Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2018-2021 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 "src/runtime/gpu/cl/operators/ClWinogradConv2d.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/experimental/Types.h" |
| 30 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 31 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 32 | #include "src/core/CL/kernels/CLFillBorderKernel.h" |
| 33 | #include "src/core/CL/kernels/CLFillBorderKernel.h" |
| 34 | #include "src/core/gpu/cl/kernels/ClWinogradFilterTransformKernel.h" |
| 35 | #include "src/core/gpu/cl/kernels/ClWinogradInputTransformKernel.h" |
| 36 | #include "src/core/gpu/cl/kernels/ClWinogradOutputTransformKernel.h" |
| 37 | #include "src/core/helpers/MemoryHelpers.h" |
| 38 | #include "src/runtime/gpu/cl/utils/ClAuxTensorHandler.h" |
| 39 | #include "support/Cast.h" |
| 40 | |
| 41 | using namespace arm_compute::experimental; |
| 42 | |
| 43 | namespace arm_compute |
| 44 | { |
| 45 | namespace opencl |
| 46 | { |
| 47 | namespace |
| 48 | { |
| 49 | Size2D winograd_output_tile(const Size2D &input_dims, const Size2D &kernel_dims, DataLayout data_layout) |
| 50 | { |
| 51 | Size2D output_tile = Size2D{}; |
| 52 | |
| 53 | const unsigned int kernel_max_dim = std::max(kernel_dims.width, kernel_dims.height); |
| 54 | |
| 55 | // Check if the input spatial dimensions are smaller than 4 |
| 56 | const bool is_input_lt4_nchw = (input_dims.width <= 4 && input_dims.height <= 4) && (data_layout == DataLayout::NCHW); |
| 57 | |
| 58 | if(kernel_max_dim == 3U) |
| 59 | { |
| 60 | if(kernel_dims == Size2D(3U, 3U)) |
| 61 | { |
| 62 | output_tile = is_input_lt4_nchw ? Size2D(2U, 2U) : Size2D(4U, 4U); |
| 63 | } |
| 64 | else if(kernel_dims == Size2D(3U, 1U)) |
| 65 | { |
| 66 | output_tile = is_input_lt4_nchw ? Size2D(2U, 1U) : Size2D(4U, 1U); |
| 67 | } |
| 68 | else |
| 69 | { |
| 70 | output_tile = is_input_lt4_nchw ? Size2D(1U, 2U) : Size2D(1U, 4U); |
| 71 | } |
| 72 | } |
| 73 | else if(kernel_max_dim == 5U) |
| 74 | { |
| 75 | output_tile = Size2D(kernel_dims.width == 1 ? 1U : 4U, |
| 76 | kernel_dims.height == 1 ? 1U : 4U); |
| 77 | } |
| 78 | else if(kernel_max_dim == 7U) |
| 79 | { |
| 80 | output_tile = Size2D(kernel_dims.width == 1 ? 1U : 2U, |
| 81 | kernel_dims.height == 1 ? 1U : 2U); |
| 82 | } |
| 83 | |
| 84 | return output_tile; |
| 85 | } |
| 86 | |
| 87 | bool check_support_fast_math(const Size2D &output_tile, const Size2D &kernel_size) |
| 88 | { |
| 89 | // Check if we want to configure a Winograd configuration which requires fast math |
| 90 | using WinogradConfiguration = std::pair<std::pair<int, int>, std::pair<int, int>>; |
| 91 | |
| 92 | std::vector<WinogradConfiguration> fast_math_winograd = |
| 93 | { |
| 94 | WinogradConfiguration(std::pair<int, int>(4, 4), std::pair<int, int>(5, 5)), |
| 95 | WinogradConfiguration(std::pair<int, int>(2, 2), std::pair<int, int>(7, 7)) |
| 96 | }; |
| 97 | |
| 98 | auto p = std::make_pair(std::pair<int, int>(output_tile.width, output_tile.height), |
| 99 | std::pair<int, int>(kernel_size.width, kernel_size.height)); |
| 100 | |
| 101 | return std::find(fast_math_winograd.begin(), fast_math_winograd.end(), p) != fast_math_winograd.end(); |
| 102 | } |
| 103 | |
| 104 | Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, |
| 105 | const ActivationLayerInfo &act_info, bool enable_fast_math) |
| 106 | { |
| 107 | // Get indeces for the width and height |
| 108 | const size_t idx_width = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH); |
| 109 | const size_t idx_height = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT); |
| 110 | |
| 111 | // Input shape, kernel size and output tile |
| 112 | const Size2D input_dims = Size2D(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height]); |
| 113 | const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); |
| 114 | const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, src->data_layout()); |
| 115 | |
| 116 | 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"); |
| 117 | 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"); |
| 118 | |
| 119 | // Check if the Winograd configuration requires fast math |
| 120 | if(!enable_fast_math) |
| 121 | { |
| 122 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); //disable winograd for fp16 if fast math is false. |
| 123 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); |
| 124 | } |
| 125 | |
| 126 | const WinogradInfo winograd_info = WinogradInfo(output_tile, |
| 127 | kernel_size, |
| 128 | input_dims, |
| 129 | conv_info, |
| 130 | src->data_layout()); |
| 131 | |
| 132 | // Validate input transform |
| 133 | const TensorShape input0_shape = misc::shape_calculator::compute_winograd_input_transform_shape(*src, winograd_info); |
| 134 | const TensorInfo input0 = src->clone()->set_tensor_shape(input0_shape); |
| 135 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWinogradInputTransformKernel::validate(src, &input0, winograd_info)); |
| 136 | |
| 137 | // Validate filter transform |
| 138 | const TensorShape input1_shape = misc::shape_calculator::compute_winograd_filter_transform_shape(*weights, winograd_info); |
| 139 | const TensorInfo input1 = weights->clone()->set_tensor_shape(input1_shape); |
| 140 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWinogradFilterTransformKernel::validate(weights, &input1, winograd_info)); |
| 141 | |
| 142 | // Validate batched matrix multiply |
| 143 | TensorShape batched_mm_output_shape = input0.tensor_shape(); |
| 144 | batched_mm_output_shape[0] = input1.tensor_shape()[0]; |
| 145 | const TensorInfo batched_mm_output = input0.clone()->set_tensor_shape(batched_mm_output_shape); |
| 146 | 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, |
| 147 | GEMMLowpOutputStageInfo(), (src->data_type() == DataType::F16)))); |
| 148 | |
| 149 | // Configure output transform |
| 150 | ARM_COMPUTE_RETURN_ON_ERROR(kernels::ClWinogradOutputTransformKernel::validate(&batched_mm_output, biases, dst, winograd_info, act_info)); |
| 151 | return Status{}; |
| 152 | } |
| 153 | |
| 154 | } // namespace |
| 155 | |
| 156 | ClWinogradConv2d::ClWinogradConv2d() |
| 157 | : _batched_mm(), |
| 158 | _input_transform(std::make_unique<kernels::ClWinogradInputTransformKernel>()), |
| 159 | _filter_transform(std::make_unique<kernels::ClWinogradFilterTransformKernel>()), |
| 160 | _output_transform(std::make_unique<kernels::ClWinogradOutputTransformKernel>()), |
| 161 | _border_handler(), |
| 162 | _input0(), |
| 163 | _input1(), |
| 164 | _batched_mm_output(), |
| 165 | _is_prepared(false), |
| 166 | _aux_mem() |
| 167 | { |
| 168 | } |
| 169 | |
| 170 | ClWinogradConv2d::~ClWinogradConv2d() = default; |
| 171 | |
| 172 | void ClWinogradConv2d::configure(const ClCompileContext &compile_context, ITensorInfo *src, ITensorInfo *weights, ITensorInfo *biases, ITensorInfo *dst, |
| 173 | const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math) |
| 174 | { |
| 175 | ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, weights, biases, dst, conv_info, act_info, enable_fast_math)); |
| 176 | // Get indices for the width and height |
| 177 | const size_t idx_width = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::WIDTH); |
| 178 | const size_t idx_height = get_data_layout_dimension_index(src->data_layout(), DataLayoutDimension::HEIGHT); |
| 179 | |
| 180 | // Input shape, kernel size and output tile |
| 181 | const Size2D input_dims = Size2D(src->tensor_shape()[idx_width], src->tensor_shape()[idx_height]); |
| 182 | const Size2D kernel_size = Size2D(weights->tensor_shape()[idx_width], weights->tensor_shape()[idx_height]); |
| 183 | const Size2D output_tile = winograd_output_tile(input_dims, kernel_size, src->data_layout()); |
| 184 | |
| 185 | // Check if the Winograd configuration requires fast math |
| 186 | if(!enable_fast_math) |
| 187 | { |
| 188 | ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32); //disable winograd for fp16 if fast math is false. |
| 189 | ARM_COMPUTE_ERROR_ON_MSG(check_support_fast_math(output_tile, kernel_size), "This Winograd configuration requires enable_fast_math=true"); |
| 190 | } |
| 191 | const WinogradInfo winograd_info = WinogradInfo(output_tile, |
| 192 | kernel_size, |
| 193 | input_dims, |
| 194 | conv_info, |
| 195 | src->data_layout()); |
| 196 | |
| 197 | _is_prepared = false; |
| 198 | |
| 199 | // Configure input transform |
| 200 | _input_transform->configure(compile_context, src, &_input0, winograd_info); |
| 201 | _border_handler.configure(compile_context, src, _input_transform->border_size(), BorderMode::CONSTANT, PixelValue()); |
| 202 | |
| 203 | // Configure filter transform |
| 204 | _filter_transform->configure(compile_context, weights, &_input1, winograd_info); |
| 205 | |
| 206 | // Configure batched matrix multiply |
| 207 | _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, |
| 208 | false, false, |
| 209 | GEMMLowpOutputStageInfo(), |
| 210 | (src->data_type() == DataType::F16))); |
| 211 | |
| 212 | // Configure output transform |
| 213 | _output_transform->configure(compile_context, &_batched_mm_output, biases, dst, winograd_info, act_info); |
| 214 | |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 215 | _aux_mem = _batched_mm.workspace(); |
| 216 | const MemoryLifetime wino_wei_lifetm = std::any_of(std::begin(_aux_mem), std::end(_aux_mem), [](const auto & r) |
| 217 | { |
| 218 | return (r.lifetime == MemoryLifetime::Persistent) && (r.size > 0); |
| 219 | }) ? |
| 220 | MemoryLifetime::Prepare : |
| 221 | MemoryLifetime::Persistent; |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 222 | _aux_mem.push_back(MemoryInfo(offset_int_vec(2), MemoryLifetime::Temporary, _input0.total_size())); |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 223 | _aux_mem.push_back(MemoryInfo(offset_int_vec(3), wino_wei_lifetm, _input1.total_size())); |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 224 | _aux_mem.push_back(MemoryInfo(offset_int_vec(4), MemoryLifetime::Temporary, _batched_mm_output.total_size())); |
| 225 | } |
| 226 | |
| 227 | Status ClWinogradConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info, |
| 228 | const ActivationLayerInfo &act_info, bool enable_fast_math) |
| 229 | { |
| 230 | ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info, act_info, enable_fast_math)); |
| 231 | return Status{}; |
| 232 | } |
| 233 | |
| 234 | void ClWinogradConv2d::run(ITensorPack &tensors) |
| 235 | { |
| 236 | prepare(tensors); |
| 237 | |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 238 | auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_0)); |
| 239 | auto biases = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_2)); |
| 240 | auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST)); |
| 241 | |
| 242 | CLAuxTensorHandler input0(offset_int_vec(2), _input0, tensors, true); |
| 243 | CLAuxTensorHandler input1(offset_int_vec(3), _input1, tensors, true); |
| 244 | CLAuxTensorHandler batched_mm_output(offset_int_vec(4), _batched_mm_output, tensors, true); |
| 245 | |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 246 | // Run input transform |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 247 | ITensorPack pack_it |
| 248 | { |
| 249 | { TensorType::ACL_SRC, src }, |
| 250 | { TensorType::ACL_DST, input0.get() }, |
| 251 | }; |
| 252 | CLScheduler::get().enqueue_op(_border_handler, pack_it); |
| 253 | CLScheduler::get().enqueue_op(*_input_transform, pack_it); |
| 254 | |
| 255 | // Run batched matrix multiplication |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 256 | ITensorPack pack_mm = tensors; |
| 257 | pack_mm.add_const_tensor(TensorType::ACL_SRC_0, input0.get()); |
| 258 | pack_mm.add_tensor(TensorType::ACL_DST, batched_mm_output.get()); |
| 259 | if(_aux_mem[3].lifetime == MemoryLifetime::Prepare) |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 260 | { |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 261 | pack_mm.remove_tensor(TensorType::ACL_SRC_1); |
| 262 | } |
| 263 | else |
| 264 | { |
| 265 | pack_mm.add_const_tensor(TensorType::ACL_SRC_1, input1.get()); |
| 266 | } |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 267 | _batched_mm.run(pack_mm); |
| 268 | |
| 269 | // Run output transform |
| 270 | ITensorPack pack_ot |
| 271 | { |
| 272 | { TensorType::ACL_SRC_0, batched_mm_output.get() }, |
| 273 | { TensorType::ACL_SRC_1, biases }, |
| 274 | { TensorType::ACL_DST, dst }, |
| 275 | }; |
| 276 | CLScheduler::get().enqueue_op(*_output_transform, pack_ot); |
| 277 | } |
| 278 | |
| 279 | void ClWinogradConv2d::prepare(ITensorPack &tensors) |
| 280 | { |
| 281 | if(!_is_prepared) |
| 282 | { |
| 283 | auto weights = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC_1)); |
| 284 | ICLTensor *in1_aux = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(offset_int_vec(3))); |
| 285 | |
| 286 | CLAuxTensorHandler input1(_input1, *in1_aux); |
| 287 | ITensorPack pack_ft |
| 288 | { |
| 289 | { TensorType::ACL_SRC, weights }, |
| 290 | { TensorType::ACL_DST, input1.get() }, |
| 291 | }; |
| 292 | // Run filter transform and mark original weights as unused |
| 293 | CLScheduler::get().enqueue_op(*_filter_transform, pack_ft, false); |
| 294 | weights->mark_as_unused(); |
| 295 | |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 296 | // Prepare GEMM and release reshaped weights if marked unused by ClGemm |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 297 | ITensorPack mm_prepare_pack = tensors; |
| 298 | mm_prepare_pack.add_tensor(ACL_SRC_1, input1.get()); |
| 299 | _batched_mm.prepare(mm_prepare_pack); |
Manuel Bottini | c6f4ec3 | 2021-05-18 18:41:56 +0100 | [diff] [blame] | 300 | |
| 301 | CLScheduler::get().queue().finish(); |
| 302 | _is_prepared = true; |
| 303 | } |
| 304 | } |
| 305 | |
| 306 | experimental::MemoryRequirements ClWinogradConv2d::workspace() const |
| 307 | { |
| 308 | return _aux_mem; |
| 309 | } |
| 310 | } // namespace opencl |
| 311 | } // namespace arm_compute |