Compute Library
 22.11
FunctionHelpers.h
Go to the documentation of this file.
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 #ifndef ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_FUNCTION_HELPERS_H
25 #define ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_FUNCTION_HELPERS_H
26 
27 #include "arm_compute/core/experimental/IPostOp.h"
28 #include "arm_compute/core/experimental/PostOps.h"
29 #include "arm_compute/graph/Logger.h"
30 #include "arm_compute/graph/Tensor.h"
31 #include "arm_compute/graph/TypePrinter.h"
32 #include "arm_compute/graph/Types.h"
33 #include "arm_compute/graph/Utils.h"
34 #include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationFunction.h"
35 #include "arm_compute/graph/backends/FusedConvolutionBatchNormalizationWithPostOpsFunction.h"
36 #include "arm_compute/graph/backends/FusedDepthwiseConvolutionBatchNormalizationFunction.h"
37 #include "arm_compute/graph/backends/Utils.h"
38 #include "arm_compute/graph/nodes/Nodes.h"
39 
40 #include "arm_compute/core/Error.h"
41 #include "arm_compute/core/Helpers.h"
42 #include "arm_compute/core/ITensorInfo.h"
43 #include "support/Cast.h"
44 
45 namespace arm_compute
46 {
47 namespace graph
48 {
49 namespace backends
50 {
51 namespace detail
52 {
53 /** Returns backing tensor of a given tensor
54  *
55  * @tparam TargetInfo Target information
56  *
57  * @param[in] tensor Tensor to extract the backing tensor from
58  *
59  * @return Backing tensor if present else nullptr
60  */
61 template <typename TargetInfo>
62 typename TargetInfo::TensorType *get_backing_tensor(arm_compute::graph::Tensor *tensor)
63 {
64  typename TargetInfo::TensorType *backing_tensor = nullptr;
65  if(tensor != nullptr)
66  {
67  ARM_COMPUTE_ERROR_ON(tensor->desc().target != TargetInfo::TargetType);
68  // Get backing tensor handle
69  ITensorHandle *tensor_handle = tensor->handle();
70  // Get backing tensor
71  backing_tensor = (tensor_handle != nullptr) ? arm_compute::utils::cast::polymorphic_cast<typename TargetInfo::TensorType *>(&tensor_handle->tensor()) : nullptr;
72  }
73 
74  return backing_tensor;
75 }
76 
77 template <typename TargetInfo>
78 void validate_node(const INode &node, size_t num_expected_inputs, size_t num_expected_outputs)
79 {
80  ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating " << node.type()
81  << " Target: " << TargetInfo::TargetType
82  << " ID: " << node.id()
83  << node.name()
84  << std::endl);
85 
86  ARM_COMPUTE_ERROR_ON(TargetInfo::TargetType != node.assigned_target());
87  ARM_COMPUTE_ERROR_ON(node.num_inputs() != num_expected_inputs);
88  ARM_COMPUTE_ERROR_ON(node.num_outputs() != num_expected_outputs);
89  ARM_COMPUTE_UNUSED(node, num_expected_inputs, num_expected_outputs);
90 }
91 
92 /** Creates a backend activation layer function
93  *
94  * @tparam ActivationLayerFunction Backend activation function
95  * @tparam TargetInfo Target-specific information
96  *
97  * @param[in] node Node to create the backend function for
98  *
99  * @return Backend activation layer function
100  */
101 template <typename ActivationLayerFunction, typename TargetInfo>
102 std::unique_ptr<IFunction> create_activation_layer(ActivationLayerNode &node)
103 {
104  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
105 
106  // Extract IO and info
107  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
108  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
109  const ActivationLayerInfo act_info = node.activation_info();
110 
111  // Create function
112  auto func = std::make_unique<ActivationLayerFunction>();
113  func->configure(input, output, act_info);
114 
115  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
116  << node.name()
117  << " Type: " << node.type()
118  << " Target: " << TargetInfo::TargetType
119  << " Data Type: " << input->info()->data_type()
120  << " Shape: " << input->info()->tensor_shape()
121  << " Activation function: " << act_info.activation()
122  << " a: " << act_info.a()
123  << " b: " << act_info.b()
124  << " InPlace : " << is_in_place_operation(input, output)
125  << std::endl);
126 
127  return std::move(func);
128 }
129 
130 /** Creates a backend argminmax layer function
131  *
132  * @tparam ArgMinMaxLayerFunction Backend activation function
133  * @tparam TargetInfo Target-specific information
134  *
135  * @param[in] node Node to create the backend function for
136  *
137  * @return Backend argminmax layer function
138  */
139 template <typename ArgMinMaxLayerFunction, typename TargetInfo>
140 std::unique_ptr<IFunction> create_arg_min_max_layer(ArgMinMaxLayerNode &node)
141 {
142  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
143 
144  // Extract IO and info
145  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
146  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
147  const ReductionOperation op = node.reduction_operation();
148  unsigned int axis = node.axis();
149 
150  // Create function
151  auto func = std::make_unique<ArgMinMaxLayerFunction>();
152  func->configure(input, axis, output, op);
153 
154  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
155  << node.name()
156  << " Type: " << node.type()
157  << " Target: " << TargetInfo::TargetType
158  << " Data Type: " << input->info()->data_type()
159  << " Shape: " << input->info()->tensor_shape()
160  << " Reduction Operation: " << op
161  << " axis: " << axis
162  << std::endl);
163 
164  return std::move(func);
165 }
166 
167 /** Create a backend batch normalization layer function
168  *
169  * @tparam BatchNormalizationLayerFunction Backend batch normalization function
170  * @tparam TargetInfo Target-specific information
171  *
172  * @param[in] node Node to create the backend function for
173  *
174  * @return Backend batch normalization layer function
175  */
176 template <typename BatchNormalizationLayerFunction, typename TargetInfo>
177 std::unique_ptr<IFunction> create_batch_normalization_layer(BatchNormalizationLayerNode &node)
178 {
179  validate_node<TargetInfo>(node, 5 /* expected inputs */, 1 /* expected outputs */);
180 
181  // Extract IO and info
182  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
183  typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(1));
184  typename TargetInfo::TensorType *var = get_backing_tensor<TargetInfo>(node.input(2));
185  typename TargetInfo::TensorType *beta = get_backing_tensor<TargetInfo>(node.input(3));
186  typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(4));
187 
188  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
189  const float epsilon = node.epsilon();
190  const ActivationLayerInfo fused_act = node.fused_activation();
191 
192  // Create and configure function
193  auto func = std::make_unique<BatchNormalizationLayerFunction>();
194  func->configure(input, output, mean, var, beta, gamma, epsilon, fused_act);
195 
196  // Log info
197  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
198  << node.name()
199  << " Type: " << node.type()
200  << " Target: " << TargetInfo::TargetType
201  << " Data Type: " << input->info()->data_type()
202  << " Shape: " << input->info()->tensor_shape()
203  << " Epsilon: " << epsilon << " "
204  << (fused_act.enabled() ? to_string(fused_act.activation()) : "")
205  << " InPlace: " << is_in_place_operation(input, output)
206  << std::endl);
207 
208  return std::move(func);
209 }
210 
211 /** Create a backend batch normalization layer function
212  *
213  * @tparam BatchNormalizationLayerFunction Backend batch normalization function
214  * @tparam TargetInfo Target-specific information
215  *
216  * @param[in] node Node to create the backend function for
217  * @param[in] ctx Graph context
218  *
219  * @return Backend batch normalization layer function
220  */
221 template <typename FusedLayerTypes, typename TargetInfo>
222 std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_layer(FusedConvolutionBatchNormalizationNode &node, GraphContext &ctx)
223 {
224  validate_node<TargetInfo>(node, 7 /* expected inputs */, 1 /* expected outputs */);
225 
226  // Extract IO and info
227  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
228  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
229  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
230  typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(3));
231  typename TargetInfo::TensorType *var = get_backing_tensor<TargetInfo>(node.input(4));
232  typename TargetInfo::TensorType *beta = get_backing_tensor<TargetInfo>(node.input(5));
233  typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(6));
234 
235  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
236 
237  const PadStrideInfo conv_info = node.convolution_info();
238  const unsigned int num_groups = node.num_groups();
239  const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled;
240  const ActivationLayerInfo fused_act = node.fused_activation();
241  const float epsilon = node.epsilon();
242 
243  // Create and configure function (we assume that functions have been validated before creation)
244  std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
245  std::unique_ptr<IFunction> func;
246  std::string func_name;
247 
248  using FType = FusedConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>;
249 
250  // Create and configure function
251  std::tie(func, func_name) = create_named_memory_managed_function<FType>(
252  std::string("FusedConvolutionBatchNormalizationLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, fused_act);
253 
254  // Log info
255  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
256  << node.name()
257  << " Type: " << node.type()
258  << " Target: " << TargetInfo::TargetType
259  << " Data Type: " << input->info()->data_type()
260  << " Input shape: " << input->info()->tensor_shape()
261  << " Weights shape: " << weights->info()->tensor_shape()
262  << " Output shape: " << output->info()->tensor_shape()
263  << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
264  << std::endl);
265  return std::move(func);
266 }
267 
268 /** Create a backend fused depthwise convolution batch normalization layer function
269  *
270  * @tparam FusedLayerTypes Fused layer types
271  * @tparam TargetInfo Target-specific information
272  *
273  * @param[in] node Node to create the backend function for
274  * @param[in] ctx Graph context
275  *
276  * @return Backend fused depthwise convolution batch normalization layer function
277  */
278 template <typename FusedLayerTypes, typename TargetInfo>
279 std::unique_ptr<IFunction> create_fused_depthwise_convolution_batch_normalization_layer(FusedDepthwiseConvolutionBatchNormalizationNode &node, GraphContext &ctx)
280 {
281  validate_node<TargetInfo>(node, 7 /* expected inputs */, 1 /* expected outputs */);
282 
283  // Extract IO and info
284  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
285  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
286  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
287  typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(3));
288  typename TargetInfo::TensorType *var = get_backing_tensor<TargetInfo>(node.input(4));
289  typename TargetInfo::TensorType *beta = get_backing_tensor<TargetInfo>(node.input(5));
290  typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(6));
291 
292  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
293 
294  const PadStrideInfo conv_info = node.convolution_info();
295  const unsigned int depth_multiplier = node.depth_multiplier();
296  const ActivationLayerInfo fused_act = node.fused_activation();
297  const float epsilon = node.epsilon();
298 
299  // Create and configure function (we assume that functions have been validated before creation)
300  std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
301  std::unique_ptr<IFunction> func;
302  std::string func_name;
303 
304  using FType = FusedDepthwiseConvolutionBatchNormalizationFunction<TargetInfo, FusedLayerTypes>;
305 
306  // Create and configure function
307  std::tie(func, func_name) = create_named_memory_managed_function<FType>(
308  std::string("FusedDepthwiseConvolutionBatchNormalizationLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, depth_multiplier, fused_act);
309 
310  // Log info
311  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
312  << node.name()
313  << " Type: " << node.type()
314  << " Target: " << TargetInfo::TargetType
315  << " Data Type: " << input->info()->data_type()
316  << " Input shape: " << input->info()->tensor_shape()
317  << " Weights shape: " << weights->info()->tensor_shape()
318  << " Output shape: " << output->info()->tensor_shape()
319  << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
320  << std::endl);
321  return std::move(func);
322 }
323 
324 /** Create a backend bounding box transform layer function
325  *
326  * @tparam BoundingBoxTransformLayerFunction Backend bounding box transform function
327  * @tparam TargetInfo Target-specific information
328  *
329  * @param[in] node Node to create the backend function for
330  *
331  * @return Backend bounding box transform layer function
332  */
333 template <typename BoundingBoxTransformLayerFunction, typename TargetInfo>
334 std::unique_ptr<IFunction> create_bounding_box_transform_layer(BoundingBoxTransformLayerNode &node)
335 {
336  validate_node<TargetInfo>(node, 2 /* expected inputs */, 1 /* expected outputs */);
337 
338  // Extract IO and info
339  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
340  typename TargetInfo::TensorType *deltas = get_backing_tensor<TargetInfo>(node.input(1));
341  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
342  const BoundingBoxTransformInfo bbox_info = node.info();
343 
344  // Create and configure function
345  auto func = std::make_unique<BoundingBoxTransformLayerFunction>();
346  func->configure(input, output, deltas, bbox_info);
347 
348  // Log info
349  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
350  << node.name()
351  << " Type: " << node.type()
352  << " Target: " << TargetInfo::TargetType
353  << " Data Type: " << input->info()->data_type()
354  << " Shape: " << input->info()->tensor_shape()
355  << " BoundingBox Info img W: " << bbox_info.img_width() << " "
356  << " BoundingBox Info img H: " << bbox_info.img_height() << " "
357  << std::endl);
358 
359  return std::move(func);
360 }
361 
362 /** Create a backend channel shuffle layer function
363  *
364  * @tparam ChannelShuffleLayerFunction Backend channel shuffle function
365  * @tparam TargetInfo Target-specific information
366  *
367  * @param[in] node Node to create the backend function for
368  *
369  * @return Backend channel shuffle layer function
370  */
371 template <typename ChannelShuffleLayerFunction, typename TargetInfo>
372 std::unique_ptr<IFunction> create_channel_shuffle_layer(ChannelShuffleLayerNode &node)
373 {
374  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
375 
376  // Extract IO and info
377  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
378  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
379  const unsigned int num_groups = node.num_groups();
380 
381  // Create function
382  auto func = std::make_unique<ChannelShuffleLayerFunction>();
383  func->configure(input, output, num_groups);
384 
385  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
386  << node.name()
387  << " Type: " << node.type()
388  << " Target: " << TargetInfo::TargetType
389  << " Data Type: " << input->info()->data_type()
390  << " Shape: " << input->info()->tensor_shape()
391  << " Num groups: " << num_groups
392  << std::endl);
393 
394  return std::move(func);
395 }
396 
397 /** Create a backend layer concatenate function
398  *
399  * @tparam ConcatenateLayerFunction Backend concatenate function
400  * @tparam TargetInfo Target-specific information
401  *
402  * @param[in] node Node to create the backend function for
403  *
404  * @return Backend concatenate layer function
405  */
406 template <typename ConcatenateLayerFunction, typename TargetInfo>
407 std::unique_ptr<arm_compute::IFunction> create_concatenate_layer(ConcatenateLayerNode &node)
408 {
409  ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating Concatenate node with ID : " << node.id() << " and Name: " << node.name() << std::endl);
410  ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
411 
412  // Return nullptr if depth concatenate is switched off
413  if(!node.is_enabled())
414  {
415  return nullptr;
416  }
417 
418  // Extract IO and info
419  std::vector<typename TargetInfo::SrcTensorType *> inputs;
420  for(unsigned int i = 0; i < node.num_inputs(); ++i)
421  {
422  inputs.push_back(get_backing_tensor<TargetInfo>(node.input(i)));
423  }
424  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
425  const DataLayout data_layout = node.output(0) != nullptr ? node.output(0)->desc().layout : DataLayout::UNKNOWN;
426  const size_t concat_axis = get_dimension_idx(data_layout, node.concatenation_axis());
427 
428  // Create and configure function
429  auto func = std::make_unique<ConcatenateLayerFunction>();
430  func->configure(inputs, output, concat_axis);
431 
432  // Log info
433  const bool is_quantized = is_data_type_quantized_asymmetric(output->info()->data_type());
434  std::ostringstream qss;
435  if(is_quantized)
436  {
437  qss << " Output QuantInfo: " << output->info()->quantization_info();
438  }
439  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
440  << node.name()
441  << " Type: " << node.type()
442  << " Target: " << TargetInfo::TargetType
443  << " Data Type: " << output->info()->data_type()
444  << " Shape: " << output->info()->tensor_shape()
445  << " Num Inputs: " << inputs.size()
446  << " Axis: " << concat_axis
447  << qss.str()
448  << std::endl);
449 
450  return std::move(func);
451 }
452 
453 /** Create a backend convolution layer function
454  *
455  * @tparam ConvolutionLayerFunctions Backend convolution functions
456  * @tparam TargetInfo Target-specific information
457  *
458  * @param[in] node Node to create the backend function for
459  * @param[in] ctx Graph context
460  *
461  * @return Backend convolution layer function
462  */
463 template <typename ConvolutionLayerFunctions, typename TargetInfo>
464 std::unique_ptr<IFunction> create_convolution_layer(ConvolutionLayerNode &node, GraphContext &ctx)
465 {
466  validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
467 
468  // Extract IO and info
469  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
470  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
471  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
472  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
473 
474  const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
475 
476  if(is_quantized)
477  {
478  biases->info()->set_data_type(DataType::S32);
479  }
480 
481  const PadStrideInfo conv_info = node.convolution_info();
482  const unsigned int num_groups = node.num_groups();
483  const ConvolutionMethod conv_algorithm = node.convolution_method();
484  const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled;
485  const ActivationLayerInfo fused_act = node.fused_activation();
486 
487  // Create and configure function (we assume that functions have been validated before creation)
488  std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
489  std::unique_ptr<IFunction> func;
490  std::string func_name;
491 
492  if(conv_algorithm == ConvolutionMethod::Winograd)
493  {
494  ARM_COMPUTE_ERROR_ON_MSG(num_groups != 1, "WinogradConvolutionLayer does not support grouping!");
495  std::tie(func, func_name) = create_named_memory_managed_function<typename ConvolutionLayerFunctions::WinogradConvolutionLayer>(
496  std::string("WinogradConvolutionLayer"), mm,
497  input, weights, biases, output, conv_info, fused_act, fast_math);
498  }
499  else if(conv_algorithm == ConvolutionMethod::Direct)
500  {
501  ARM_COMPUTE_ERROR_ON_MSG(num_groups != 1, "DirectConvolutionLayer does not support grouping!");
502  std::tie(func, func_name) = create_named_function<typename ConvolutionLayerFunctions::DirectConvolutionLayer>(
503  std::string("DirectConvolutionLayer"),
504  input, weights, biases, output, conv_info, fused_act);
505  }
506  else if(conv_algorithm == ConvolutionMethod::GEMM)
507  {
508  std::tie(func, func_name) = create_named_memory_managed_function<typename ConvolutionLayerFunctions::GEMMConvolutionLayer>(
509  std::string("GEMMConvolutionLayer"), mm,
510  input, weights, biases, output, conv_info,
511  WeightsInfo(), Size2D(1U, 1U), fused_act, num_groups);
512  }
513  else
514  {
515  std::tie(func, func_name) = create_named_memory_managed_function<typename ConvolutionLayerFunctions::GenericConvolutionLayer>(
516  std::string("GenericConvolutionLayer"), mm,
517  input, weights, biases, output, conv_info,
518  WeightsInfo(), Size2D(1U, 1U), fused_act, fast_math, num_groups);
519  }
520 
521  // Log info
522  std::ostringstream qss;
523  if(is_quantized)
524  {
525  qss << " Input QuantInfo: " << input->info()->quantization_info()
526  << " Weights QuantInfo: " << weights->info()->quantization_info()
527  << " Output QuantInfo: " << output->info()->quantization_info();
528  }
529  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
530  << node.name()
531  << " Type: " << func_name
532  << " Target: " << TargetInfo::TargetType
533  << " Data Type: " << input->info()->data_type()
534  << " Groups: " << num_groups
535  << " Input shape: " << input->info()->tensor_shape()
536  << " Weights shape: " << weights->info()->tensor_shape()
537  << " Output shape: " << output->info()->tensor_shape()
538  << qss.str()
539  << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
540  << std::endl);
541  return std::move(func);
542 }
543 
544 /** Create a backend convolution layer function with post operator
545  *
546  * @tparam ConvolutionLayerFunctions Backend convolution functions
547  * @tparam TargetInfo Target-specific information
548  *
549  * @param[in] node Node to create the backend function for
550  * @param[in] ctx Graph context
551  *
552  * @return Backend convolution layer function
553  */
554 template <typename ConvolutionLayerFunctions, typename TargetInfo>
555 std::unique_ptr<IFunction> create_fused_convolution_with_post_op(FusedConvolutionWithPostOpNode &node, GraphContext &ctx)
556 {
557  validate_node<TargetInfo>(node, 4 /* expected inputs */, 1 /* expected outputs */);
558 
559  // Extract IO and info
560  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
561  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
562  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
563  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
564 
565  const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
566 
567  if(is_quantized)
568  {
569  biases->info()->set_data_type(DataType::S32);
570  }
571 
572  const PadStrideInfo conv_info = node.convolution_info();
573  const unsigned int num_groups = node.num_groups();
574  const ActivationLayerInfo fused_act = node.fused_activation();
575 
576  experimental::PostOpList<typename TargetInfo::TensorType *> post_ops;
577 
578  auto &post_op_info_list = node.post_op_info_list();
579  for(const auto &post_op_info : post_op_info_list)
580  {
581  switch(post_op_info->type())
582  {
583  case PostOpType::Activation:
584  {
585  const auto act_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoActivation *>(post_op_info.get());
586  post_ops.template push_back_op<experimental::PostOpAct<typename TargetInfo::TensorType *>>(act_info->_act);
587  break;
588  }
589  case PostOpType::Eltwise_Add:
590  {
591  typename TargetInfo::TensorType *add_input = get_backing_tensor<TargetInfo>(node.input(3));
592  const auto eltwise_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoEltwiseAdd *>(post_op_info.get());
593  post_ops.template push_back_op<experimental::PostOpEltwiseAdd<typename TargetInfo::TensorType *>>(add_input, eltwise_info->_prev_op_dst_pos, eltwise_info->_policy);
594  break;
595  }
596  default:
597  {
598  ARM_COMPUTE_ERROR("Unsupported PostOpType");
599  }
600  }
601  }
602 
603  // Create and configure function (we assume that functions have been validated before creation)
604  std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
605  std::unique_ptr<IFunction> func;
606  std::string func_name;
607 
608  // Fuse convolution with post ops is only supported for conv1x1, which is only implemented as gemmconv2d
609  std::tie(func, func_name) = create_named_memory_managed_function<typename ConvolutionLayerFunctions::GEMMConvolutionLayer>(
610  std::string("GEMMConvolutionLayer"), mm,
611  input, weights, biases, output, conv_info,
612  WeightsInfo(), Size2D(1U, 1U), fused_act, num_groups, post_ops);
613 
614  // Log info
615  std::ostringstream qss;
616  if(is_quantized)
617  {
618  qss << " Input QuantInfo: " << input->info()->quantization_info()
619  << " Weights QuantInfo: " << weights->info()->quantization_info()
620  << " Output QuantInfo: " << output->info()->quantization_info();
621  }
622  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
623  << node.name()
624  << " Type: " << func_name
625  << " Target: " << TargetInfo::TargetType
626  << " Data Type: " << input->info()->data_type()
627  << " Groups: " << num_groups
628  << " Input shape: " << input->info()->tensor_shape()
629  << " Weights shape: " << weights->info()->tensor_shape()
630  << " Output shape: " << output->info()->tensor_shape()
631  << qss.str()
632  << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
633  << " Post ops" << post_ops
634  << std::endl);
635  return std::move(func);
636 }
637 
638 /** Create a backend convolution batch normalization layer function with post operator
639  *
640  * @tparam FusedLayerTypes Backend convolution functions
641  * @tparam TargetInfo Target-specific information
642  *
643  * @param[in] node Node to create the backend function for
644  * @param[in] ctx Graph context
645  *
646  * @return Backend fused convolution with batch normalization layer function
647  */
648 template <typename FusedLayerTypes, typename TargetInfo>
649 std::unique_ptr<IFunction> create_fused_convolution_batch_normalization_with_post_op(FusedConvolutionBatchNormalizationWithPostOpsNode &node, GraphContext &ctx)
650 {
651  validate_node<TargetInfo>(node, 8 /* expected inputs */, 1 /* expected outputs */);
652 
653  // Extract IO and info
654  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
655  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
656  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
657  typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(3));
658  typename TargetInfo::TensorType *var = get_backing_tensor<TargetInfo>(node.input(4));
659  typename TargetInfo::TensorType *beta = get_backing_tensor<TargetInfo>(node.input(5));
660  typename TargetInfo::TensorType *gamma = get_backing_tensor<TargetInfo>(node.input(6));
661 
662  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
663 
664  const PadStrideInfo conv_info = node.convolution_info();
665  const unsigned int num_groups = node.num_groups();
666  const bool fast_math = node.fast_math_hint() == FastMathHint::Enabled;
667  const float epsilon = node.epsilon();
668 
669  experimental::PostOpList<typename TargetInfo::TensorType *> post_ops;
670 
671  auto &post_op_info_list = node.post_op_info_list();
672  for(const auto &post_op_info : post_op_info_list)
673  {
674  switch(post_op_info->type())
675  {
676  case PostOpType::Activation:
677  {
678  const auto act_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoActivation *>(post_op_info.get());
679  post_ops.template push_back_op<experimental::PostOpAct<typename TargetInfo::TensorType *>>(act_info->_act);
680  break;
681  }
682  case PostOpType::Eltwise_Add:
683  {
684  typename TargetInfo::TensorType *add_input = get_backing_tensor<TargetInfo>(node.input(3));
685  const auto eltwise_info = utils::cast::polymorphic_downcast<const ConvPostOpInfoEltwiseAdd *>(post_op_info.get());
686  post_ops.template push_back_op<experimental::PostOpEltwiseAdd<typename TargetInfo::TensorType *>>(add_input, eltwise_info->_prev_op_dst_pos, eltwise_info->_policy);
687  break;
688  }
689  default:
690  {
691  ARM_COMPUTE_ERROR("Unsupported PostOpType");
692  }
693  }
694  }
695 
696  // Create and configure function (we assume that functions have been validated before creation)
697  std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
698  std::unique_ptr<IFunction> func;
699  std::string func_name;
700 
701  using FType = FusedConvolutionBatchNormalizationWithPostOpsFunction<TargetInfo, FusedLayerTypes>;
702 
703  // Create and configure function
704  std::tie(func, func_name) = create_named_memory_managed_function<FType>(
705  std::string("FusedConvolutionBatchNormalizationLayerWithPostOpsLayer"), mm, input, weights, biases, output, mean, var, beta, gamma, epsilon, conv_info, num_groups, fast_math, post_ops);
706 
707  // Log info
708  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
709  << node.name()
710  << " Type: " << node.type()
711  << " Target: " << TargetInfo::TargetType
712  << " Data Type: " << input->info()->data_type()
713  << " Input shape: " << input->info()->tensor_shape()
714  << " Weights shape: " << weights->info()->tensor_shape()
715  << " Output shape: " << output->info()->tensor_shape()
716  << " Post Ops:" << post_ops
717  << std::endl);
718  return std::move(func);
719 }
720 
721 /** Create a backend deconvolution layer function
722  *
723  * @tparam DeconvolutionLayerFunction Backend deconvolution function
724  * @tparam TargetInfo Target-specific information
725  *
726  * @param[in] node Node to create the backend function for
727  * @param[in] ctx Graph context
728  *
729  * @return Backend deconvolution layer function
730  */
731 template <typename DeconvolutionLayerFunction, typename TargetInfo>
732 std::unique_ptr<IFunction> create_deconvolution_layer(DeconvolutionLayerNode &node, GraphContext &ctx)
733 {
734  validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
735 
736  // Extract IO and info
737  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
738  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
739  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
740  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
741 
742  const PadStrideInfo deconv_info = node.deconvolution_info();
743 
744  // Create and configure function (we assume that functions have been validated before creation)
745  std::shared_ptr<IMemoryManager> mm = get_memory_manager(ctx, TargetInfo::TargetType);
746  std::unique_ptr<IFunction> func;
747 
748  std::tie(func, std::ignore) = create_named_memory_managed_function<DeconvolutionLayerFunction>(
749  std::string(), mm,
750  input, weights, biases, output, deconv_info);
751 
752  // Log info
753  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
754  << node.name()
755  << " Type: " << node.type()
756  << " Target: " << TargetInfo::TargetType
757  << " Data Type: " << input->info()->data_type()
758  << " Input shape: " << input->info()->tensor_shape()
759  << " Weights shape: " << weights->info()->tensor_shape()
760  << " Output shape: " << output->info()->tensor_shape()
761  << std::endl);
762  return func;
763 }
764 
765 /** Create a backend layer depth-wise convolution function
766  *
767  * @tparam DepthwiseConvolutionLayerFunctions Backend depthwise convolution function
768  * @tparam TargetInfo Target-specific information
769  *
770  * @param[in] node Node to create the backend function for
771  *
772  * @return Backend depth-wise convolution layer function
773  */
774 template <typename DepthwiseConvolutionLayer, typename TargetInfo>
775 std::unique_ptr<IFunction> create_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node)
776 {
777  validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
778 
779  // Extract IO and info
780  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
781  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
782  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
783  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
784 
785  const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
786 
787  if(is_quantized)
788  {
789  biases->info()->set_data_type(DataType::S32);
790  }
791 
792  const PadStrideInfo conv_info = node.convolution_info();
793  const unsigned int depth_multiplier = node.depth_multiplier();
794  const ActivationLayerInfo fused_act = node.fused_activation();
795 
796  // Create and configure function (we assume that functions have been validated before creation)
797  std::unique_ptr<IFunction> func;
798  std::string func_name;
799 
800  std::tie(func, func_name) = create_named_function<DepthwiseConvolutionLayer>(
801  std::string("DepthwiseConvolutionLayer"),
802  input, weights, biases, output, conv_info, depth_multiplier, fused_act);
803 
804  // Log info
805  std::ostringstream qss;
806  if(is_quantized)
807  {
808  qss << " Input QuantInfo: " << input->info()->quantization_info()
809  << " Weights QuantInfo: " << weights->info()->quantization_info()
810  << " Output QuantInfo: " << output->info()->quantization_info();
811  }
812  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
813  << node.name()
814  << " Type: " << func_name
815  << " Target: " << TargetInfo::TargetType
816  << " Data Type: " << input->info()->data_type()
817  << " Input shape: " << input->info()->tensor_shape()
818  << " Weights shape: " << weights->info()->tensor_shape()
819  << " Output shape: " << output->info()->tensor_shape()
820  << " Depth multiplier: " << depth_multiplier
821  << qss.str()
822  << (fused_act.enabled() ? " " + to_string(fused_act.activation()) : "")
823  << std::endl);
824  return std::move(func);
825 }
826 
827 /** Create a backend depth to space layer function
828  *
829  * @tparam DepthToSpaceLayerNode Function Backend depth to space function
830  * @tparam TargetInfo Target-specific information
831  *
832  * @param[in] node Node to create the backend function for
833  *
834  * @return Backend depth to space layer function
835  */
836 template <typename DepthToSpaceLayerFunction, typename TargetInfo>
837 std::unique_ptr<IFunction> create_depth_to_space_layer(DepthToSpaceLayerNode &node)
838 {
839  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
840 
841  // Extract IO and info
842  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
843  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
844 
845  ARM_COMPUTE_ERROR_ON(input == nullptr);
846  ARM_COMPUTE_ERROR_ON(output == nullptr);
847 
848  // Create and configure function
849  auto func = std::make_unique<DepthToSpaceLayerFunction>();
850  func->configure(input, output, node.block_shape());
851 
852  // Log info
853  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
854  << node.name()
855  << " Type: " << node.type()
856  << " Target: " << TargetInfo::TargetType
857  << " Data Type: " << input->info()->data_type()
858  << " Input shape: " << input->info()->tensor_shape()
859  << " Block Size: " << node.block_shape()
860  << " Output shape: " << output->info()->tensor_shape()
861  << std::endl);
862 
863  return std::move(func);
864 }
865 
866 /** Create a backend dequantize layer function
867  *
868  * @tparam DequantizationLayer Function Backend dequantize function
869  * @tparam TargetInfo Target-specific information
870  *
871  * @param[in] node Node to create the backend function for
872  *
873  * @return Backend dequantize layer function
874  */
875 template <typename DequantizationLayerFunction, typename TargetInfo>
876 std::unique_ptr<IFunction> create_dequantization_layer(DequantizationLayerNode &node)
877 {
878  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
879 
880  // Extract IO and info
881  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
882  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
883 
884  ARM_COMPUTE_ERROR_ON(input == nullptr);
885  ARM_COMPUTE_ERROR_ON(output == nullptr);
886 
887  // Create and configure function
888  auto func = std::make_unique<DequantizationLayerFunction>();
889  func->configure(input, output);
890 
891  // Log info
892  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
893  << node.name()
894  << " Type: " << node.type()
895  << " Target: " << TargetInfo::TargetType
896  << " Data Type: " << input->info()->data_type()
897  << " Input shape: " << input->info()->tensor_shape()
898  << " Input quantization info: " << output->info()->quantization_info()
899  << " Output shape: " << output->info()->tensor_shape()
900  << std::endl);
901 
902  return std::move(func);
903 }
904 /** Create a backend detection output layer function
905  *
906  * @tparam DetectionOutputLayer Function Backend detection output function
907  * @tparam TargetInfo Target-specific information
908  *
909  * @param[in] node Node to create the backend function for
910  *
911  * @return Backend detection output layer function
912  */
913 template <typename DetectionOutputLayerFunction, typename TargetInfo>
914 std::unique_ptr<IFunction> create_detection_output_layer(DetectionOutputLayerNode &node)
915 {
916  validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
917 
918  // Extract IO and info
919  typename TargetInfo::TensorType *input0 = get_backing_tensor<TargetInfo>(node.input(0));
920  typename TargetInfo::TensorType *input1 = get_backing_tensor<TargetInfo>(node.input(1));
921  typename TargetInfo::TensorType *input2 = get_backing_tensor<TargetInfo>(node.input(2));
922  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
923  const DetectionOutputLayerInfo detect_info = node.detection_output_info();
924 
925  ARM_COMPUTE_ERROR_ON(input0 == nullptr);
926  ARM_COMPUTE_ERROR_ON(input1 == nullptr);
927  ARM_COMPUTE_ERROR_ON(input2 == nullptr);
928  ARM_COMPUTE_ERROR_ON(output == nullptr);
929 
930  // Create and configure function
931  auto func = std::make_unique<DetectionOutputLayerFunction>();
932  func->configure(input0, input1, input2, output, detect_info);
933 
934  // Log info
935  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
936  << node.name()
937  << " Type: " << node.type()
938  << " Target: " << TargetInfo::TargetType
939  << " Data Type: " << input0->info()->data_type()
940  << " Input0 shape: " << input0->info()->tensor_shape()
941  << " Input1 shape: " << input1->info()->tensor_shape()
942  << " Input2 shape: " << input2->info()->tensor_shape()
943  << " Output shape: " << output->info()->tensor_shape()
944  << " DetectionOutputLayer info: " << detect_info
945  << std::endl);
946 
947  return std::move(func);
948 }
949 
950 /** Create a backend detection post process layer function
951  *
952  * @tparam DetectionPostProcessLayerFunction Backend detection output function
953  * @tparam TargetInfo Target-specific information
954  *
955  * @param[in] node Node to create the backend function for
956  *
957  * @return Backend detection post process layer function
958  */
959 template <typename DetectionPostProcessLayerFunction, typename TargetInfo>
960 std::unique_ptr<IFunction> create_detection_post_process_layer(DetectionPostProcessLayerNode &node)
961 {
962  validate_node<TargetInfo>(node, 3 /* expected inputs */, 4 /* expected outputs */);
963 
964  // Extract IO and info
965  typename TargetInfo::TensorType *input0 = get_backing_tensor<TargetInfo>(node.input(0));
966  typename TargetInfo::TensorType *input1 = get_backing_tensor<TargetInfo>(node.input(1));
967  typename TargetInfo::TensorType *input2 = get_backing_tensor<TargetInfo>(node.input(2));
968  typename TargetInfo::TensorType *output0 = get_backing_tensor<TargetInfo>(node.output(0));
969  typename TargetInfo::TensorType *output1 = get_backing_tensor<TargetInfo>(node.output(1));
970  typename TargetInfo::TensorType *output2 = get_backing_tensor<TargetInfo>(node.output(2));
971  typename TargetInfo::TensorType *output3 = get_backing_tensor<TargetInfo>(node.output(3));
972  const DetectionPostProcessLayerInfo detect_info = node.detection_post_process_info();
973 
974  ARM_COMPUTE_ERROR_ON(input0 == nullptr);
975  ARM_COMPUTE_ERROR_ON(input1 == nullptr);
976  ARM_COMPUTE_ERROR_ON(input2 == nullptr);
977  ARM_COMPUTE_ERROR_ON(output0 == nullptr);
978  ARM_COMPUTE_ERROR_ON(output1 == nullptr);
979  ARM_COMPUTE_ERROR_ON(output2 == nullptr);
980  ARM_COMPUTE_ERROR_ON(output3 == nullptr);
981 
982  // Create and configure function
983  auto func = std::make_unique<DetectionPostProcessLayerFunction>();
984  func->configure(input0, input1, input2, output0, output1, output2, output3, detect_info);
985 
986  // Log info
987  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
988  << node.name()
989  << " Type: " << node.type()
990  << " Target: " << TargetInfo::TargetType
991  << " Data Type: " << input0->info()->data_type()
992  << " Input0 shape: " << input0->info()->tensor_shape()
993  << " Input1 shape: " << input1->info()->tensor_shape()
994  << " Input2 shape: " << input2->info()->tensor_shape()
995  << " Output0 shape: " << output0->info()->tensor_shape()
996  << " Output1 shape: " << output1->info()->tensor_shape()
997  << " Output2 shape: " << output2->info()->tensor_shape()
998  << " Output3 shape: " << output3->info()->tensor_shape()
999  << " DetectionPostProcessLayer info: " << detect_info
1000  << std::endl);
1001 
1002  return std::move(func);
1003 }
1004 
1005 /** Create a backend element-wise operation layer function
1006  *
1007  * @tparam EltwiseFunctions Backend element-wise function
1008  * @tparam TargetInfo Target-specific information
1009  *
1010  * @param[in] node Node to create the backend function for
1011  *
1012  * @return Backend element-wise operation layer function
1013  */
1014 template <typename EltwiseFunctions, typename TargetInfo>
1015 std::unique_ptr<IFunction> create_eltwise_layer(EltwiseLayerNode &node)
1016 {
1017  validate_node<TargetInfo>(node, 2 /* expected inputs */, 1 /* expected outputs */);
1018 
1019  // Extract IO and info
1020  typename TargetInfo::TensorType *input1 = get_backing_tensor<TargetInfo>(node.input(0));
1021  typename TargetInfo::TensorType *input2 = get_backing_tensor<TargetInfo>(node.input(1));
1022  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1023  const EltwiseOperation eltwise_op = node.eltwise_operation();
1024  const ConvertPolicy convert_policy = node.convert_policy();
1025  const ActivationLayerInfo act_info = node.fused_activation();
1026  ARM_COMPUTE_ERROR_ON(input1 == nullptr);
1027  ARM_COMPUTE_ERROR_ON(input2 == nullptr);
1028  ARM_COMPUTE_ERROR_ON(output == nullptr);
1029 
1030  std::unique_ptr<IFunction> func = nullptr;
1031  std::string func_name;
1032  if(eltwise_op == EltwiseOperation::Add)
1033  {
1034  std::tie(func, func_name) = create_named_function<typename EltwiseFunctions::Addition>(
1035  std::string("ArithmeticAddition"),
1036  input1, input2, output, convert_policy, act_info);
1037  }
1038  else if(eltwise_op == EltwiseOperation::Sub)
1039  {
1040  std::tie(func, func_name) = create_named_function<typename EltwiseFunctions::Subtraction>(
1041  std::string("ArithmeticSubtraction"),
1042  input1, input2, output, convert_policy, act_info);
1043  }
1044  else if(eltwise_op == EltwiseOperation::Mul)
1045  {
1046  std::tie(func, func_name) = create_named_function<typename EltwiseFunctions::Multiplication>(
1047  std::string("PixelWiseMultiplication"),
1048  input1, input2, output, 1.f, convert_policy, node.rounding_policy(), act_info);
1049  }
1050  else if(eltwise_op == EltwiseOperation::Max)
1051  {
1052  std::tie(func, func_name) = create_named_function<typename EltwiseFunctions::Maximum>(
1053  std::string("ElementwiseMaximum"),
1054  input1, input2, output, act_info);
1055  }
1056  else if(eltwise_op == EltwiseOperation::Div)
1057  {
1058  std::tie(func, func_name) = create_named_function<typename EltwiseFunctions::Division>(
1059  std::string("ArithmeticDivision"),
1060  input1, input2, output, act_info);
1061  }
1062  else
1063  {
1064  ARM_COMPUTE_ERROR("Unsupported element-wise operation!");
1065  }
1066 
1067  // Log info
1068  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1069  << node.name()
1070  << " Type: " << node.type()
1071  << " Target: " << TargetInfo::TargetType
1072  << " Operation: " << func_name
1073  << " Data Type: " << input1->info()->data_type()
1074  << " Shape: " << input1->info()->tensor_shape()
1075  << std::endl);
1076 
1077  return std::move(func);
1078 }
1079 
1080 /** Create a backend unary element-wise operation layer function
1081  *
1082  * @tparam UnaryEltwiseFunctions Backend unary element-wise function
1083  * @tparam TargetInfo Target-specific information
1084  *
1085  * @param[in] node Node to create the backend function for
1086  *
1087  * @return Backend unary element-wise operation layer function
1088  */
1089 template <typename UnaryEltwiseFunctions, typename TargetInfo>
1090 std::unique_ptr<IFunction> create_unary_eltwise_layer(UnaryEltwiseLayerNode &node)
1091 {
1092  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1093 
1094  // Extract IO and info
1095  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1096  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1097  const UnaryEltwiseOperation eltwise_op = node.eltwise_descriptor().op;
1098 
1099  ARM_COMPUTE_ERROR_ON(input == nullptr);
1100  ARM_COMPUTE_ERROR_ON(output == nullptr);
1101 
1102  std::unique_ptr<IFunction> func = nullptr;
1103  std::string func_name;
1104  if(eltwise_op == UnaryEltwiseOperation::Exp)
1105  {
1106  std::tie(func, func_name) = create_named_function<typename UnaryEltwiseFunctions::Exp>(
1107  std::string("Exp"),
1108  input, output);
1109  }
1110  else
1111  {
1112  ARM_COMPUTE_ERROR("Unsupported unary element-wise operation!");
1113  }
1114 
1115  // Log info
1116  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1117  << node.name()
1118  << " Type: " << node.type()
1119  << " Target: " << TargetInfo::TargetType
1120  << " Operation: " << func_name
1121  << " Data Type: " << input->info()->data_type()
1122  << " Shape: " << input->info()->tensor_shape()
1123  << std::endl);
1124 
1125  return std::move(func);
1126 }
1127 
1128 /** Create a backend flatten layer function
1129  *
1130  * @tparam FlattenLayerFunction Backend flatten function
1131  * @tparam TargetInfo Target-specific information
1132  *
1133  * @param[in] node Node to create the backend function for
1134  *
1135  * @return Backend flatten layer function
1136  */
1137 template <typename FlattenLayerFunction, typename TargetInfo>
1138 std::unique_ptr<IFunction> create_flatten_layer(FlattenLayerNode &node)
1139 {
1140  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1141 
1142  // Extract IO and info
1143  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1144  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1145 
1146  ARM_COMPUTE_ERROR_ON(input == nullptr);
1147  ARM_COMPUTE_ERROR_ON(output == nullptr);
1148 
1149  // Create and configure function
1150  auto func = std::make_unique<FlattenLayerFunction>();
1151  func->configure(input, output);
1152 
1153  // Log info
1154  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1155  << node.name()
1156  << " Type: " << node.type()
1157  << " Target: " << TargetInfo::TargetType
1158  << " Data Type: " << input->info()->data_type()
1159  << " Input shape: " << input->info()->tensor_shape()
1160  << " Output shape: " << output->info()->tensor_shape()
1161  << std::endl);
1162 
1163  return std::move(func);
1164 }
1165 
1166 /** Create a backend fully connected layer function
1167  *
1168  * @tparam FullyConnectedLayerFunction Backend fully-connected function
1169  * @tparam TargetInfo Target-specific information
1170  *
1171  * @param[in] node Node to create the backend function for
1172  * @param[in] ctx Graph context
1173  *
1174  * @return Backend fully connected layer function
1175  */
1176 template <typename FullyConnectedLayerFunction, typename TargetInfo>
1177 std::unique_ptr<IFunction> create_fully_connected_layer(FullyConnectedLayerNode &node, GraphContext &ctx)
1178 {
1179  validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
1180 
1181  // Extract IO and info
1182  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1183  typename TargetInfo::TensorType *weights = get_backing_tensor<TargetInfo>(node.input(1));
1184  typename TargetInfo::TensorType *biases = get_backing_tensor<TargetInfo>(node.input(2));
1185  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1186  FullyConnectedLayerInfo fc_info = node.info();
1187  fc_info.enable_fast_math = (node.fast_math_hint() == FastMathHint::Enabled);
1188 
1189  ARM_COMPUTE_ERROR_ON(input == nullptr);
1190  ARM_COMPUTE_ERROR_ON(weights == nullptr);
1191  ARM_COMPUTE_ERROR_ON(output == nullptr);
1192 
1193  // Create and configure function
1194  auto wm = get_weights_manager(ctx, TargetInfo::TargetType);
1195  auto mm = get_memory_manager(ctx, TargetInfo::TargetType);
1196  auto func = std::make_unique<FullyConnectedLayerFunction>(mm, wm.get());
1197  func->configure(input, weights, biases, output, fc_info);
1198 
1199  const bool is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
1200 
1201  // Log info
1202  std::ostringstream qss;
1203  if(is_quantized)
1204  {
1205  qss << " Input QuantInfo: " << input->info()->quantization_info()
1206  << " Weights QuantInfo: " << weights->info()->quantization_info()
1207  << " Output QuantInfo: " << output->info()->quantization_info();
1208  }
1209  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1210  << node.name()
1211  << " Type: " << node.type()
1212  << " Target: " << TargetInfo::TargetType
1213  << " Data Type: " << input->info()->data_type()
1214  << qss.str()
1215  << " Input shape: " << input->info()->tensor_shape()
1216  << " Weights shape: " << weights->info()->tensor_shape()
1217  << " Output shape: " << output->info()->tensor_shape()
1218  << std::endl);
1219 
1220  return std::move(func);
1221 }
1222 
1223 /** Create a backend generate proposals layer function
1224  *
1225  * @tparam GenerateProposalsLayerFunction Backend generate proposals function
1226  * @tparam TargetInfo Target-specific information
1227  *
1228  * @param[in] node Node to create the backend function for
1229  * @param[in] ctx Graph context
1230  *
1231  * @return Backend generate proposals layer function
1232  */
1233 template <typename GenerateProposalsLayerFunction, typename TargetInfo>
1234 std::unique_ptr<IFunction> create_generate_proposals_layer(GenerateProposalsLayerNode &node, GraphContext &ctx)
1235 {
1236  validate_node<TargetInfo>(node, 3 /* expected inputs */, 3 /* expected outputs */);
1237 
1238  // Extract IO and info
1239  typename TargetInfo::TensorType *scores = get_backing_tensor<TargetInfo>(node.input(0));
1240  typename TargetInfo::TensorType *deltas = get_backing_tensor<TargetInfo>(node.input(1));
1241  typename TargetInfo::TensorType *anchors = get_backing_tensor<TargetInfo>(node.input(2));
1242  typename TargetInfo::TensorType *proposals = get_backing_tensor<TargetInfo>(node.output(0));
1243  typename TargetInfo::TensorType *scores_out = get_backing_tensor<TargetInfo>(node.output(1));
1244  typename TargetInfo::TensorType *num_valid_proposals = get_backing_tensor<TargetInfo>(node.output(2));
1245  const GenerateProposalsInfo info = node.info();
1246 
1247  ARM_COMPUTE_ERROR_ON(scores == nullptr);
1248  ARM_COMPUTE_ERROR_ON(deltas == nullptr);
1249  ARM_COMPUTE_ERROR_ON(anchors == nullptr);
1250  ARM_COMPUTE_ERROR_ON(proposals == nullptr);
1251  ARM_COMPUTE_ERROR_ON(scores_out == nullptr);
1252 
1253  // Create and configure function
1254  auto func = std::make_unique<GenerateProposalsLayerFunction>(get_memory_manager(ctx, TargetInfo::TargetType));
1255  func->configure(scores, deltas, anchors, proposals, scores_out, num_valid_proposals, info);
1256 
1257  // Log info
1258  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated " << node.type()
1259  << " Target " << TargetInfo::TargetType
1260  << " Data Type: " << scores->info()->data_type()
1261  << " Scores shape: " << scores->info()->tensor_shape()
1262  << " Deltas shape: " << deltas->info()->tensor_shape()
1263  << " Anchors shape: " << anchors->info()->tensor_shape()
1264  << " Proposals shape: " << proposals->info()->tensor_shape()
1265  << " Num valid proposals shape: " << num_valid_proposals->info()->tensor_shape()
1266  << " Scores Out shape: " << scores_out->info()->tensor_shape()
1267  << std::endl);
1268 
1269  return std::move(func);
1270 }
1271 
1272 /** Create a backend l2 normalization layer function
1273  *
1274  * @tparam NormalizationLayerFunction Backend normalization function
1275  * @tparam TargetInfo Target-specific information
1276  *
1277  * @param[in] node Node to create the backend function for
1278  * @param[in] ctx Graph context
1279  *
1280  * @return Backend normalization layer function
1281  */
1282 template <typename L2NormalizeLayerFunction, typename TargetInfo>
1283 std::unique_ptr<IFunction> create_l2_normalize_layer(L2NormalizeLayerNode &node, GraphContext &ctx)
1284 {
1285  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1286 
1287  // Extract IO and info
1288  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1289  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1290  int axis = node.axis();
1291  float epsilon = node.epsilon();
1292 
1293  ARM_COMPUTE_ERROR_ON(input == nullptr);
1294  ARM_COMPUTE_ERROR_ON(output == nullptr);
1295 
1296  // Create and configure function
1297  auto mm = get_memory_manager(ctx, TargetInfo::TargetType);
1298  auto func = std::make_unique<L2NormalizeLayerFunction>(mm);
1299  func->configure(input, output, axis, epsilon);
1300 
1301  // Log info
1302  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1303  << node.name()
1304  << " Type: " << node.type()
1305  << " Target: " << TargetInfo::TargetType
1306  << " Data Type: " << input->info()->data_type()
1307  << " Input shape: " << input->info()->tensor_shape()
1308  << " Output shape: " << output->info()->tensor_shape()
1309  << " Axis: " << axis
1310  << " Epsilon: " << epsilon
1311  << std::endl);
1312 
1313  return std::move(func);
1314 }
1315 
1316 /** Create a backend normalization layer function
1317  *
1318  * @tparam NormalizationLayerFunction Backend normalization function
1319  * @tparam TargetInfo Target-specific information
1320  *
1321  * @param[in] node Node to create the backend function for
1322  * @param[in] ctx Graph context
1323  *
1324  * @return Backend normalization layer function
1325  */
1326 template <typename NormalizationLayerFunction, typename TargetInfo>
1327 std::unique_ptr<IFunction> create_normalization_layer(NormalizationLayerNode &node, GraphContext &ctx)
1328 {
1329  ARM_COMPUTE_UNUSED(ctx);
1330 
1331  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1332 
1333  // Extract IO and info
1334  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1335  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1336  const NormalizationLayerInfo norm_info = node.normalization_info();
1337  ARM_COMPUTE_ERROR_ON(input == nullptr);
1338  ARM_COMPUTE_ERROR_ON(output == nullptr);
1339 
1340  // Create and configure function
1341  auto func = std::make_unique<NormalizationLayerFunction>();
1342  func->configure(input, output, norm_info);
1343 
1344  // Log info
1345  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1346  << node.name()
1347  << " Type: " << node.type()
1348  << " Target: " << TargetInfo::TargetType
1349  << " Data Type: " << input->info()->data_type()
1350  << " Input shape: " << input->info()->tensor_shape()
1351  << " Output shape: " << output->info()->tensor_shape()
1352  << " Normalization info: " << norm_info.type()
1353  << std::endl);
1354 
1355  return std::move(func);
1356 }
1357 
1358 /** Create a backend normalize planar YUV layer function
1359  *
1360  * @tparam NormalizePlanarYUVLayerFunction Backend normalize planar YUV function
1361  * @tparam TargetInfo Target-specific information
1362  *
1363  * @param[in] node Node to create the backend function for
1364  *
1365  * @return Backend normalize plnar YUV layer function
1366  */
1367 template <typename NormalizePlanarYUVLayerFunction, typename TargetInfo>
1368 std::unique_ptr<IFunction> create_normalize_planar_yuv_layer(NormalizePlanarYUVLayerNode &node)
1369 {
1370  validate_node<TargetInfo>(node, 3 /* expected inputs */, 1 /* expected outputs */);
1371 
1372  // Extract IO and info
1373  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1374  typename TargetInfo::TensorType *mean = get_backing_tensor<TargetInfo>(node.input(1));
1375  typename TargetInfo::TensorType *std = get_backing_tensor<TargetInfo>(node.input(2));
1376  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1377  ARM_COMPUTE_ERROR_ON(input == nullptr);
1378  ARM_COMPUTE_ERROR_ON(mean == nullptr);
1379  ARM_COMPUTE_ERROR_ON(std == nullptr);
1380  ARM_COMPUTE_ERROR_ON(output == nullptr);
1381 
1382  // Create and configure function
1383  auto func = std::make_unique<NormalizePlanarYUVLayerFunction>();
1384  func->configure(input, output, mean, std);
1385 
1386  // Log info
1387  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1388  << node.name()
1389  << " Type: " << node.type()
1390  << " Target: " << TargetInfo::TargetType
1391  << " Data Type: " << input->info()->data_type()
1392  << " Shape: " << input->info()->tensor_shape()
1393  << std::endl);
1394 
1395  return std::move(func);
1396 }
1397 
1398 /** Create a backend pad layer function
1399  *
1400  * @tparam PadLayerFunction Backend pad function
1401  * @tparam TargetInfo Target-specific information
1402  *
1403  * @param[in] node Node to create the backend function for
1404  *
1405  * @return Backend pad layer function
1406  */
1407 template <typename PadLayerFunction, typename TargetInfo>
1408 std::unique_ptr<IFunction> create_pad_layer(PadLayerNode &node)
1409 {
1410  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1411 
1412  // Extract IO and info
1413  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1414  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1415  const PaddingList &padding = node.padding();
1416  const PixelValue pad_value = node.pad_value();
1417  ARM_COMPUTE_ERROR_ON(input == nullptr);
1418  ARM_COMPUTE_ERROR_ON(output == nullptr);
1419 
1420  // Create and configure function
1421  auto func = std::make_unique<PadLayerFunction>();
1422  func->configure(input, output, padding, pad_value);
1423 
1424  // Log info
1425  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1426  << node.name()
1427  << " Type: " << node.type()
1428  << " Target: " << TargetInfo::TargetType
1429  << " Data Type: " << input->info()->data_type()
1430  << " Input shape: " << input->info()->tensor_shape()
1431  << " Output shape: " << output->info()->tensor_shape()
1432  << std::endl);
1433 
1434  return std::move(func);
1435 }
1436 
1437 /** Create a backend permute layer function
1438  *
1439  * @tparam PermuteLayerFunction Backend permute function
1440  * @tparam TargetInfo Target-specific information
1441  *
1442  * @param[in] node Node to create the backend function for
1443  *
1444  * @return Backend permute layer function
1445  */
1446 template <typename PermuteLayerFunction, typename TargetInfo>
1447 std::unique_ptr<IFunction> create_permute_layer(PermuteLayerNode &node)
1448 {
1449  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1450 
1451  // Extract IO and info
1452  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1453  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1454  const PermutationVector &perm = node.permutation_vector();
1455  ARM_COMPUTE_ERROR_ON(input == nullptr);
1456  ARM_COMPUTE_ERROR_ON(output == nullptr);
1457 
1458  // Create and configure function
1459  auto func = std::make_unique<PermuteLayerFunction>();
1460  func->configure(input, output, perm);
1461 
1462  // Log info
1463  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1464  << node.name()
1465  << " Type: " << node.type()
1466  << " Target: " << TargetInfo::TargetType
1467  << " Data Type: " << input->info()->data_type()
1468  << " Input shape: " << input->info()->tensor_shape()
1469  << " Output shape: " << output->info()->tensor_shape()
1470  << " Permutation vector: " << perm
1471  << std::endl);
1472 
1473  return std::move(func);
1474 }
1475 
1476 /** Create a backend pooling layer function
1477  *
1478  * @tparam PoolingLayerFunction Backend pooling function
1479  * @tparam TargetInfo Target-specific information
1480  *
1481  * @param[in] node Node to create the backend function for
1482  *
1483  * @return Backend pooling layer function
1484  */
1485 template <typename PoolingLayerFunction, typename TargetInfo>
1486 std::unique_ptr<IFunction> create_pooling_layer(PoolingLayerNode &node)
1487 {
1488  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1489 
1490  // Extract IO and info
1491  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1492  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1493  const PoolingLayerInfo pool_info = node.pooling_info();
1494  ARM_COMPUTE_ERROR_ON(input == nullptr);
1495  ARM_COMPUTE_ERROR_ON(output == nullptr);
1496 
1497  // Create and configure function
1498  auto func = std::make_unique<PoolingLayerFunction>();
1499  func->configure(input, output, pool_info);
1500 
1501  // Log info
1502  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1503  << node.name()
1504  << " Type: " << node.type()
1505  << " Target: " << TargetInfo::TargetType
1506  << " Data Type: " << input->info()->data_type()
1507  << " Input shape: " << input->info()->tensor_shape()
1508  << " Output shape: " << output->info()->tensor_shape()
1509  << " Pooling info: " << pool_info.pool_type
1510  << std::endl);
1511 
1512  return std::move(func);
1513 }
1514 
1515 /** Create a backend PRelu layer function
1516  *
1517  * @tparam PReluFunction Backend PRelu function
1518  * @tparam TargetInfo Target-specific information
1519  *
1520  * @param[in] node Node to create the backend function for
1521  *
1522  * @return Backend PRelu layer function
1523  */
1524 template <typename PReluFunction, typename TargetInfo>
1525 std::unique_ptr<IFunction> create_prelu_layer(PReluLayerNode &node)
1526 {
1527  validate_node<TargetInfo>(node, 2 /* expected inputs */, 1 /* expected outputs */);
1528 
1529  // Extract IO and info
1530  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1531  typename TargetInfo::TensorType *alpha = get_backing_tensor<TargetInfo>(node.input(1));
1532  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1533  ARM_COMPUTE_ERROR_ON(input == nullptr || alpha == nullptr);
1534  ARM_COMPUTE_ERROR_ON(output == nullptr);
1535 
1536  // Create and configure function
1537  auto func = std::make_unique<PReluFunction>();
1538  func->configure(input, alpha, output);
1539 
1540  // Log info
1541  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1542  << node.name()
1543  << " Type: " << node.type()
1544  << " Target: " << TargetInfo::TargetType
1545  << " Data Type: " << input->info()->data_type()
1546  << " Input shape: " << input->info()->tensor_shape()
1547  << " Output shape: " << output->info()->tensor_shape()
1548  << std::endl);
1549 
1550  return std::move(func);
1551 }
1552 
1553 /** Create a backend print layer function
1554  *
1555  * @tparam TargetInfo Target-specific information
1556  *
1557  * @param[in] node Node to create the backend function for
1558  *
1559  * @return Backend print layer function
1560  */
1561 template <typename TargetInfo>
1562 std::unique_ptr<IFunction> create_print_layer(PrintLayerNode &node)
1563 {
1564  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1565 
1566  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1567  ARM_COMPUTE_ERROR_ON(input == nullptr);
1568  ARM_COMPUTE_UNUSED(input);
1569 
1570  // Log info
1571  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1572  << node.name()
1573  << " Type: " << node.type()
1574  << " Target: " << TargetInfo::TargetType
1575  << " Data Type: " << input->info()->data_type()
1576  << " Input shape: " << input->info()->tensor_shape()
1577  << std::endl);
1578 
1579  return nullptr;
1580 }
1581 
1582 /** Create a backend priorbox layer function
1583  *
1584  * @tparam PriorBoxLayerFunction Backend priorbox function
1585  * @tparam TargetInfo Target-specific information
1586  *
1587  * @param[in] node Node to create the backend function for
1588  *
1589  * @return Backend priorbox layer function
1590  */
1591 template <typename PriorBoxLayerFunction, typename TargetInfo>
1592 std::unique_ptr<IFunction> create_priorbox_layer(PriorBoxLayerNode &node)
1593 {
1594  validate_node<TargetInfo>(node, 2 /* expected inputs */, 1 /* expected outputs */);
1595 
1596  // Extract IO and info
1597  typename TargetInfo::TensorType *input0 = get_backing_tensor<TargetInfo>(node.input(0));
1598  typename TargetInfo::TensorType *input1 = get_backing_tensor<TargetInfo>(node.input(1));
1599  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1600  const PriorBoxLayerInfo prior_info = node.priorbox_info();
1601  ARM_COMPUTE_ERROR_ON(input0 == nullptr);
1602  ARM_COMPUTE_ERROR_ON(input1 == nullptr);
1603  ARM_COMPUTE_ERROR_ON(output == nullptr);
1604 
1605  // Create and configure function
1606  auto func = std::make_unique<PriorBoxLayerFunction>();
1607  func->configure(input0, input1, output, prior_info);
1608 
1609  // Log info
1610  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1611  << node.name()
1612  << " Type: " << node.type()
1613  << " Target: " << TargetInfo::TargetType
1614  << " Data Type: " << input0->info()->data_type()
1615  << " Input0 shape: " << input0->info()->tensor_shape()
1616  << " Input1 shape: " << input1->info()->tensor_shape()
1617  << " Output shape: " << output->info()->tensor_shape()
1618  << " PriorBoxLayer info: " << prior_info
1619  << std::endl);
1620 
1621  return std::move(func);
1622 }
1623 
1624 /** Create a backend quantization layer function
1625  *
1626  * @tparam QuantizationLayerFunction Backend quantization function
1627  * @tparam TargetInfo Target-specific information
1628  *
1629  * @param[in] node Node to create the backend function for
1630  *
1631  * @return Backend quantization layer function
1632  */
1633 template <typename QuantizationLayerFunction, typename TargetInfo>
1634 std::unique_ptr<IFunction> create_quantization_layer(QuantizationLayerNode &node)
1635 {
1636  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1637 
1638  // Extract IO and info
1639  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1640  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1641  ARM_COMPUTE_ERROR_ON(input == nullptr);
1642  ARM_COMPUTE_ERROR_ON(output == nullptr);
1643 
1644  // Create and configure function
1645  auto func = std::make_unique<QuantizationLayerFunction>();
1646  func->configure(input, output);
1647 
1648  // Log info
1649  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1650  << node.name()
1651  << " Type: " << node.type()
1652  << " Target: " << TargetInfo::TargetType
1653  << " Data Type: " << input->info()->data_type()
1654  << " Input shape: " << input->info()->tensor_shape()
1655  << " Output shape: " << output->info()->tensor_shape()
1656  << std::endl);
1657 
1658  return std::move(func);
1659 }
1660 
1661 /** Create a backend reduction operation layer function
1662  *
1663  * @tparam ReductionOperationFunction Backend reduction operation function
1664  * @tparam TargetInfo Target-specific information
1665  *
1666  * @param[in] node Node to create the backend function for
1667  * @param[in] ctx Graph context
1668  *
1669  * @return Backend reduction sum layer function
1670  */
1671 template <typename ReductionOperationFunction, typename TargetInfo>
1672 std::unique_ptr<IFunction> create_reduction_operation_layer(ReductionLayerNode &node, GraphContext &ctx)
1673 {
1674  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1675 
1676  // Extract IO and info
1677  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1678  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1679  ReductionOperation op = node.op();
1680  int axis = node.axis();
1681  bool keep_dims = node.keep_dims();
1682  ARM_COMPUTE_ERROR_ON(input == nullptr);
1683  ARM_COMPUTE_ERROR_ON(output == nullptr);
1684 
1685  // Create and configure function
1686  auto func = std::make_unique<ReductionOperationFunction>(get_memory_manager(ctx, TargetInfo::TargetType));
1687  func->configure(input, output, axis, op, keep_dims);
1688 
1689  // Log info
1690  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1691  << node.name()
1692  << " Type: " << node.type()
1693  << " Target: " << TargetInfo::TargetType
1694  << " Data Type: " << input->info()->data_type()
1695  << " Input shape: " << input->info()->tensor_shape()
1696  << " Output shape: " << output->info()->tensor_shape()
1697  << " Operation: " << op
1698  << " Axis: " << axis
1699  << " Keep dimensions:" << keep_dims
1700  << std::endl);
1701 
1702  return std::move(func);
1703 }
1704 
1705 /** Create a backend reorg layer function
1706  *
1707  * @tparam ReorgLayerFunction Backend reorg function
1708  * @tparam TargetInfo Target-specific information
1709  *
1710  * @param[in] node Node to create the backend function for
1711  *
1712  * @return Backend reshape layer function
1713  */
1714 template <typename ReorgLayerFunction, typename TargetInfo>
1715 std::unique_ptr<IFunction> create_reorg_layer(ReorgLayerNode &node)
1716 {
1717  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1718 
1719  // Extract IO and info
1720  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1721  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1722  ARM_COMPUTE_ERROR_ON(input == nullptr);
1723  ARM_COMPUTE_ERROR_ON(output == nullptr);
1724 
1725  // Create and configure function
1726  auto func = std::make_unique<ReorgLayerFunction>();
1727  func->configure(input, output, node.stride());
1728 
1729  // Log info
1730  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1731  << node.name()
1732  << " Type: " << node.type()
1733  << " Target: " << TargetInfo::TargetType
1734  << " Data Type: " << input->info()->data_type()
1735  << " Input shape: " << input->info()->tensor_shape()
1736  << " Output shape: " << output->info()->tensor_shape()
1737  << std::endl);
1738 
1739  return std::move(func);
1740 }
1741 
1742 /** Create a backend reshape layer function
1743  *
1744  * @tparam ReshapeLayerFunction Backend reshape function
1745  * @tparam TargetInfo Target-specific information
1746  *
1747  * @param[in] node Node to create the backend function for
1748  *
1749  * @return Backend reshape layer function
1750  */
1751 template <typename ReshapeLayerFunction, typename TargetInfo>
1752 std::unique_ptr<IFunction> create_reshape_layer(ReshapeLayerNode &node)
1753 {
1754  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1755 
1756  // Extract IO and info
1757  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1758  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1759  ARM_COMPUTE_ERROR_ON(input == nullptr);
1760  ARM_COMPUTE_ERROR_ON(output == nullptr);
1761 
1762  // Create and configure function
1763  auto func = std::make_unique<ReshapeLayerFunction>();
1764  func->configure(input, output);
1765 
1766  // Log info
1767  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1768  << node.name()
1769  << " Type: " << node.type()
1770  << " Target: " << TargetInfo::TargetType
1771  << " Data Type: " << input->info()->data_type()
1772  << " Input shape: " << input->info()->tensor_shape()
1773  << " Output shape: " << output->info()->tensor_shape()
1774  << std::endl);
1775 
1776  return std::move(func);
1777 }
1778 
1779 /** Create a backend resize layer function
1780  *
1781  * @tparam ResizeLayerFunction Backend resize function
1782  * @tparam TargetInfo Target-specific information
1783  *
1784  * @param[in] node Node to create the backend function for
1785  *
1786  * @return Backend resize layer function
1787  */
1788 template <typename ResizeLayerFunction, typename TargetInfo>
1789 std::unique_ptr<IFunction> create_resize_layer(ResizeLayerNode &node)
1790 {
1791  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1792 
1793  // Extract IO and info
1794  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1795  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1796  ARM_COMPUTE_ERROR_ON(input == nullptr);
1797  ARM_COMPUTE_ERROR_ON(output == nullptr);
1798  const InterpolationPolicy policy = node.policy();
1799 
1800  // Create and configure function
1801  auto func = std::make_unique<ResizeLayerFunction>();
1802  func->configure(input, output, ScaleKernelInfo{ policy, BorderMode::CONSTANT, PixelValue(), SamplingPolicy::CENTER, false, false });
1803 
1804  // Log info
1805  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1806  << node.name()
1807  << " Type: " << node.type()
1808  << " Target: " << TargetInfo::TargetType
1809  << " Data Type: " << input->info()->data_type()
1810  << " Input shape: " << input->info()->tensor_shape()
1811  << " Output shape: " << output->info()->tensor_shape()
1812  << " Interpolation: " << policy
1813  << std::endl);
1814 
1815  return std::move(func);
1816 }
1817 
1818 /** Create a backend ROI align layer function
1819  *
1820  * @tparam ROIAlignLayerFunction ROI Align function
1821  * @tparam TargetInfo Target-specific information
1822  *
1823  * @param[in] node Node to create the backend function for
1824  *
1825  * @return ROI Align layer function
1826  */
1827 template <typename ROIAlignLayerFunction, typename TargetInfo>
1828 std::unique_ptr<IFunction> create_roi_align_layer(ROIAlignLayerNode &node)
1829 {
1830  validate_node<TargetInfo>(node, 2 /* expected inputs */, 1 /* expected outputs */);
1831 
1832  // Extract IO and info
1833  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1834  typename TargetInfo::TensorType *rois = get_backing_tensor<TargetInfo>(node.input(1));
1835  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1836  ARM_COMPUTE_ERROR_ON(input == nullptr);
1837  ARM_COMPUTE_ERROR_ON(output == nullptr);
1838  ARM_COMPUTE_ERROR_ON(rois == nullptr);
1839 
1840  const ROIPoolingLayerInfo pool_info = node.pooling_info();
1841 
1842  // Create and configure function
1843  auto func = std::make_unique<ROIAlignLayerFunction>();
1844 
1845  func->configure(input, rois, output, pool_info);
1846 
1847  // Log info
1848  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1849  << node.name()
1850  << " Type: " << node.type()
1851  << " Target: " << TargetInfo::TargetType
1852  << " Data Type: " << input->info()->data_type()
1853  << " Input shape: " << input->info()->tensor_shape()
1854  << " Output shape: " << output->info()->tensor_shape()
1855  << " ROIs shape: " << rois->info()->tensor_shape()
1856  << " ROIPooling width: " << pool_info.pooled_width()
1857  << " ROIPooling height: " << pool_info.pooled_height()
1858  << std::endl);
1859 
1860  return std::move(func);
1861 }
1862 
1863 /** Create a backend slice layer function
1864  *
1865  * @tparam SliceLayerFunction Backend slice function
1866  * @tparam TargetInfo Target-specific information
1867  *
1868  * @param[in] node Node to create the backend function for
1869  *
1870  * @return Backend slice layer function
1871  */
1872 template <typename SliceLayerFunction, typename TargetInfo>
1873 std::unique_ptr<IFunction> create_slice_layer(SliceLayerNode &node)
1874 {
1875  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1876 
1877  // Extract IO and info
1878  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1879  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1880  ARM_COMPUTE_ERROR_ON(input == nullptr);
1881  ARM_COMPUTE_ERROR_ON(output == nullptr);
1882 
1883  // Create and configure function
1884  auto func = std::make_unique<SliceLayerFunction>();
1885  func->configure(input, output, node.starts(), node.ends());
1886 
1887  // Log info
1888  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1889  << node.name()
1890  << " Type: " << node.type()
1891  << " Target: " << TargetInfo::TargetType
1892  << " Data Type: " << input->info()->data_type()
1893  << " Input shape: " << input->info()->tensor_shape()
1894  << " Output shape: " << output->info()->tensor_shape()
1895  << std::endl);
1896 
1897  return std::move(func);
1898 }
1899 
1900 /** Create a backend softmax layer function
1901  *
1902  * @tparam SoftmaxLayerFunction Backend softmax function
1903  * @tparam TargetInfo Target-specific information
1904  *
1905  * @param[in] node Node to create the backend function for
1906  * @param[in] ctx Graph context
1907  *
1908  * @return Backend softmax layer function
1909  */
1910 template <typename SoftmaxLayerFunction, typename TargetInfo>
1911 std::unique_ptr<IFunction> create_softmax_layer(SoftmaxLayerNode &node, GraphContext &ctx)
1912 {
1913  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1914 
1915  // Extract IO and info
1916  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1917  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1918  const float beta = node.beta();
1919  ARM_COMPUTE_ERROR_ON(input == nullptr);
1920  ARM_COMPUTE_ERROR_ON(output == nullptr);
1921 
1922  // Create and configure function
1923  auto func = std::make_unique<SoftmaxLayerFunction>(get_memory_manager(ctx, TargetInfo::TargetType));
1924  func->configure(input, output, beta);
1925 
1926  // Log info
1927  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1928  << node.name()
1929  << " Type: " << node.type()
1930  << " Target: " << TargetInfo::TargetType
1931  << " Data Type: " << input->info()->data_type()
1932  << " Input shape: " << input->info()->tensor_shape()
1933  << " Output shape: " << output->info()->tensor_shape()
1934  << std::endl);
1935 
1936  return std::move(func);
1937 }
1938 
1939 /** Create a backend layer stack function
1940  *
1941  * @tparam StackLayerFunction Backend stack function
1942  * @tparam TargetInfo Target-specific information
1943  *
1944  * @param[in] node Node to create the backend function for
1945  *
1946  * @return Backend stack layer function
1947  */
1948 template <typename StackLayerFunction, typename TargetInfo>
1949 std::unique_ptr<arm_compute::IFunction> create_stack_layer(StackLayerNode &node)
1950 {
1951  ARM_COMPUTE_LOG_GRAPH_VERBOSE("Creating Stack node with ID : " << node.id() << " and Name: " << node.name() << std::endl);
1952  ARM_COMPUTE_ERROR_ON(node.num_outputs() != 1);
1953 
1954  // Extract IO and info
1955  std::vector<typename TargetInfo::TensorType *> inputs;
1956  for(unsigned int i = 0; i < node.num_inputs(); ++i)
1957  {
1958  inputs.push_back(get_backing_tensor<TargetInfo>(node.input(i)));
1959  }
1960  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1961  const int axis = node.axis();
1962 
1963  // Create and configure function
1964  auto func = std::make_unique<StackLayerFunction>();
1965  func->configure(inputs, axis, output);
1966 
1967  // Log info
1968  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
1969  << node.name()
1970  << " Type: " << node.type()
1971  << " Target: " << TargetInfo::TargetType
1972  << " Data Type: " << output->info()->data_type()
1973  << " Inputs shape: " << inputs[0]->info()->tensor_shape()
1974  << " Output shape: " << output->info()->tensor_shape()
1975  << " Num Inputs: " << inputs.size()
1976  << " Axis: " << axis
1977  << std::endl);
1978 
1979  return std::move(func);
1980 }
1981 
1982 /** Create a backend slice layer function
1983  *
1984  * @tparam StridedSliceLayerFunction Backend strided slice function
1985  * @tparam TargetInfo Target-specific information
1986  *
1987  * @param[in] node Node to create the backend function for
1988  *
1989  * @return Backend strided slice layer function
1990  */
1991 template <typename StridedSliceLayerFunction, typename TargetInfo>
1992 std::unique_ptr<IFunction> create_strided_slice_layer(StridedSliceLayerNode &node)
1993 {
1994  validate_node<TargetInfo>(node, 1 /* expected inputs */, 1 /* expected outputs */);
1995 
1996  // Extract IO and info
1997  typename TargetInfo::TensorType *input = get_backing_tensor<TargetInfo>(node.input(0));
1998  typename TargetInfo::TensorType *output = get_backing_tensor<TargetInfo>(node.output(0));
1999  Coordinates starts = node.starts();
2000  Coordinates ends = node.ends();
2001  BiStrides strides = node.strides();
2002  StridedSliceLayerInfo info = node.strided_slice_info();
2003 
2004  ARM_COMPUTE_ERROR_ON(input == nullptr);
2005  ARM_COMPUTE_ERROR_ON(output == nullptr);
2006 
2007  // Create and configure function
2008  auto func = std::make_unique<StridedSliceLayerFunction>();
2009  func->configure(input, output, starts, ends, strides, info.begin_mask(), info.end_mask(), info.shrink_axis_mask());
2010 
2011  // Log info
2012  ARM_COMPUTE_LOG_GRAPH_INFO("Instantiated "
2013  << node.name()
2014  << " Type: " << node.type()
2015  << " Target: " << TargetInfo::TargetType
2016  << " Data Type: " << input->info()->data_type()
2017  << " Input shape: " << input->info()->tensor_shape()
2018  << " Output shape: " << output->info()->tensor_shape()
2019  << std::endl);
2020 
2021  return std::move(func);
2022 }
2023 } // namespace detail
2024 } // namespace backends
2025 } // namespace graph
2026 } // namespace arm_compute
2027 
2028 #endif /* ARM_COMPUTE_GRAPH_BACKENDS_DETAIL_FUNCTION_HELPERS_H */
arm_compute::graph::backends::detail::create_normalization_layer
std::unique_ptr< IFunction > create_normalization_layer(NormalizationLayerNode &node, GraphContext &ctx)
Create a backend normalization layer function.
Definition: FunctionHelpers.h:1327
arm_compute::graph::backends::FusedConvolutionBatchNormalizationWithPostOpsFunction
Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE...
Definition: FusedConvolutionBatchNormalizationWithPostOpsFunction.h:40
arm_compute::graph::backends::detail::create_dequantization_layer
std::unique_ptr< IFunction > create_dequantization_layer(DequantizationLayerNode &node)
Create a backend dequantize layer function.
Definition: FunctionHelpers.h:876
arm_compute::graph::INode::name
std::string name() const
Returns node&#39;s name.
Definition: INode.cpp:108
arm_compute::PixelValue
Class describing the value of a pixel for any image format.
Definition: PixelValue.h:34
arm_compute::InterpolationPolicy
InterpolationPolicy
Interpolation method.
Definition: Types.h:411
arm_compute::graph::EltwiseLayerNode
Eltwise Layer node.
Definition: EltwiseLayerNode.h:34
arm_compute::graph::backends::detail::create_arg_min_max_layer
std::unique_ptr< IFunction > create_arg_min_max_layer(ArgMinMaxLayerNode &node)
Creates a backend argminmax layer function.
Definition: FunctionHelpers.h:140
arm_compute::GenerateProposalsInfo
Generate Proposals Information class.
Definition: Types.h:1441
arm_compute::test::validation::data_layout
const auto data_layout
Definition: ConvolutionLayer.cpp:403
arm_compute::test::validation::post_ops
experimental::PostOpList< ITensorInfo * > post_ops
Definition: ConvolutionLayer.cpp:410
arm_compute::graph::EltwiseOperation
EltwiseOperation
Supported Element-wise operations.
Definition: Types.h:109
arm_compute::graph::ConvolutionMethod::Direct
Deep direct convolution.
arm_compute::graph::backends::detail::create_slice_layer
std::unique_ptr< IFunction > create_slice_layer(SliceLayerNode &node)
Create a backend slice layer function.
Definition: FunctionHelpers.h:1873
arm_compute::graph::BatchNormalizationLayerNode
Batch Normalization Layer node.
Definition: BatchNormalizationLayerNode.h:34
arm_compute::graph::FusedDepthwiseConvolutionBatchNormalizationNode
Fused Depthwise Convolution Batch Normalization node.
Definition: FusedDepthwiseConvolutionBatchNormalizationNode.h:34
arm_compute::graph::NormalizationLayerNode
Normalization Layer node.
Definition: NormalizationLayerNode.h:34
arm_compute::StridedSliceLayerInfo::end_mask
int32_t end_mask() const
Definition: Types.h:1947
arm_compute::graph::backends::detail::create_eltwise_layer
std::unique_ptr< IFunction > create_eltwise_layer(EltwiseLayerNode &node)
Create a backend element-wise operation layer function.
Definition: FunctionHelpers.h:1015
arm_compute::ActivationLayerInfo::enabled
bool enabled() const
Check if initialised.
Definition: Types.h:1694
arm_compute::PaddingList
std::vector< PaddingInfo > PaddingList
List of padding information.
Definition: Types.h:453
arm_compute::ReductionOperation
ReductionOperation
Available reduction operations.
Definition: Types.h:476
arm_compute::graph::ConcatenateLayerNode::concatenation_axis
DataLayoutDimension concatenation_axis() const
Concatenation axis parameter accessor.
Definition: ConcatenateLayerNode.cpp:54
arm_compute::graph::Tensor::handle
ITensorHandle * handle()
Backend tensor handle accessor.
Definition: Tensor.cpp:55
arm_compute::graph::backends::detail::create_batch_normalization_layer
std::unique_ptr< IFunction > create_batch_normalization_layer(BatchNormalizationLayerNode &node)
Create a backend batch normalization layer function.
Definition: FunctionHelpers.h:177
arm_compute::graph::backends::detail::create_depth_to_space_layer
std::unique_ptr< IFunction > create_depth_to_space_layer(DepthToSpaceLayerNode &node)
Create a backend depth to space layer function.
Definition: FunctionHelpers.h:837
ARM_COMPUTE_ERROR
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
arm_compute::graph::backends::detail::create_fused_convolution_batch_normalization_layer
std::unique_ptr< IFunction > create_fused_convolution_batch_normalization_layer(FusedConvolutionBatchNormalizationNode &node, GraphContext &ctx)
Create a backend batch normalization layer function.
Definition: FunctionHelpers.h:222
arm_compute::graph::INode::assigned_target
Target assigned_target() const
Returns assigned target for this node.
Definition: INode.cpp:199
arm_compute::TensorType
TensorType
Memory type.
Definition: Types.h:38
arm_compute::graph::backends::detail::create_prelu_layer
std::unique_ptr< IFunction > create_prelu_layer(PReluLayerNode &node)
Create a backend PRelu layer function.
Definition: FunctionHelpers.h:1525
arm_compute::test::validation::conv_info
const auto conv_info
Definition: ConvolutionLayer.cpp:404
arm_compute::graph::backends::detail::create_depthwise_convolution_layer
std::unique_ptr< IFunction > create_depthwise_convolution_layer(DepthwiseConvolutionLayerNode &node)
Create a backend layer depth-wise convolution function.
Definition: FunctionHelpers.h:775
Types.h
arm_compute::graph::INode::num_outputs
size_t num_outputs() const
Returns number of outputs of the node.
Definition: INode.cpp:184
arm_compute::graph::backends::detail::create_resize_layer
std::unique_ptr< IFunction > create_resize_layer(ResizeLayerNode &node)
Create a backend resize layer function.
Definition: FunctionHelpers.h:1789
arm_compute::graph::backends::FusedConvolutionBatchNormalizationFunction
Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE...
Definition: FusedConvolutionBatchNormalizationFunction.h:39
arm_compute::NormalizationLayerInfo
Normalization Layer Information class.
Definition: Types.h:1847
arm_compute::graph::NormalizePlanarYUVLayerNode
Batch Normalization Layer node.
Definition: NormalizePlanarYUVLayerNode.h:34
arm_compute::graph::FlattenLayerNode
Flatten Layer node.
Definition: FlattenLayerNode.h:34
arm_compute::graph::backends::detail::create_permute_layer
std::unique_ptr< IFunction > create_permute_layer(PermuteLayerNode &node)
Create a backend permute layer function.
Definition: FunctionHelpers.h:1447
ARM_COMPUTE_ERROR_ON
#define ARM_COMPUTE_ERROR_ON(cond)
If the condition is true then an error message is printed and an exception thrown.
Definition: Error.h:466
arm_compute::graph::backends::detail::create_generate_proposals_layer
std::unique_ptr< IFunction > create_generate_proposals_layer(GenerateProposalsLayerNode &node, GraphContext &ctx)
Create a backend generate proposals layer function.
Definition: FunctionHelpers.h:1234
arm_compute::graph::StackLayerNode::type
NodeType type() const override
Returns node&#39;s type.
Definition: StackLayerNode.cpp:105
arm_compute::graph::backends::detail::create_l2_normalize_layer
std::unique_ptr< IFunction > create_l2_normalize_layer(L2NormalizeLayerNode &node, GraphContext &ctx)
Create a backend l2 normalization layer function.
Definition: FunctionHelpers.h:1283
arm_compute::FullyConnectedLayerInfo
Fully connected layer info.
Definition: Types.h:1809
arm_compute::graph::FastMathHint::Enabled
Fast math enabled for Convolution layer.
arm_compute::graph::ReorgLayerNode
Reorg Layer node.
Definition: ReorgLayerNode.h:34
arm_compute::graph::backends::detail::create_flatten_layer
std::unique_ptr< IFunction > create_flatten_layer(FlattenLayerNode &node)
Create a backend flatten layer function.
Definition: FunctionHelpers.h:1138
arm_compute::ROIPoolingLayerInfo::pooled_width
unsigned int pooled_width() const
Get the pooled width of the layer.
Definition: Types.h:1413
arm_compute::graph::backends::FusedDepthwiseConvolutionBatchNormalizationFunction
Wrapper function to first apply {NE, CL}BatchNormalizationLayer on the weights and then run {NE...
Definition: FusedDepthwiseConvolutionBatchNormalizationFunction.h:39
arm_compute::graph::backends::detail::create_pad_layer
std::unique_ptr< IFunction > create_pad_layer(PadLayerNode &node)
Create a backend pad layer function.
Definition: FunctionHelpers.h:1408
arm_compute::graph::SliceLayerNode
Slice Layer node.
Definition: SliceLayerNode.h:36
arm_compute::graph::GraphContext
Graph context.
Definition: GraphContext.h:57
arm_compute::ActivationLayerInfo
Activation Layer Information class.
Definition: Types.h:1639
arm_compute::graph::PermuteLayerNode
Permute Layer node.
Definition: PermuteLayerNode.h:34
arm_compute::graph::PoolingLayerNode
Pooling Layer node.
Definition: PoolingLayerNode.h:34
arm_compute::FullyConnectedLayerInfo::enable_fast_math
bool enable_fast_math
Enable fast math computation.
Definition: Types.h:1818
ARM_COMPUTE_LOG_GRAPH_INFO
#define ARM_COMPUTE_LOG_GRAPH_INFO(x)
Definition: Logger.h:54
arm_compute::graph::backends::detail::create_stack_layer
std::unique_ptr< arm_compute::IFunction > create_stack_layer(StackLayerNode &node)
Create a backend layer stack function.
Definition: FunctionHelpers.h:1949
arm_compute
Copyright (c) 2017-2022 Arm Limited.
Definition: introduction.dox:24
arm_compute::graph::FusedConvolutionBatchNormalizationNode
Batch Normalization node.
Definition: FusedConvolutionBatchNormalizationNode.h:34
arm_compute::SamplingPolicy::CENTER
Samples are taken at pixel center.
FusedConvolutionBatchNormalizationFunction.h
arm_compute::graph::EltwiseOperation::Sub
Arithmetic subtraction.
arm_compute::graph::BoundingBoxTransformLayerNode
Bounding Box Transform Layer node.
Definition: BoundingBoxTransformLayerNode.h:35
arm_compute::test::validation::input
auto input
Definition: LSTMLayerQuantized.cpp:486
Cast.h
arm_compute::graph::PadLayerNode
Pad Layer node.
Definition: PadLayerNode.h:34
arm_compute::WeightsInfo
Convolution Layer Weights Information class.
Definition: Types.h:2073
arm_compute::graph::backends::detail::create_softmax_layer
std::unique_ptr< IFunction > create_softmax_layer(SoftmaxLayerNode &node, GraphContext &ctx)
Create a backend softmax layer function.
Definition: FunctionHelpers.h:1911
arm_compute::Format::S32
1 channel, 1 S32 per channel
arm_compute::graph::UnaryEltwiseOperation::Exp
Exp.
arm_compute::graph::Tensor::desc
TensorDescriptor & desc()
TensorInfo metadata accessor.
Definition: Tensor.cpp:40
arm_compute::graph::ConvolutionLayerNode
Convolution Layer node.
Definition: ConvolutionLayerNode.h:34
arm_compute::graph::StackLayerNode
Stack Layer node.
Definition: StackLayerNode.h:34
arm_compute::graph::EltwiseOperation::Div
Arithmetic division.
arm_compute::graph::backends::detail::create_fused_depthwise_convolution_batch_normalization_layer
std::unique_ptr< IFunction > create_fused_depthwise_convolution_batch_normalization_layer(FusedDepthwiseConvolutionBatchNormalizationNode &node, GraphContext &ctx)
Create a backend fused depthwise convolution batch normalization layer function.
Definition: FunctionHelpers.h:279
arm_compute::graph::ConvolutionMethod::GEMM
GEMM based convolution.
arm_compute::graph::INode
Node interface.
Definition: INode.h:46
arm_compute::graph::backends::detail::create_convolution_layer
std::unique_ptr< IFunction > create_convolution_layer(ConvolutionLayerNode &node, GraphContext &ctx)
Create a backend convolution layer function.
Definition: FunctionHelpers.h:464
ARM_COMPUTE_UNUSED
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
arm_compute::graph::backends::detail::create_roi_align_layer
std::unique_ptr< IFunction > create_roi_align_layer(ROIAlignLayerNode &node)
Create a backend ROI align layer function.
Definition: FunctionHelpers.h:1828
arm_compute::graph::QuantizationLayerNode
Quantization Layer node.
Definition: QuantizationLayerNode.h:34
arm_compute::graph::ConcatenateLayerNode::is_enabled
bool is_enabled() const
Enabled parameter accessor.
Definition: ConcatenateLayerNode.cpp:49
arm_compute::graph::INode::output
Tensor * output(size_t idx) const
Returns the tensor of a given output of the node.
Definition: INode.cpp:159
Utils.h
arm_compute::graph::PriorBoxLayerNode
PriorBox Layer node.
Definition: PriorBoxLayerNode.h:34
arm_compute::graph::DequantizationLayerNode
Dequantize Layer node.
Definition: DequantizationLayerNode.h:38
ARM_COMPUTE_ERROR_ON_MSG
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
Definition: Error.h:456
arm_compute::test::validation::num_groups
const unsigned int num_groups
Definition: Im2Col.cpp:153
arm_compute::graph::EltwiseOperation::Max
Arithmetic maximum.
arm_compute::Coordinates
Coordinates of an item.
Definition: Coordinates.h:37
FusedDepthwiseConvolutionBatchNormalizationFunction.h
arm_compute::utils::cast::U
U
Definition: SaturateCast.h:57
arm_compute::PoolingLayerInfo
Pooling Layer Information struct.
Definition: Types.h:1200
arm_compute::graph::backends::detail::validate_node
void validate_node(const INode &node, size_t num_expected_inputs, size_t num_expected_outputs)
Definition: FunctionHelpers.h:78
arm_compute::graph::backends::detail::create_reduction_operation_layer
std::unique_ptr< IFunction > create_reduction_operation_layer(ReductionLayerNode &node, GraphContext &ctx)
Create a backend reduction operation layer function.
Definition: FunctionHelpers.h:1672
arm_compute::graph::backends::detail::create_priorbox_layer
std::unique_ptr< IFunction > create_priorbox_layer(PriorBoxLayerNode &node)
Create a backend priorbox layer function.
Definition: FunctionHelpers.h:1592
arm_compute::graph::INode::id
NodeID id() const
Returns node&#39;s ID.
Definition: INode.cpp:103
PostOps.h
arm_compute::PriorBoxLayerInfo
PriorBox layer info.
Definition: Types.h:825
arm_compute::graph::backends::detail::create_detection_output_layer
std::unique_ptr< IFunction > create_detection_output_layer(DetectionOutputLayerNode &node)
Create a backend detection output layer function.
Definition: FunctionHelpers.h:914
ITensorInfo.h
arm_compute::graph::ChannelShuffleLayerNode
Channel Shuffle Layer node.
Definition: ChannelShuffleLayerNode.h:34
arm_compute::graph::EltwiseOperation::Add
Arithmetic addition.
arm_compute::graph::ReshapeLayerNode
Reshape Layer node.
Definition: ReshapeLayerNode.h:34
arm_compute::PadStrideInfo
Padding and stride information class.
Definition: Types.h:669
arm_compute::graph::ConcatenateLayerNode
Concatenation Layer node.
Definition: ConcatenateLayerNode.h:34
arm_compute::graph::ArgMinMaxLayerNode
Arg Min/Max Layer node.
Definition: ArgMinMaxLayerNode.h:34
Error.h
arm_compute::to_string
std::string to_string(const T &val)
Fallback method: try to use std::to_string:
Definition: TypePrinter.h:80
arm_compute::experimental::PostOpType::Activation
arm_compute::graph::FusedConvolutionBatchNormalizationWithPostOpsNode
Batch Normalization node.
Definition: FusedConvolutionBatchNormalizationWithPostOpsNode.h:34
arm_compute::BoundingBoxTransformInfo
Bounding Box Transform information class.
Definition: Types.h:1572
arm_compute::graph::SoftmaxLayerNode
Softmax Layer node.
Definition: SoftmaxLayerNode.h:34
arm_compute::graph::L2NormalizeLayerNode
L2Normalize Layer node.
Definition: L2NormalizeLayerNode.h:34
arm_compute::graph::DepthToSpaceLayerNode
DepthToSpace Layer node.
Definition: DepthToSpaceLayerNode.h:34
arm_compute::StridedSliceLayerInfo
Definition: Types.h:1926
arm_compute::graph::backends::detail::create_detection_post_process_layer
std::unique_ptr< IFunction > create_detection_post_process_layer(DetectionPostProcessLayerNode &node)
Create a backend detection post process layer function.
Definition: FunctionHelpers.h:960
arm_compute::graph::ITensorHandle
Tensor handle interface object.
Definition: ITensorHandle.h:38
arm_compute::graph::backends::detail::create_deconvolution_layer
std::unique_ptr< IFunction > create_deconvolution_layer(DeconvolutionLayerNode &node, GraphContext &ctx)
Create a backend deconvolution layer function.
Definition: FunctionHelpers.h:732
arm_compute::graph::TensorDescriptor::target
Target target
Target.
Definition: TensorDescriptor.h:113
arm_compute::graph::ReductionLayerNode
Reduction Operation node.
Definition: ReductionLayerNode.h:34
arm_compute::graph::backends::detail::create_unary_eltwise_layer
std::unique_ptr< IFunction > create_unary_eltwise_layer(UnaryEltwiseLayerNode &node)
Create a backend unary element-wise operation layer function.
Definition: FunctionHelpers.h:1090
arm_compute::is_data_type_quantized_asymmetric
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1052
arm_compute::graph::backends::is_in_place_operation
bool is_in_place_operation(void *input, void *output)
Checks if an operation is in place.
Definition: Utils.h:77
arm_compute::Strides
Strides of an item in bytes.
Definition: Strides.h:37
arm_compute::graph::backends::detail::create_print_layer
std::unique_ptr< IFunction > create_print_layer(PrintLayerNode &node)
Create a backend print layer function.
Definition: FunctionHelpers.h:1562
arm_compute::graph::ROIAlignLayerNode
ROI Align node.
Definition: ROIAlignLayerNode.h:36
arm_compute::graph::PReluLayerNode
PRelu Layer node.
Definition: PReluLayerNode.h:34
arm_compute::GPUTarget::UNKNOWN
arm_compute::DetectionOutputLayerInfo
Detection Output layer info.
Definition: Types.h:962
arm_compute::graph::DetectionPostProcessLayerNode
DetectionPostProcess Layer node.
Definition: DetectionPostProcessLayerNode.h:34
TypePrinter.h
arm_compute::graph::ActivationLayerNode
Activation Layer node.
Definition: ActivationLayerNode.h:34
arm_compute::test::validation::info
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
arm_compute::graph::backends::detail::create_fully_connected_layer
std::unique_ptr< IFunction > create_fully_connected_layer(FullyConnectedLayerNode &node, GraphContext &ctx)
Create a backend fully connected layer function.
Definition: FunctionHelpers.h:1177
arm_compute::graph::backends::detail::create_activation_layer
std::unique_ptr< IFunction > create_activation_layer(ActivationLayerNode &node)
Creates a backend activation layer function.
Definition: FunctionHelpers.h:102
arm_compute::ScaleKernelInfo
Definition: KernelDescriptors.h:187
arm_compute::graph::DetectionOutputLayerNode
DetectionOutput Layer node.
Definition: DetectionOutputLayerNode.h:34
arm_compute::graph::StackLayerNode::axis
int axis() const
Stack axis parameter accessor.
Definition: StackLayerNode.cpp:45
arm_compute::StridedSliceLayerInfo::shrink_axis_mask
int32_t shrink_axis_mask() const
Definition: Types.h:1953
IPostOp.h
arm_compute::graph::backends::detail::create_strided_slice_layer
std::unique_ptr< IFunction > create_strided_slice_layer(StridedSliceLayerNode &node)
Create a backend slice layer function.
Definition: FunctionHelpers.h:1992
arm_compute::PaddingMode::CONSTANT
arm_compute::ROIPoolingLayerInfo::pooled_height
unsigned int pooled_height() const
Get the pooled height of the layer.
Definition: Types.h:1418
arm_compute::ROIPoolingLayerInfo
ROI Pooling Layer Information class.
Definition: Types.h:1398
arm_compute::Size2D
Class for specifying the size of an image or rectangle.
Definition: Size2D.h:34
arm_compute::graph::backends::detail::create_quantization_layer
std::unique_ptr< IFunction > create_quantization_layer(QuantizationLayerNode &node)
Create a backend quantization layer function.
Definition: FunctionHelpers.h:1634
arm_compute::graph::UnaryEltwiseLayerNode
Unary Eltwise Layer node.
Definition: EltwiseLayerNode.h:91
arm_compute::graph::backends::get_memory_manager
std::shared_ptr< IMemoryManager > get_memory_manager(GraphContext &ctx, Target target)
Returns the memory manager for a given target.
Definition: Utils.h:89
arm_compute::graph::backends::detail::create_normalize_planar_yuv_layer
std::unique_ptr< IFunction > create_normalize_planar_yuv_layer(NormalizePlanarYUVLayerNode &node)
Create a backend normalize planar YUV layer function.
Definition: FunctionHelpers.h:1368
ARM_COMPUTE_LOG_GRAPH_VERBOSE
#define ARM_COMPUTE_LOG_GRAPH_VERBOSE(x)
Definition: Logger.h:50
arm_compute::graph::ConvolutionMethod
ConvolutionMethod
Supported Convolution layer methods.
Definition: Types.h:126
arm_compute::graph::backends::detail::create_bounding_box_transform_layer
std::unique_ptr< IFunction > create_bounding_box_transform_layer(BoundingBoxTransformLayerNode &node)
Create a backend bounding box transform layer function.
Definition: FunctionHelpers.h:334
arm_compute::DetectionPostProcessLayerInfo
Detection Output layer info.
Definition: Types.h:1081
arm_compute::graph::INode::num_inputs
size_t num_inputs() const
Returns number of inputs of the node.
Definition: INode.cpp:179
arm_compute::graph::GenerateProposalsLayerNode
Generate Proposals Layer node.
Definition: GenerateProposalsLayerNode.h:35
arm_compute::graph::backends::detail::get_backing_tensor
TargetInfo::TensorType * get_backing_tensor(arm_compute::graph::Tensor *tensor)
Returns backing tensor of a given tensor.
Definition: FunctionHelpers.h:62
tensor
CLTensor * tensor
Pointer to the auxiliary tensor.
Definition: ClWorkloadRuntime.cpp:66
arm_compute::graph::DeconvolutionLayerNode
Deconvolution Layer node.
Definition: DeconvolutionLayerNode.h:35
Helpers.h
arm_compute::graph::FullyConnectedLayerNode
Fully Connected Layer node.
Definition: FullyConnectedLayerNode.h:34
arm_compute::graph::DepthwiseConvolutionLayerNode
Depthwise Convolution Layer node.
Definition: DepthwiseConvolutionLayerNode.h:34
Nodes.h
arm_compute::graph::StridedSliceLayerNode
Slice Layer node.
Definition: StridedSliceLayerNode.h:36
arm_compute::graph::ConvolutionMethod::Winograd
Winograd based convolution.
arm_compute::ActivationLayerInfo::activation
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1679
arm_compute::graph::ResizeLayerNode
Resize Layer node.
Definition: ResizeLayerNode.h:34
arm_compute::graph::INode::type
virtual NodeType type() const =0
Returns node&#39;s type.
Utils.h
arm_compute::graph::EltwiseOperation::Mul
Arithmetic multiplication.
arm_compute::graph::PrintLayerNode
Print Layer node.
Definition: PrintLayerNode.h:37
arm_compute::graph::backends::get_weights_manager
std::shared_ptr< IWeightsManager > get_weights_manager(GraphContext &ctx, Target target)
Returns the weights manager for a given target.
Definition: Utils.h:102
arm_compute::graph::INode::input
Tensor * input(size_t idx) const
Returns the tensor of a given input of the node.
Definition: INode.cpp:151
arm_compute::graph::ConcatenateLayerNode::type
NodeType type() const override
Returns node&#39;s type.
Definition: ConcatenateLayerNode.cpp:130
arm_compute::graph::TensorDescriptor::layout
DataLayout layout
Data layout.
Definition: TensorDescriptor.h:111
FusedConvolutionBatchNormalizationWithPostOpsFunction.h
arm_compute::graph::backends::detail::create_fused_convolution_with_post_op
std::unique_ptr< IFunction > create_fused_convolution_with_post_op(FusedConvolutionWithPostOpNode &node, GraphContext &ctx)
Create a backend convolution layer function with post operator.
Definition: FunctionHelpers.h:555
arm_compute::graph::FusedConvolutionWithPostOpNode
Convolution node.
Definition: FusedConvolutionWithPostOpNode.h:36
arm_compute::graph::backends::detail::create_pooling_layer
std::unique_ptr< IFunction > create_pooling_layer(PoolingLayerNode &node)
Create a backend pooling layer function.
Definition: FunctionHelpers.h:1486
arm_compute::DataLayout
DataLayout
[DataLayout enum definition]
Definition: Types.h:113
arm_compute::graph::backends::detail::create_concatenate_layer
std::unique_ptr< arm_compute::IFunction > create_concatenate_layer(ConcatenateLayerNode &node)
Create a backend layer concatenate function.
Definition: FunctionHelpers.h:407
arm_compute::StridedSliceLayerInfo::begin_mask
int32_t begin_mask() const
Definition: Types.h:1941
arm_compute::experimental::PostOpList
A sequence of PostOps that can be appended to the end of other operators.
Definition: IPostOp.h:119
arm_compute::ConvertPolicy
ConvertPolicy
Policy to handle integer overflow.
Definition: Types.h:404
arm_compute::graph::backends::detail::create_reshape_layer
std::unique_ptr< IFunction > create_reshape_layer(ReshapeLayerNode &node)
Create a backend reshape layer function.
Definition: FunctionHelpers.h:1752
arm_compute::graph::get_dimension_idx
size_t get_dimension_idx(DataLayout data_layout, const DataLayoutDimension data_layout_dimension)
Get index of a tensor&#39;s given dimension depending on its layout.
Definition: Utils.cpp:148
arm_compute::graph::backends::detail::create_fused_convolution_batch_normalization_with_post_op
std::unique_ptr< IFunction > create_fused_convolution_batch_normalization_with_post_op(FusedConvolutionBatchNormalizationWithPostOpsNode &node, GraphContext &ctx)
Create a backend convolution batch normalization layer function with post operator.
Definition: FunctionHelpers.h:649
arm_compute::graph::backends::detail::create_reorg_layer
std::unique_ptr< IFunction > create_reorg_layer(ReorgLayerNode &node)
Create a backend reorg layer function.
Definition: FunctionHelpers.h:1715
arm_compute::graph::Tensor
Tensor object.
Definition: Tensor.h:41
arm_compute::graph::backends::detail::create_channel_shuffle_layer
std::unique_ptr< IFunction > create_channel_shuffle_layer(ChannelShuffleLayerNode &node)
Create a backend channel shuffle layer function.
Definition: FunctionHelpers.h:372
arm_compute::quantization::epsilon
constexpr float epsilon
Definition: AsymmHelpers.cpp:37
arm_compute::graph::UnaryEltwiseOperation
UnaryEltwiseOperation
Supported Unary Element-wise operations.
Definition: Types.h:120