Compute Library
 21.11
CpuFullyConnected.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 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  */
25 
32 #include "src/common/utils/Log.h"
41 
42 namespace arm_compute
43 {
44 namespace cpu
45 {
46 using namespace arm_compute::experimental;
48 
49 namespace
50 {
51 // Get min, max bound of a quantized asymmetric dst tensor, with the effect of fused activation
52 std::pair<PixelValue, PixelValue> get_quantized_asymmetric_output_min_max(const QuantizationInfo &q_info, const ActivationLayerInfo &act_info, DataType data_type)
53 {
54  PixelValue type_min{};
55  PixelValue type_max{};
56  std::tie(type_min, type_max) = get_min_max(data_type);
57  const UniformQuantizationInfo q_unif = q_info.uniform();
58 
59  if(act_info.enabled())
60  {
61  switch(act_info.activation())
62  {
64  type_min = PixelValue(q_unif.offset);
65  break;
67  type_min = PixelValue(q_unif.offset);
68  type_max = PixelValue(act_info.a(), data_type, q_info);
69  break;
71  type_min = PixelValue(act_info.b(), data_type, q_info);
72  type_max = PixelValue(act_info.a(), data_type, q_info);
73  break;
74  default:
75  ARM_COMPUTE_ERROR("Activation function not supported.");
76  break;
77  }
78  }
79 
80  return std::make_pair(type_min, type_max);
81 }
82 
83 Status get_gemmlowp_output_stage_info(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const ActivationLayerInfo &act,
84  GEMMLowpOutputStageInfo &gemmlowp_output_stage_info)
85 {
86  const auto data_type = src->data_type();
87  const QuantizationInfo oq_info = dst->quantization_info();
88  const UniformQuantizationInfo iq_unif = src->quantization_info().uniform();
89  const UniformQuantizationInfo wq_unif = weights->quantization_info().uniform();
90  const UniformQuantizationInfo oq_unif = oq_info.uniform();
91 
92  float multiplier = (iq_unif.scale * wq_unif.scale) / oq_unif.scale;
93  int32_t output_multiplier;
94  int32_t output_shift;
95 
96  ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
97 
98  PixelValue type_min{};
99  PixelValue type_max{};
100  std::tie(type_min, type_max) = get_quantized_asymmetric_output_min_max(oq_info, act, data_type);
101 
102  gemmlowp_output_stage_info.gemmlowp_multiplier = output_multiplier;
103  gemmlowp_output_stage_info.gemmlowp_shift = output_shift;
104  gemmlowp_output_stage_info.gemmlowp_offset = oq_unif.offset;
105  gemmlowp_output_stage_info.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
106  gemmlowp_output_stage_info.gemmlowp_min_bound = type_min.get<int32_t>();
107  gemmlowp_output_stage_info.gemmlowp_max_bound = type_max.get<int32_t>();
108 
109  return Status{};
110 }
111 
112 Status validate_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, const ActivationLayerInfo &act, bool enable_fast_math)
113 {
114  if(is_data_type_quantized_asymmetric(src->data_type()))
115  {
116  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
117  // Extract and negate src and weights offset
118  const QuantizationInfo src_quantization_info(src->quantization_info().uniform().scale, -src->quantization_info().uniform().offset);
119  const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset);
120 
121  GEMMLowpOutputStageInfo gemmlowp_output_stage_info;
122  ARM_COMPUTE_RETURN_ON_ERROR(get_gemmlowp_output_stage_info(src, weights, dst, act, gemmlowp_output_stage_info));
123 
124  GEMMInfo gemm_info;
125  gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info);
126  gemm_info.set_fast_math(enable_fast_math);
127 
128  // Validate gemmlowp function
129  TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info);
130  TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info);
132  &weights_info,
133  biases,
134  dst,
135  gemm_info));
136  }
137  else
138  {
139  GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */);
140  gemm_info.set_fast_math(enable_fast_math);
141  ARM_COMPUTE_RETURN_ON_ERROR(CpuGemm::validate(src, weights, biases, dst, 1.f, 1.0f, gemm_info));
142  }
143 
144  return Status{};
145 }
146 } // namespace
147 
149  : _flatten(nullptr),
150  _convert_weights(nullptr),
151  _transpose_weights(nullptr),
152  _mm_gemm(nullptr),
153  _mm_gemmlowp(nullptr),
154  _flattened_src(),
155  _converted_weights(),
156  _reshaped_weights(),
157  _trans_weights(),
158  _trans_weights_idx(AuxTensorIdx::Count),
159  _aux_mem(Count),
160  _needs_weights_conversion(false),
161  _needs_weights_reshape(false),
162  _is_fc_after_conv(false),
163  _is_quantized_asymmetric(false),
164  _is_prepared(false),
165  _enable_fast_math(false)
166 
167 {
168 }
169 
171 
172 void CpuFullyConnected::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act)
173 {
174  if(_is_quantized_asymmetric)
175  {
176  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
177  // Extract and negate src and weights offset
178  const QuantizationInfo src_quantization_info(src->quantization_info().uniform().scale, -src->quantization_info().uniform().offset);
179  const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset);
180 
181  TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info);
182  TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info);
183 
184  // Configure gemmlowp function and output stage for asymmetric quantized types
185  GEMMLowpOutputStageInfo gemmlowp_output_stage_info;
186  const Status status = get_gemmlowp_output_stage_info(&src_info, &weights_info, dst, act, gemmlowp_output_stage_info);
188 
190  gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info);
191  gemm_info.set_activation_info(act);
192  gemm_info.set_fast_math(_enable_fast_math);
193  _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>();
194  _mm_gemmlowp->configure(&src_info, &weights_info, biases, dst, gemm_info);
195  }
196  else
197  {
198  // Configure matrix multiply kernel
199  GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */);
200  gemm_info.set_activation_info(act);
201  gemm_info.set_fast_math(_enable_fast_math);
202  _mm_gemm = std::make_unique<CpuGemm>();
203  _mm_gemm->configure(src, weights, biases, dst, 1.f, 1.0f, gemm_info);
204  }
205 }
206 
207 void CpuFullyConnected::configure_conv_fc(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act)
208 {
209  ARM_COMPUTE_ERROR_ON((weights->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
210 
211  // If the fully connected layer is called after a convolution layer, the src tensor must be linearized
212 
213  // Initialize output tensor for flatten
214  auto_init_if_empty(_flattened_src, src->clone()->set_tensor_shape(compute_flatten_shape(src)));
215 
216  _flatten = std::make_unique<CpuFlatten>();
217  _flatten->configure(src, &_flattened_src);
218 
219  // Configure matrix multiply kernel
220  configure_mm(&_flattened_src, weights, biases, dst, act);
221 }
222 
223 void CpuFullyConnected::configure_fc_fc(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act)
224 {
225  ARM_COMPUTE_ERROR_ON(src->dimension(0) != weights->dimension(1));
226 
227  // Configure matrix multiply kernel
228  configure_mm(src, weights, biases, dst, act);
229 }
230 
231 void CpuFullyConnected::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst,
232  FullyConnectedLayerInfo fc_info)
233 {
234  // Perform validate step
235  ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
237  weights,
238  biases != nullptr ? biases : nullptr,
239  dst,
240  fc_info));
241  ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, fc_info);
242 
243  _needs_weights_conversion = false;
244  _needs_weights_reshape = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
245  _needs_weights_reshape = _needs_weights_reshape && !fc_info.retain_internal_weights;
246  _is_fc_after_conv = true;
247  _is_quantized_asymmetric = is_data_type_quantized_asymmetric(src->data_type());
248  _is_prepared = false;
249  _trans_weights_idx = AuxTensorIdx::Count;
250  _enable_fast_math = fc_info.enable_fast_math;
251 
252  // With the Fully Connected layer we can have 4 different cases:
253  // 1) Convolution layer -> Fully Connected layer without batches
254  // 2) Fully Connected layer -> Fully Connected layer without batches
255  // 3) Convolution layer -> Fully Connected layer with batches
256  // 4) Fully Connected layer -> Fully Connected layer with batches
257 
258  const ITensorInfo *weights_to_use = weights;
259 
260  // Check if we have a fully connected layer with batches
261  const bool is_batched_fc_layer = dst->dimension(1) > 1;
262  if(is_batched_fc_layer)
263  {
264  _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3,
265  src->tensor_shape().cend(),
266  dst->tensor_shape().cbegin() + 1));
267  }
268  else
269  {
270  _is_fc_after_conv = src->num_dimensions() > 1;
271  }
272 
273  // Reshape weights if needed
274  if(_needs_weights_reshape)
275  {
276  // Reshape the weights
277  _transpose_weights = std::make_unique<kernels::CpuTransposeKernel>();
278  _transpose_weights->configure(weights, &_reshaped_weights);
279  weights_to_use = &_reshaped_weights;
280  _trans_weights_idx = AuxTensorIdx::TransposedWeights;
281  }
282 
283  // Convert weights if needed
284  if(_is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout))
285  {
286  // Convert weights
287  _convert_weights = std::make_unique<CpuConvertFullyConnectedWeights>();
288  _convert_weights->configure(weights_to_use,
289  &_converted_weights,
290  src->tensor_shape(),
291  fc_info.weights_trained_layout);
292 
293  weights_to_use = &_converted_weights;
294  _needs_weights_conversion = true;
295  _trans_weights_idx = AuxTensorIdx::ConvertedWeights;
296  }
297 
298  if(_is_fc_after_conv)
299  {
300  // Fully Connected layer after a Convolution Layer without batches
301  configure_conv_fc(src, weights_to_use, biases, dst, fc_info.activation_info);
302  }
303  else
304  {
305  // Fully Connected layer after a Fully Connected Layer without batches
306  configure_fc_fc(src, weights_to_use, biases, dst, fc_info.activation_info);
307  }
308 
309  // Retain the tensorinfo with the weights to use
310  if(_needs_weights_reshape || _needs_weights_conversion)
311  {
312  _trans_weights = *weights_to_use;
313  }
314 
315  // Set auxiliary memory requirements
316  auto gemm_mem_req = (_is_quantized_asymmetric) ? _mm_gemmlowp->workspace() : _mm_gemm->workspace();
317  for(unsigned int i = 0; i < gemm_mem_req.size(); ++i)
318  {
319  _aux_mem[i] = gemm_mem_req[i];
320  }
321 
322  if(_aux_mem[Pretranspose].size > 0)
323  {
324  // Release permuted weights at the end of prepare as they are further transposed by the assembly dispatch
325  // Do not release them if biases are dynamic and data type is quantized, since the weights tensor will be used for biases offset calculation
326  _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), (_is_quantized_asymmetric
327  && biases && !(biases->are_values_constant())) ?
328  MemoryLifetime::Persistent :
329  MemoryLifetime::Prepare,
330  _reshaped_weights.total_size());
331  _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Prepare, _converted_weights.total_size());
332  }
333  else
334  {
335  _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), _needs_weights_conversion ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _reshaped_weights.total_size());
336  _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Persistent, _converted_weights.total_size());
337  }
338  _aux_mem[FlattenedSrc] = MemoryInfo(offset_int_vec(FlattenedSrc), MemoryLifetime::Temporary, _flattened_src.total_size());
339 }
340 
341 Status CpuFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
342  FullyConnectedLayerInfo fc_info)
343 {
345  ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst);
352 
353  bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
354  bool is_fc_after_conv = true;
355 
356  const ITensorInfo &flatten_src = TensorInfo(src->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(src)));
357  const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
358  const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
359 
360  // With the Fully Connected layer we can have 4 different cases:
361  // 1) Convolution layer -> Fully Connected layer without batches
362  // 2) Fully Connected layer -> Fully Connected layer without batches
363  // 3) Convolution layer -> Fully Connected layer with batches
364  // 4) Fully Connected layer -> Fully Connected layer with batches
365 
366  const ITensorInfo *src_to_use = src;
367  const ITensorInfo *weights_to_use = weights;
368 
369  // Check if we have a fully connected layer with batches
370  const bool is_batched_fc_layer = dst->dimension(1) > 1;
371 
372  if(biases != nullptr)
373  {
376  {
378  }
379  else
380  {
382  }
383  }
384 
385  if(is_batched_fc_layer)
386  {
387  is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3,
388  src->tensor_shape().cend(),
389  dst->tensor_shape().cbegin() + 1));
390  }
391  else
392  {
393  is_fc_after_conv = src->num_dimensions() > 1;
394  }
395 
396  if(!weights_reshaped)
397  {
398  // Validate reshape weights kernel
400  weights_to_use = &reshaped_weights;
401  }
402 
403  if(is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout))
404  {
405  // Validate convert weights kernel
407  &converted_weights,
408  src->tensor_shape(),
409  fc_info.weights_trained_layout));
410  weights_to_use = &converted_weights;
411  }
412 
413  if(is_fc_after_conv)
414  {
415  // Fully Connected layer after a Convolution Layer without batches
416  ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
417 
418  // Validate flatten kernel
420  src_to_use = &flatten_src;
421  }
422  else
423  {
424  // Fully Connected layer after a Fully Connected Layer without batches
425  ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension(1));
426  }
427  // Validate matrix multiply kernel
428  ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(src_to_use, weights_to_use, biases, dst, fc_info.activation_info, fc_info.enable_fast_math));
429 
430  return Status{};
431 }
432 
434 {
435  prepare(tensors);
436 
437  auto src = tensors.get_const_tensor(ACL_SRC_0);
438 
439  CpuAuxTensorHandler flattened_src(offset_int_vec(FlattenedSrc), _flattened_src, tensors, false);
440  CpuAuxTensorHandler transformed_wei(offset_int_vec(_trans_weights_idx), _trans_weights, tensors, false);
441 
442  // Linearize src if it comes from a convolutional layer
443  if(_is_fc_after_conv)
444  {
445  ITensorPack flatten_pack{ { ACL_SRC, src }, { ACL_DST, flattened_src.get() } };
446  _flatten->run(flatten_pack);
447  }
448 
449  ITensorPack gemm_pack = tensors;
450  gemm_pack.add_const_tensor(ACL_SRC_0, (_is_fc_after_conv) ? flattened_src.get() : src);
451  if(_needs_weights_reshape || _needs_weights_conversion)
452  {
453  gemm_pack.add_const_tensor(ACL_SRC_1, transformed_wei.get());
454  }
455 
456  // Run matrix multiply
457  if(_is_quantized_asymmetric)
458  {
459  _mm_gemmlowp->run(gemm_pack);
460  }
461  else
462  {
463  _mm_gemm->run(gemm_pack);
464  }
465 }
466 
468 {
469  if(!_is_prepared)
470  {
471  auto weights = tensors.get_const_tensor(ACL_SRC_1);
472 
473  CpuAuxTensorHandler reshaped_weights(offset_int_vec(TransposedWeights), _reshaped_weights, tensors, false);
474  CpuAuxTensorHandler converted_weights(offset_int_vec(ConvertedWeights), _converted_weights, tensors, false);
475 
476  // Pointer to current weights
477  const ITensor *cur_weights = weights;
478 
479  // Reshape of the weights (happens only once)
480  if(_needs_weights_reshape)
481  {
482  // Run reshape weights kernel and mark weights as unused
483  ITensorPack transpose_pack{ { ACL_SRC, weights }, { ACL_DST, reshaped_weights.get() } };
484  NEScheduler::get().schedule_op(_transpose_weights.get(), Window::DimY, _transpose_weights->window(), transpose_pack);
485 
486  cur_weights->mark_as_unused();
487  cur_weights = reshaped_weights.get();
488  }
489 
490  // Convert weights if needed (happens only once)
491  if(_needs_weights_conversion)
492  {
493  ITensorPack convert_pack{ { ACL_SRC, cur_weights }, { ACL_DST, converted_weights.get() } };
494  _convert_weights->run(convert_pack);
495 
496  cur_weights->mark_as_unused();
497  cur_weights = converted_weights.get();
498  }
499 
500  ITensorPack gemm_pack = tensors;
501  gemm_pack.add_const_tensor(ACL_SRC_1, cur_weights);
502 
503  // Prepare GEMM prepare and release unused weights
504  if(!_is_quantized_asymmetric)
505  {
506  _mm_gemm->prepare(gemm_pack);
507  }
508  else
509  {
510  _mm_gemmlowp->prepare(gemm_pack);
511  }
512 
513  _is_prepared = true;
514  }
515 }
516 
518 {
519  return _aux_mem;
520 }
521 } // namespace cpu
522 } // namespace arm_compute
bool is_data_type_quantized(DataType dt)
Check if a given data type is of quantized type.
Definition: Utils.h:981
virtual size_t num_dimensions() const =0
The number of dimensions of the tensor (rank)
void prepare(ITensorPack &tensors) override
Prepare the function for executing.
Quantize using a fixed point multiplication.
void set_activation_info(const ActivationLayerInfo &activation_info)
Set activation layer info.
Definition: Types.h:2155
std::unique_ptr< ITensorInfo > clone() const override
Provide a clone of the current object of class T.
Definition: TensorInfo.cpp:282
void run(ITensorPack &tensors) override
Run the kernels contained in the function.
bool enabled() const
Check if initialised.
Definition: Types.h:1559
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
void add_const_tensor(int id, const ITensor *tensor)
Add const tensor to the pack.
Definition: ITensorPack.cpp:49
bool retain_internal_weights
Retain internal reshaped weights.
Definition: Types.h:1580
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
virtual void schedule_op(ICPPKernel *kernel, const Hints &hints, const Window &window, ITensorPack &tensors)=0
Runs the kernel in the same thread as the caller synchronously.
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of CpuGemm.
Definition: CpuGemm.cpp:152
1 channel, 1 F32 per channel
#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
Fully connected layer info.
Definition: Types.h:1572
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
Status class.
Definition: Error.h:52
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Activation Layer Information class.
Definition: Types.h:1509
Interface for CPU tensor.
Definition: ITensor.h:36
bool enable_fast_math
Enable fast math computation.
Definition: Types.h:1581
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2021 Arm Limited.
std::vector< MemoryInfo > MemoryRequirements
Definition: Types.h:132
1 channel, 1 F16 per channel
static Status validate(const ITensorInfo *src, const ITensorInfo *dst)
Static function to check if given info will lead to a valid configuration.
Definition: CpuFlatten.cpp:42
TensorShape compute_transposed_shape(const ITensorInfo &input)
Calculate the transposed shape of a tensor.
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:159
void mark_as_unused() const
Marks a tensor as unused.
Definition: ITensor.cpp:168
1 channel, 1 S32 per channel
TensorShape compute_flatten_shape(const ITensorInfo *input)
Calculate the flattened output shape of a tensor.
const DataType data_type
Definition: Im2Col.cpp:150
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
Definition: ITensorPack.cpp:54
Quantization information.
static Status validate(const ITensorInfo *src, const ITensorInfo *dst, const TensorShape &original_src_shape, DataLayout data_layout)
Static function to check if given info will lead to a valid configuration.
experimental::MemoryRequirements workspace() const override
Return the memory requirements required by the workspace.
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
static Status validate(const ITensorInfo *src, const ITensorInfo *dst)
Static function to check if given info will lead to a valid configuration.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
quantized, asymmetric fixed-point 8-bit number unsigned
bool are_weights_reshaped
Reshape the weights tensor if false.
Definition: Types.h:1579
size_t total_size() const override
Returns the total size of the tensor in bytes.
Definition: TensorInfo.h:250
void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Set the input and output tensors.
UniformQuantizationInfo uniform() const
Return per layer quantization info.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
GEMMLowp output stage info.
Definition: Types.h:1922
virtual bool are_values_constant() const =0
Flag indicating whether the values of the tensor are constant, meaning that they can change on kernel...
ActivationLayerInfo activation_info
Fused activation to apply after the matrix multiplication.
Definition: Types.h:1575
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1003
TensorInfo src_info(src_shape, 1, data_type)
std::array< T, num_max_dimensions >::const_iterator cend() const
Returns a read-only (constant) iterator that points one past the last element in the dimension array...
Definition: Dimensions.h:255
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
std::array< T, num_max_dimensions >::const_iterator cbegin() const
Returns a read-only (constant) iterator that points to the first element in the dimension array...
Definition: Dimensions.h:231
static Status validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Static function to check if given info will lead to a valid configuration of CpuFullyConnected.
DataLayout weights_trained_layout
Layout that the weights have been trained with.
Definition: Types.h:1577
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
bool transpose_weights
Transpose weights if true.
Definition: Types.h:1578
Tensor packing service.
Definition: ITensorPack.h:39
#define ARM_COMPUTE_LOG_PARAMS(...)
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:157
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:43
int offset_int_vec(int offset)
Definition: MemoryHelpers.h:38
GEMM information class.
Definition: Types.h:1974
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1544
quantized, asymmetric fixed-point 8-bit number signed
static constexpr size_t num_max_dimensions
Number of dimensions the tensor has.
Definition: Dimensions.h:46
DataType
Available data types.
Definition: Types.h:79
std::tuple< PixelValue, PixelValue > get_min_max(DataType dt)
Compute the mininum and maximum values a data type can take.
Definition: Utils.h:564
ErrorCode error_code() const
Gets error code.
Definition: Error.h:89
virtual DataLayout data_layout() const =0
Get the data layout of the tensor.
void set_fast_math(bool fast_math)
Set fast math flag.
Definition: Types.h:2115
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:94
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *dst, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration.