requires fast math |
| 166 | if(!enable_fast_math) |
| 167 | { |
| 168 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); |
| 169 | } |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 170 | |
| 171 | const WinogradInfo winograd_info = WinogradInfo(output_tile, |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 172 | kernel_size, |
| 173 | input_dims, |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 174 | conv_info, |
| 175 | input->data_layout()); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 176 | |
| 177 | // Validate input transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 178 | const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*input, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 179 | const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 180 | ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradInputTransform::validate(input, &input0, winograd_info)); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 181 | |
| 182 | // Validate filter transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 183 | const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 184 | const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 185 | ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradFilterTransformKernel::validate(weights, &input1, winograd_info)); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 186 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 187 | // Validate batched matrix multiply |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 188 | TensorShape batched_mm_output_shape = input0.tensor_shape(); |
| 189 | batched_mm_output_shape[0] = input1.tensor_shape()[0]; |
| 190 | const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); |
| 191 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input0, &input1, nullptr, &batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/))); |
| 192 | |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 193 | // Configure output transform |
| 194 | ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradOutputTransformKernel::validate(&batched_mm_output, biases, output, winograd_info)); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 195 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 196 | // Validate Activation Layer |
| 197 | if(act_info.enabled()) |
| 198 | { |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 199 | ARM_COMPUTE_RETURN_ON_ERROR(CLActivationLayer::validate(output, nullptr, act_info)); |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 200 | } |
| 201 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 202 | return Status{}; |
| 203 | } |
| 204 | |
| 205 | void CLWinogradConvolutionLayer::run() |
| 206 | { |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 207 | prepare(); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 208 | |
| 209 | _memory_group.acquire(); |
| 210 | |
| 211 | // Run input transform |
| 212 | _input_transform.run(); |
| 213 | |
| 214 | // Run batched matrix multiplication |
| 215 | _batched_mm.run(); |
| 216 | |
| 217 | // Run output transform |
| 218 | CLScheduler::get().enqueue(_output_transform); |
| 219 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 220 | if(_is_activationlayer_enabled) |
| 221 | { |
| 222 | _activationlayer_function.run(); |
| 223 | } |
| 224 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 225 | _memory_group.release(); |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 226 | } |
Georgios Pinitas | 82b5148 | 2018-04-24 15:14:12 +0100 | [diff] [blame] | 227 | |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 228 | void CLWinogradConvolutionLayer::prepare() |
| 229 | { |
| 230 | if(!_is_prepared) |
| 231 | { |
| 232 | // Run filter transform and mark original weights as unused |
| 233 | _input1.allocator()->allocate(); |
| 234 | CLScheduler::get().enqueue(_filter_transform, false); |
| 235 | _original_weights->mark_as_unused(); |
| 236 | |
| 237 | // Prepare GEMM and release reshaped weights if marked unused by CLGEMM |
| 238 | _batched_mm.prepare(); |
| 239 | if(!_input1.is_used()) |
| 240 | { |
| 241 | _input1.allocator()->free(); |
| 242 | } |
| 243 | |
| 244 | CLScheduler::get().queue().finish(); |
| 245 | _is_prepared = true; |
| 246 | } |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 247 | } |