1 /*
2 * Copyright (c) 2021-2022 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/cpu/operators/CpuWinogradConv2d.h"
25 #include "arm_compute/core/Error.h"
26 #include "arm_compute/core/Utils.h"
27 #include "arm_compute/core/Validate.h"
28 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
29 #include "arm_compute/core/utils/quantization/AsymmHelpers.h"
30 #include "arm_compute/runtime/FunctionDescriptors.h"
31 #include "arm_compute/runtime/NEON/NEScheduler.h"
32 #include "src/common/utils/Log.h"
33 #include "src/core/CPP/Validate.h"
34 #include "src/core/NEON/kernels/assembly/winograd.hpp"
35 #include "src/core/NEON/kernels/convolution/common/tensor.hpp"
36 #include "src/core/NEON/kernels/convolution/common/utils.hpp"
37 #include "src/core/helpers/MemoryHelpers.h"
38 #include "src/core/helpers/WindowHelpers.h"
39 #include "src/core/utils/AssemblyUtils.h"
40 #include "src/cpu/kernels/CpuWinogradConv2dKernel.h"
41 #include "src/cpu/kernels/assembly/arm_gemm.hpp"
42 #include "src/cpu/operators/CpuActivation.h"
43 #include "src/cpu/operators/CpuPermute.h"
44 #include "src/cpu/utils/CpuAuxTensorHandler.h"
45 #include "support/Cast.h"
46
47 namespace arm_compute
48 {
49 namespace cpu
50 {
51 using namespace arm_compute::experimental;
52 using namespace arm_compute::utils::cast;
53
54 namespace
55 {
internal_get_shape(const ITensorInfo * in)56 inline Tensor4DShape internal_get_shape(const ITensorInfo *in)
57 {
58 const DataLayout data_layout = in->data_layout();
59 const int in_width = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH));
60 const int in_height = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT));
61 const int in_channels = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL));
62 const int in_batches = in->dimension(get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES));
63
64 return Tensor4DShape{ in_batches, in_height, in_width, in_channels };
65 }
66
validate_arguments(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * dst,const PadStrideInfo & conv_info)67 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const PadStrideInfo &conv_info)
68 {
69 ARM_COMPUTE_UNUSED(dst, weights);
70 ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(src);
71
72 ARM_COMPUTE_RETURN_ERROR_ON_MSG(conv_info.stride().first != 1 || conv_info.stride().second != 1, "Winograd layer only supports unit strides.");
73 if(biases != nullptr)
74 {
75 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, biases);
76 ARM_COMPUTE_RETURN_ERROR_ON(biases->num_dimensions() > 1);
77 }
78 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F16, DataType::F32);
79 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(src, weights);
80 return Status{};
81 }
82
get_winograd_kernel_implementation(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,bool enable_fast_math,arm_conv::winograd::WinogradImpl * winograd_impl,std::unique_ptr<arm_conv::ConvolutionArgs> & conv_args)83 bool get_winograd_kernel_implementation(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst,
84 const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math,
85 arm_conv::winograd::WinogradImpl *winograd_impl, std::unique_ptr<arm_conv::ConvolutionArgs> &conv_args)
86 {
87 arm_conv::winograd::WinogradConfig winograd_cfg;
88 arm_gemm::GemmConfig cfg;
89
90 const DataType data_type = src->data_type();
91 Tensor4DShape in_shape{ internal_get_shape(src) };
92 Tensor4DShape out_shape{ internal_get_shape(dst) };
93 Tensor4DShape kernel_shape{ internal_get_shape(weights) };
94 uint32_t nthreads = NEScheduler::get().num_threads();
95 // Get configuration arguments for Winograd
96 winograd_cfg.output_rows = 0;
97 winograd_cfg.output_cols = 0;
98 conv_args = std::make_unique<arm_conv::ConvolutionArgs>(
99 in_shape.n_batches,
100 arm_conv::Shape2D{ static_cast<uint32_t>(in_shape.n_rows), static_cast<uint32_t>(in_shape.n_cols) },
101 in_shape.n_channels,
102 conv_info.pad_top(),
103 conv_info.pad_left(),
104 arm_conv::Shape2D{ static_cast<uint32_t>(out_shape.n_rows), static_cast<uint32_t>(out_shape.n_cols) },
105 out_shape.n_channels,
106 arm_conv::Shape2D{ static_cast<uint32_t>(kernel_shape.n_rows), static_cast<uint32_t>(kernel_shape.n_cols) },
107 assembly_utils::map_to_arm_gemm_activation(act_info));
108
109 bool success = false;
110 if(data_type == DataType::F32)
111 {
112 success = arm_conv::winograd::get_implementation<float>(
113 *winograd_impl, &CPUInfo::get(), *conv_args, nthreads, enable_fast_math, &winograd_cfg, nullptr);
114 }
115 #if defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
116 else if(data_type == DataType::F16)
117 {
118 success = arm_conv::winograd::get_implementation<__fp16>(
119 *winograd_impl, &CPUInfo::get(), *conv_args, nthreads, enable_fast_math, &winograd_cfg, nullptr);
120 }
121 #endif // defined(__aarch64__) && defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
122 else
123 {
124 success = false;
125 }
126 return success;
127 }
fuse_function_supported(const ActivationLayerInfo & act_info)128 inline bool fuse_function_supported(const ActivationLayerInfo &act_info)
129 {
130 return act_info.activation() == ActivationLayerInfo::ActivationFunction::RELU || act_info.activation() == ActivationLayerInfo::ActivationFunction::BOUNDED_RELU;
131 }
132 } // namespace
133
CpuWinogradConv2d()134 CpuWinogradConv2d::CpuWinogradConv2d()
135
136 : _gemm_function(std::make_unique<CpuGemm>()),
137 _activation_func(std::make_unique<CpuActivation>()),
138 _transform_input_kernel(nullptr),
139 _transform_output_kernel(nullptr),
140 _permute_input(std::make_unique<CpuPermute>()),
141 _permute_output(std::make_unique<CpuPermute>()),
142 _permute_weights(std::make_unique<CpuPermute>()),
143 _aux_mem(AuxTensorIdx::Count),
144 _conv_args{ nullptr },
145 _winograd_impl{},
146 _data_layout(),
147 _winograd_transformed_input{},
148 _winograd_transformed_output{},
149 _winograd_transformed_weights{},
150 _input_workspace(),
151 _output_workspace(),
152 _weights_hwio(),
153 _input_nhwc(),
154 _output_nhwc(),
155 _is_prepared{ false },
156 _run_activation{ false }
157 {
158 }
159
160 CpuWinogradConv2d::~CpuWinogradConv2d() = default;
161
configure(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,bool enable_fast_math)162 void CpuWinogradConv2d::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst,
163 const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math)
164 {
165 ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
166 ARM_COMPUTE_ERROR_THROW_ON(validate(src, weights, biases, dst, conv_info, act_info, enable_fast_math));
167 ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, conv_info, act_info, enable_fast_math);
168 ARM_COMPUTE_UNUSED(biases);
169 const DataType data_type = src->data_type();
170 uint32_t nthreads = NEScheduler::get().num_threads();
171 _data_layout = src->data_layout();
172 const Tensor4DShape kernel_shape{ internal_get_shape(weights) };
173
174 bool success = get_winograd_kernel_implementation(src, weights, dst, conv_info, act_info, enable_fast_math, &_winograd_impl, _conv_args);
175
176 ARM_COMPUTE_EXIT_ON_MSG_VAR(!success, "Unsupported kernel size: %d x %d.\n", kernel_shape.n_rows, kernel_shape.n_cols);
177 ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using input transform: %s\n", _winograd_impl.input_transform->get_name().c_str());
178 ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using weight transform: %s\n", _winograd_impl.input_transform->get_name().c_str());
179 ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using output transform: %s\n", _winograd_impl.input_transform->get_name().c_str());
180
181 const bool has_impl = ((_winograd_impl.input_transform != nullptr) && (_winograd_impl.output_transform != nullptr) && (_winograd_impl.gemm_args != nullptr));
182 if(has_impl)
183 {
184 // Determine how much working space is required, allocate it.
185 const size_t input_workspace_size = _winograd_impl.input_transform->get_working_space_size(*_conv_args, nthreads);
186 const size_t output_workspace_size = _winograd_impl.output_transform->get_working_space_size(*_conv_args, nthreads);
187
188 TensorInfo input_workspace_info(TensorShape(input_workspace_size), 1, DataType::U8);
189 TensorInfo output_workspace_info(TensorShape(output_workspace_size), 1, DataType::U8);
190 _input_workspace = input_workspace_info;
191 _output_workspace = output_workspace_info;
192
193 const auto &wds = _winograd_impl.winograd_spec;
194
195 // Preparing winograd transformed input tensor
196 const size_t data_type_size = src->element_size();
197 const uint32_t m = _winograd_impl.gemm_args->_Msize; // Total number of tiles
198 const uint32_t k = _winograd_impl.gemm_args->_Ksize; // Input channels
199 const uint32_t n = _winograd_impl.gemm_args->_Nsize; // Output channels
200 const uint32_t n_gemms = _winograd_impl.gemm_args->_nmulti;
201 const uint32_t n_batches = _winograd_impl.gemm_args->_nbatches;
202 constexpr size_t storage_alignment = 64;
203
204 const TensorShape a_shape(k, m, n_batches, n_gemms);
205 Strides a_strides(data_type_size);
206 a_strides.set(1, data_type_size * _winograd_impl.winograd_spec.input_ld_row);
207 a_strides.set(2, data_type_size * _winograd_impl.winograd_spec.input_ld_batch);
208 a_strides.set(3, data_type_size * _winograd_impl.winograd_spec.input_ld_matrix);
209
210 const TensorShape b_shape(n, k, n_gemms);
211 Strides b_strides(data_type_size);
212 b_strides.set(1, data_type_size * _winograd_impl.winograd_spec.weight_ld_row);
213 b_strides.set(2, data_type_size * _winograd_impl.winograd_spec.weight_ld_matrix);
214
215 const TensorShape d_shape(n, m, n_batches, n_gemms);
216 Strides d_strides(data_type_size);
217 d_strides.set(1, data_type_size * _winograd_impl.winograd_spec.output_ld_row);
218 d_strides.set(2, data_type_size * _winograd_impl.winograd_spec.output_ld_batch);
219 d_strides.set(3, data_type_size * _winograd_impl.winograd_spec.output_ld_matrix);
220
221 TensorInfo a_info{};
222 TensorInfo b_info{};
223 TensorInfo d_info{};
224 a_info.init(a_shape, 1, data_type, a_strides, 0, wds.input_matrix_size_bytes);
225 b_info.init(b_shape, 1, data_type, b_strides, 0, wds.weight_matrix_size_bytes);
226 d_info.init(d_shape, 1, data_type, d_strides, 0, wds.output_matrix_size_bytes);
227
228 _winograd_transformed_input = a_info;
229 _winograd_transformed_weights = b_info;
230 _winograd_transformed_output = d_info;
231
232 PermutationVector weights_permutation_vector(3U, 0U, 1U, 2U);
233
234 // Configure the kernel to transform the input tensor from NCHW -> NHWC
235 if(_data_layout == DataLayout::NCHW)
236 {
237 _permute_input->configure(src, &_input_nhwc, PermutationVector(2U, 0U, 1U));
238 weights_permutation_vector = PermutationVector(3U, 2U, 0U, 1U);
239 }
240
241 // Re-order a weight tensor from [Output feature map x Input feature map x Height x Width] to [Height x Width x Input feature map x Output feature map]
242 _permute_weights->configure(weights, &_weights_hwio, weights_permutation_vector);
243
244 // Reorder the convoluted output to ACL's ordering NCHW
245 if(_data_layout == DataLayout::NCHW)
246 {
247 // configure and allocate dst tensor to be used to convert from winograd domain to spatial domain when calling to reshape_output()
248 TensorInfo info(TensorShape(dst->dimension(2), dst->dimension(0),
249 dst->dimension(1), dst->dimension(3)),
250 1, dst->data_type());
251 _output_nhwc = info;
252 _permute_output->configure(&_output_nhwc, dst, PermutationVector(1U, 2U, 0U));
253 }
254
255 // Configure input transform kernel
256 _transform_input_kernel = std::make_unique<CpuWinogradConv2dTransformInputKernel>(_winograd_impl, *_conv_args, nthreads);
257
258 // Configure GEMM function
259 _gemm_function->configure(&_winograd_transformed_input, &_winograd_transformed_weights, nullptr, &_winograd_transformed_output, 1.0f, 0.f);
260
261 // Configure output transform kernel
262 _transform_output_kernel = std::make_unique<CpuWinogradConv2dTransformOutputKernel>(_winograd_impl, *_conv_args, nthreads);
263
264 //Configure Activation Layer
265 _run_activation = act_info.enabled() && !fuse_function_supported(act_info);
266 if(_run_activation)
267 {
268 _activation_func->configure(dst, nullptr, act_info);
269 }
270
271 auto asm_mem_req = _gemm_function->workspace();
272 _aux_mem[GemmWorkspace] = asm_mem_req[GemmWorkspace];
273 _aux_mem[Pretranspose] = asm_mem_req[Pretranspose];
274 _aux_mem[InterleavedLHS] = asm_mem_req[InterleavedLHS];
275 _aux_mem[TransposedRHS] = asm_mem_req[TransposedRHS];
276 _aux_mem[TempResult] = asm_mem_req[TempResult];
277
278 // Request temporary memory. Overlap memory needed for Input/Output transformations as they run on different non-overlapping time-steps.
279 _aux_mem[TransformedInput] = MemoryInfo(offset_int_vec(TransformedInput), MemoryLifetime::Temporary, wds.input_matrix_size_bytes, storage_alignment);
280 _aux_mem[TransformedOutput] = MemoryInfo(offset_int_vec(TransformedOutput), MemoryLifetime::Temporary, wds.output_matrix_size_bytes, storage_alignment);
281 _aux_mem[WorkspaceIO] = MemoryInfo(offset_int_vec(WorkspaceIO), MemoryLifetime::Temporary, std::max(input_workspace_size, output_workspace_size));
282 _aux_mem[PermutedWeights] = MemoryInfo(offset_int_vec(PermutedWeights), MemoryLifetime::Prepare, _weights_hwio.total_size());
283 _aux_mem[TransformedWeights] = MemoryInfo(offset_int_vec(TransformedWeights), MemoryLifetime::Persistent, wds.weight_matrix_size_bytes, storage_alignment);
284 if(_data_layout == DataLayout::NCHW)
285 {
286 _aux_mem[PermutedInput].merge(offset_int_vec(PermutedInput), src->total_size());
287 _aux_mem[PermutedOutput].merge(offset_int_vec(PermutedOutput), dst->total_size());
288 }
289 }
290 }
validate(const ITensorInfo * src,const ITensorInfo * weights,const ITensorInfo * biases,const ITensorInfo * dst,const PadStrideInfo & conv_info,const ActivationLayerInfo & act_info,bool enable_fast_math)291 Status CpuWinogradConv2d::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
292 const PadStrideInfo &conv_info, const ActivationLayerInfo &act_info, bool enable_fast_math)
293 {
294 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst);
295 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(src, weights, biases, dst, conv_info));
296
297 // Disable winograd for fp16 if fast math is false.
298 if(!enable_fast_math)
299 {
300 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(src, 1, DataType::F32);
301 }
302
303 const Tensor4DShape kernel_shape{ internal_get_shape(weights) };
304 arm_conv::winograd::WinogradImpl winograd_impl{};
305
306 std::unique_ptr<arm_conv::ConvolutionArgs> conv_args;
307 const bool success = get_winograd_kernel_implementation(src, weights, dst, conv_info, act_info, enable_fast_math, &winograd_impl, conv_args);
308
309 ARM_COMPUTE_RETURN_ERROR_ON_MSG_VAR(success == false, "Unsupported kernel size: %d x %d.\n", kernel_shape.n_rows, kernel_shape.n_cols);
310 ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using input transform: %s\n", winograd_impl.input_transform->get_name().c_str());
311 ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using weight transform: %s\n", winograd_impl.input_transform->get_name().c_str());
312 ARM_COMPUTE_LOG_MSG_WITH_FORMAT_ACL(arm_compute::logging::LogLevel::INFO, "Using output transform: %s\n", winograd_impl.input_transform->get_name().c_str());
313 return Status{};
314 }
315
run(ITensorPack & tensors)316 void CpuWinogradConv2d::run(ITensorPack &tensors)
317 {
318 prepare(tensors);
319 auto src = tensors.get_const_tensor(ACL_SRC_0);
320 auto biases = tensors.get_const_tensor(ACL_SRC_2);
321 auto output = tensors.get_tensor(ACL_DST);
322 Window win;
323
324 const uint32_t nthreads = NEScheduler::get().num_threads();
325
326 // The Winograd transform implementation does fine-grain threading inside the transforms. Just pass thread_id and nthreads.
327 win.set(Window::DimX, Window::Dimension(0, nthreads, 1));
328
329 // Wrap the winograd-domain tensorInfos created in configuration in tensors and allocate the required memory.
330 CpuAuxTensorHandler input_nhwc(offset_int_vec(PermutedInput), _input_nhwc, tensors, true);
331 CpuAuxTensorHandler winograd_input_transformed(offset_int_vec(TransformedInput), _winograd_transformed_input, tensors, true);
332 CpuAuxTensorHandler input_workspace(offset_int_vec(WorkspaceIO), _input_workspace, tensors, true);
333 const bool is_nchw = _data_layout == DataLayout::NCHW;
334 if(is_nchw)
335 {
336 //Bring channels to the front as Winograd code expects the tensor to be in the format NHWC
337 ITensorPack pack{ { ACL_SRC, src }, { ACL_DST, input_nhwc.get() } };
338 _permute_input->run(pack);
339 }
340
341 CpuAuxTensorHandler winograd_output_transformed(offset_int_vec(TransformedOutput), _winograd_transformed_output, tensors, true);
342 CpuAuxTensorHandler output_workspace(offset_int_vec(WorkspaceIO), _output_workspace, tensors, true);
343 CpuAuxTensorHandler output_nhwc(offset_int_vec(PermutedOutput), _output_nhwc, tensors, true);
344
345 ITensorPack transform_input_pack{ { ACL_SRC, is_nchw ? input_nhwc.get() : src }, { ACL_DST, winograd_input_transformed.get() }, { ACL_INT, input_workspace.get() } };
346 NEScheduler::get().schedule_op(_transform_input_kernel.get(), Window::DimX, win, transform_input_pack);
347
348 CpuAuxTensorHandler winograd_weights_transformed(offset_int_vec(TransformedWeights), _winograd_transformed_weights, tensors, true);
349
350 // Run 16 GEMMs in multiple threads, each kernel runs one or more GEMMs
351 ITensorPack gemm_pack = tensors;
352 gemm_pack.add_const_tensor(ACL_SRC, winograd_input_transformed.get());
353 gemm_pack.add_const_tensor(ACL_SRC_1, winograd_weights_transformed.get());
354 gemm_pack.add_const_tensor(ACL_BIAS, nullptr);
355 gemm_pack.add_tensor(ACL_DST, winograd_output_transformed.get());
356 _gemm_function->run(gemm_pack);
357
358 // Output transform
359 ITensorPack transform_output_pack{ { ACL_SRC_0, winograd_output_transformed.get() }, { ACL_DST, is_nchw ? output_nhwc.get() : output }, { ACL_SRC_1, biases }, { ACL_INT, output_workspace.get() } };
360 NEScheduler::get().schedule_op(_transform_output_kernel.get(), Window::DimX, win, transform_output_pack);
361 if(is_nchw)
362 {
363 // Reorder the convoluted output to ACL's ordering NCHW
364 ITensorPack pack{ { ACL_SRC, output_nhwc.get() }, { ACL_DST, output } };
365 _permute_output->run(pack);
366 }
367 if(_run_activation)
368 {
369 ITensorPack pack{ { ACL_SRC, output }, { ACL_DST, output } };
370 _activation_func->run(pack);
371 }
372 }
373
prepare(ITensorPack & tensors)374 void CpuWinogradConv2d::prepare(ITensorPack &tensors)
375 {
376 if(!_is_prepared)
377 {
378 const ITensor *weights = tensors.get_const_tensor(ACL_SRC_1);
379 ITensor *weights_aux = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(PermutedWeights)));
380
381 CpuAuxTensorHandler permuted_weights(_weights_hwio, *weights_aux);
382 ITensorPack permute_tensors{ { ACL_SRC, weights }, { ACL_DST, permuted_weights.get() } };
383 _permute_weights->run(permute_tensors);
384 const int element_size_in_bytes = permuted_weights.get()->info()->element_size();
385 // Weights were in OHWI format, before being permuted "permuted_weights" to be in HWIO format.
386 const unsigned int height_idx = 3; // H in HWIO
387 const unsigned int width_idx = 2; // W in HWIO
388 const unsigned int channel_idx = 1; // I in HWIO
389
390 const int permuted_weight_row_stride = permuted_weights.get()->info()->strides_in_bytes()[height_idx] / element_size_in_bytes;
391 const int permuted_weight_col_stride = permuted_weights.get()->info()->strides_in_bytes()[width_idx] / element_size_in_bytes;
392 const int permuted_weight_channel_stride = permuted_weights.get()->info()->strides_in_bytes()[channel_idx] / element_size_in_bytes;
393
394 // Wrap the winograd-domain transformed weight TensorInfo in Auxiliary tensor and allocate the required memory.
395 ITensor *weights_transf = utils::cast::polymorphic_cast<ITensor *>(tensors.get_tensor(offset_int_vec(TransformedWeights)));
396 ARM_COMPUTE_ERROR_ON_NULLPTR(weights_transf);
397 CpuAuxTensorHandler winograd_transformed_weights(_winograd_transformed_weights, *weights_transf);
398
399 const void *permuted_weights_ptr;
400 void *win_wght_transf_ptr;
401
402 permuted_weights_ptr = reinterpret_cast<const void *>(permuted_weights.get()->buffer() + permuted_weights.get()->info()->offset_first_element_in_bytes());
403 win_wght_transf_ptr = reinterpret_cast<void *>(winograd_transformed_weights.get()->buffer() + winograd_transformed_weights.get()->info()->offset_first_element_in_bytes());
404
405 // Prepare Weights
406 _winograd_impl.weight_transform->execute(
407 *_conv_args,
408 permuted_weights_ptr,
409 permuted_weight_row_stride,
410 permuted_weight_col_stride,
411 permuted_weight_channel_stride,
412 win_wght_transf_ptr,
413 _winograd_impl.winograd_spec,
414 0, 1 // Thread 1 of 1
415 );
416 ITensorPack gemm_pack = tensors;
417 gemm_pack.add_const_tensor(ACL_SRC_1, winograd_transformed_weights.get());
418 _gemm_function->prepare(gemm_pack);
419 _is_prepared = 1;
420 }
421 }
workspace() const422 experimental::MemoryRequirements CpuWinogradConv2d::workspace() const
423 {
424 return _aux_mem;
425 }
426
427 } // namespace cpu
428 } // namespace arm_compute
429