Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2018-2020 Arm Limited. |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 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" |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 31 | #include "src/core/CL/kernels/CLFillBorderKernel.h" |
| 32 | #include "src/core/CL/kernels/CLGEMMMatrixMultiplyKernel.h" |
| 33 | #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedKernel.h" |
| 34 | #include "src/core/CL/kernels/CLGEMMMatrixMultiplyReshapedOnlyRHSKernel.h" |
| 35 | #include "src/core/CL/kernels/CLGEMMReshapeLHSMatrixKernel.h" |
| 36 | #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h" |
| 37 | #include "src/core/CL/kernels/CLWinogradFilterTransformKernel.h" |
| 38 | #include "src/core/CL/kernels/CLWinogradOutputTransformKernel.h" |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 39 | |
| 40 | using namespace arm_compute; |
| 41 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 42 | namespace |
| 43 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 44 | 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] | 45 | { |
| 46 | Size2D output_tile = Size2D{}; |
| 47 | |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 48 | const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height); |
| 49 | |
| 50 | // Check if the input spatial dimensions are smaller than 4 |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 51 | 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] | 52 | |
| 53 | if(kernel_max_dim == 3U) |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 54 | { |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 55 | if(kernel_dims == Size2D(3U, 3U)) |
| 56 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 57 | 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] | 58 | } |
| 59 | else if(kernel_dims == Size2D(3U, 1U)) |
| 60 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 61 | 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] | 62 | } |
| 63 | else |
| 64 | { |
Gian Marco Iodice | d30714a | 2018-08-15 16:53:27 +0100 | [diff] [blame] | 65 | 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] | 66 | } |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 67 | } |
Gian Marco Iodice | f1c2bf0 | 2018-06-13 14:05:54 +0100 | [diff] [blame] | 68 | else if(kernel_max_dim == 5U) |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 69 | { |
Gian Marco Iodice | 876be2a | 2018-07-03 12:22:09 +0100 | [diff] [blame] | 70 | output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U, |
| 71 | kernel_dims.height == 1 ? 1U : 4U); |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 72 | } |
giuros01 | 3bfacb2 | 2019-04-01 12:07:02 +0100 | [diff] [blame] | 73 | else if(kernel_max_dim == 7U) |
| 74 | { |
Gian Marco Iodice | 4fbcac6 | 2019-04-04 10:14:58 +0100 | [diff] [blame] | 75 | output_tile = Size2D(kernel_dims.width == 1 ? 1U : 2U, |
| 76 | kernel_dims.height == 1 ? 1U : 2U); |
giuros01 | 3bfacb2 | 2019-04-01 12:07:02 +0100 | [diff] [blame] | 77 | } |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 78 | |
| 79 | return output_tile; |
| 80 | } |
| 81 | |
| 82 | bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size) |
| 83 | { |
| 84 | // Check if we want to configure a Winograd configuration which requires fast math |
| 85 | using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; |
| 86 | |
| 87 | std::vector<WinogradConfiguration> fast_math_winograd = |
| 88 | { |
giuros01 | 3bfacb2 | 2019-04-01 12:07:02 +0100 | [diff] [blame] | 89 | WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)), |
| 90 | WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(7, 7)) |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 91 | }; |
| 92 | |
| 93 | auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), |
| 94 | std::pair<int, int>(kernel_size.width, kernel_size.height)); |
| 95 | |
| 96 | return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end(); |
| 97 | } |
| 98 | } // namespace |
| 99 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 100 | CLWinogradConvolutionLayer::CLWinogradConvolutionLayer(std::shared_ptr<IMemoryManager> memory_manager) |
Georgios Pinitas | 40f51a6 | 2020-11-21 03:04:18 +0000 | [diff] [blame] | 101 | : _memory_group(memory_manager), _batched_mm(memory_manager), _input_transform(), _filter_transform(std::make_unique<CLWinogradFilterTransformKernel>()), |
| 102 | _output_transform(std::make_unique<CLWinogradOutputTransformKernel>()), _input0(), _input1(), _batched_mm_output(), _original_weights(nullptr), _is_prepared(false) |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 103 | { |
| 104 | } |
| 105 | |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 106 | CLWinogradConvolutionLayer::~CLWinogradConvolutionLayer() = default; |
| 107 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 108 | void CLWinogradConvolutionLayer::configure(ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, |
| 109 | bool enable_fast_math) |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 110 | { |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 111 | configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, conv_info, act_info, enable_fast_math); |
| 112 | } |
| 113 | |
| 114 | void CLWinogradConvolutionLayer::configure(const CLCompileContext &compile_context, ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 115 | const PadStrideInfo &conv_info, |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 116 | const ActivationLayerInfo &act_info, bool enable_fast_math) |
| 117 | { |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 118 | // Get indices for the width and height |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 119 | const size_t idx_width = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::WIDTH); |
| 120 | const size_t idx_height = get_data_layout_dimension_index(input->info()->data_layout(), DataLayoutDimension::HEIGHT); |
| 121 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 122 | // Input shape, kernel size and output tile |
| 123 | const Size2D input_dims = Size2D(input->info()->tensor_shape()[idx_width], input->info()->tensor_shape()[idx_height]); |
| 124 | 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] | 125 | 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] | 126 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 127 | // Check if the Winograd configuration requires fast math |
| 128 | if(!enable_fast_math) |
| 129 | { |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 130 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); //disable winograd for fp16 if fast math is false. |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 131 | ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); |
| 132 | } |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 133 | const WinogradInfo winograd_info = WinogradInfo(output_tile, |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 134 | kernel_size, |
| 135 | input_dims, |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 136 | conv_info, |
| 137 | input->info()->data_layout()); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 138 | |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 139 | _is_prepared = false; |
| 140 | _original_weights = weights; |
| 141 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 142 | // Manage intermediate tensors |
| 143 | _memory_group.manage(&_input0); |
| 144 | _memory_group.manage(&_batched_mm_output); |
| 145 | |
| 146 | // Do not manage _input1 as it contains the weights |
| 147 | |
| 148 | // Configure input transform |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 149 | _input_transform.configure(compile_context, input, &_input0, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 150 | |
| 151 | // Configure filter transform |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 152 | _filter_transform->configure(compile_context, weights, &_input1, winograd_info); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 153 | |
| 154 | // Configure batched matrix multiply |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 155 | _batched_mm.configure(compile_context, &_input0, &_input1, nullptr, &_batched_mm_output, 1.0f, 0.0f, GEMMInfo(false, false, true /* Reshape weights only for the first run*/, 0, false, false, |
| 156 | GEMMLowpOutputStageInfo(), |
| 157 | (input->info()->data_type() == DataType::F16))); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 158 | |
| 159 | // Configure output transform |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 160 | _output_transform->configure(compile_context, &_batched_mm_output, biases, output, winograd_info, act_info); |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 161 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 162 | // Allocate temporary tensors |
| 163 | _input0.allocator()->allocate(); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 164 | _batched_mm_output.allocator()->allocate(); |
| 165 | } |
| 166 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 167 | 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] | 168 | const ActivationLayerInfo &act_info, bool enable_fast_math) |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 169 | { |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 170 | // Get indeces for the width and height |
| 171 | const size_t idx_width = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::WIDTH); |
| 172 | const size_t idx_height = get_data_layout_dimension_index(input->data_layout(), DataLayoutDimension::HEIGHT); |
| 173 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 174 | // Input shape, kernel size and output tile |
| 175 | const Size2D input_dims = Size2D(input->tensor_shape()[idx_width], input->tensor_shape()[idx_height]); |
| 176 | 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] | 177 | 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] | 178 | |
Pablo Tello | 7786f25 | 2019-08-05 12:22:37 +0100 | [diff] [blame] | 179 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_left() > (kernel_size.x() / 2u)) || (conv_info.pad_right() > (kernel_size.x() / 2u))), "Winograd only supports padding up to half kernel size"); |
| 180 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(((conv_info.pad_top() > (kernel_size.y() / 2u)) || (conv_info.pad_bottom() > (kernel_size.y() / 2u))), "Winograd only supports padding up to half kernel size"); |
| 181 | |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 182 | // Check if the Winograd configuration requires fast math |
| 183 | if(!enable_fast_math) |
| 184 | { |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 185 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32); //disable winograd for fp16 if fast math is false. |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 186 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); |
| 187 | } |
Gian Marco Iodice | e52a300 | 2018-04-11 15:59:10 +0100 | [diff] [blame] | 188 | |
| 189 | const WinogradInfo winograd_info = WinogradInfo(output_tile, |
Gian Marco Iodice | 2213d4b | 2018-04-27 10:39:06 +0100 | [diff] [blame] | 190 | kernel_size, |
| 191 | input_dims, |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 192 | conv_info, |
| 193 | input->data_layout()); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 194 | |
| 195 | // Validate input transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 196 | 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] | 197 | const TensorInfo input0 = input->clone()->set_tensor_shape(input0_shape); |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 198 | ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradInputTransform::validate(input, &input0, winograd_info)); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 199 | |
| 200 | // Validate filter transform |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 201 | 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] | 202 | const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 203 | ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradFilterTransformKernel::validate(weights, &input1, winograd_info)); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 204 | |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 205 | // Validate batched matrix multiply |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 206 | TensorShape batched_mm_output_shape = input0.tensor_shape(); |
| 207 | batched_mm_output_shape[0] = input1.tensor_shape()[0]; |
| 208 | const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); |
Vidhya Sudhan Loganathan | a25d16c | 2018-11-16 11:33:12 +0000 | [diff] [blame] | 209 | 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*/, 0, false, false, |
| 210 | GEMMLowpOutputStageInfo(), (input->data_type() == DataType::F16)))); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 211 | |
Gian Marco Iodice | 247f52c | 2018-03-22 11:24:56 +0000 | [diff] [blame] | 212 | // Configure output transform |
Gian Marco Iodice | 607339f | 2019-08-01 15:53:23 +0100 | [diff] [blame] | 213 | ARM_COMPUTE_RETURN_ON_ERROR(CLWinogradOutputTransformKernel::validate(&batched_mm_output, biases, output, winograd_info, act_info)); |
Isabella Gottardi | 3f217ec | 2018-02-12 14:59:19 +0000 | [diff] [blame] | 214 | |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 215 | return Status{}; |
| 216 | } |
| 217 | |
| 218 | void CLWinogradConvolutionLayer::run() |
| 219 | { |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 220 | prepare(); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 221 | |
Georgios Pinitas | da953f2 | 2019-04-02 17:27:03 +0100 | [diff] [blame] | 222 | MemoryGroupResourceScope scope_mg(_memory_group); |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 223 | |
| 224 | // Run input transform |
| 225 | _input_transform.run(); |
| 226 | |
| 227 | // Run batched matrix multiplication |
| 228 | _batched_mm.run(); |
| 229 | |
| 230 | // Run output transform |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 231 | CLScheduler::get().enqueue(*_output_transform); |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 232 | } |
Georgios Pinitas | 82b5148 | 2018-04-24 15:14:12 +0100 | [diff] [blame] | 233 | |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 234 | void CLWinogradConvolutionLayer::prepare() |
| 235 | { |
| 236 | if(!_is_prepared) |
| 237 | { |
| 238 | // Run filter transform and mark original weights as unused |
| 239 | _input1.allocator()->allocate(); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 240 | CLScheduler::get().enqueue(*_filter_transform, false); |
Georgios Pinitas | e043767 | 2018-05-02 14:07:55 +0100 | [diff] [blame] | 241 | _original_weights->mark_as_unused(); |
| 242 | |
| 243 | // Prepare GEMM and release reshaped weights if marked unused by CLGEMM |
| 244 | _batched_mm.prepare(); |
| 245 | if(!_input1.is_used()) |
| 246 | { |
| 247 | _input1.allocator()->free(); |
| 248 | } |
| 249 | |
| 250 | CLScheduler::get().queue().finish(); |
| 251 | _is_prepared = true; |
| 252 | } |
Gian Marco Iodice | d2fab73 | 2018-03-02 11:18:12 +0000 | [diff] [blame] | 253 | } |