Compute Library
 22.11
CpuFullyConnected.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2021-2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
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  _fixed_format(false),
167  _weight_format(arm_compute::WeightFormat::UNSPECIFIED)
168 {
169 }
170 
172 
173 void CpuFullyConnected::configure_mm(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act)
174 {
175  if(_is_quantized_asymmetric)
176  {
177  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
178  // Extract and negate src and weights offset
179  const QuantizationInfo src_quantization_info(src->quantization_info().uniform().scale, -src->quantization_info().uniform().offset);
180  const QuantizationInfo weights_quantization_info(weights->quantization_info().uniform().scale, -weights->quantization_info().uniform().offset);
181 
182  TensorInfo src_info = src->clone()->set_quantization_info(src_quantization_info);
183  TensorInfo weights_info = weights->clone()->set_quantization_info(weights_quantization_info);
184 
185  // Configure gemmlowp function and output stage for asymmetric quantized types
186  GEMMLowpOutputStageInfo gemmlowp_output_stage_info;
187  const Status status = get_gemmlowp_output_stage_info(&src_info, &weights_info, dst, act, gemmlowp_output_stage_info);
189 
191  gemm_info.set_gemmlowp_output_stage(gemmlowp_output_stage_info);
192  gemm_info.set_activation_info(act);
193  gemm_info.set_fast_math(_enable_fast_math);
194  _mm_gemmlowp = std::make_unique<CpuGemmLowpMatrixMultiplyCore>();
195  _mm_gemmlowp->configure(&src_info, &weights_info, biases, dst, gemm_info);
196  }
197  else
198  {
199  // Configure matrix multiply kernel
200  GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */);
201  gemm_info.set_activation_info(act);
202  gemm_info.set_fast_math(_enable_fast_math);
203  gemm_info.set_fixed_format(_fixed_format);
204  gemm_info.set_weight_format(_weight_format);
205  _mm_gemm = std::make_unique<CpuGemm>();
206  _mm_gemm->configure(src, weights, biases, dst, 1.f, 1.0f, gemm_info);
207  }
208 }
209 
210 void CpuFullyConnected::configure_conv_fc(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act)
211 {
212  ARM_COMPUTE_ERROR_ON((weights->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
213 
214  // If the fully connected layer is called after a convolution layer, the src tensor must be linearized
215 
216  // Initialize output tensor for flatten
217  auto_init_if_empty(_flattened_src, src->clone()->set_tensor_shape(compute_flatten_shape(src)));
218 
219  _flatten = std::make_unique<CpuFlatten>();
220  _flatten->configure(src, &_flattened_src);
221 
222  // Configure matrix multiply kernel
223  configure_mm(&_flattened_src, weights, biases, dst, act);
224 }
225 
226 void CpuFullyConnected::configure_fc_fc(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, const ActivationLayerInfo &act)
227 {
228  ARM_COMPUTE_ERROR_ON(src->dimension(0) != weights->dimension(1));
229 
230  // Configure matrix multiply kernel
231  configure_mm(src, weights, biases, dst, act);
232 }
233 
234 void CpuFullyConnected::configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst,
235  FullyConnectedLayerInfo fc_info, const WeightsInfo &weights_info)
236 {
237  // Perform validate step
238  ARM_COMPUTE_ERROR_ON_NULLPTR(src, weights, dst);
240  weights,
241  biases != nullptr ? biases : nullptr,
242  dst,
243  fc_info));
244  ARM_COMPUTE_LOG_PARAMS(src, weights, biases, dst, fc_info);
245 
246  _needs_weights_conversion = false;
247  _needs_weights_reshape = fc_info.transpose_weights ? !fc_info.are_weights_reshaped : false;
248  _needs_weights_reshape = _needs_weights_reshape && !fc_info.retain_internal_weights;
249  _is_fc_after_conv = true;
250  _is_quantized_asymmetric = is_data_type_quantized_asymmetric(src->data_type());
251  _is_prepared = false;
252  _trans_weights_idx = AuxTensorIdx::Count;
253  _enable_fast_math = fc_info.enable_fast_math;
254  _fixed_format = weights_info.weight_format() != WeightFormat::UNSPECIFIED;
255  _weight_format = weights_info.weight_format();
256 
257  // With the Fully Connected layer we can have 4 different cases:
258  // 1) Convolution layer -> Fully Connected layer without batches
259  // 2) Fully Connected layer -> Fully Connected layer without batches
260  // 3) Convolution layer -> Fully Connected layer with batches
261  // 4) Fully Connected layer -> Fully Connected layer with batches
262 
263  const ITensorInfo *weights_to_use = weights;
264 
265  // Check if we have a fully connected layer with batches
266  const bool is_batched_fc_layer = dst->dimension(1) > 1;
267  if(is_batched_fc_layer)
268  {
269  _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3, src->tensor_shape().cend(), dst->tensor_shape().cbegin() + 1));
270  }
271  else
272  {
273  _is_fc_after_conv = src->num_dimensions() > 1;
274  }
275 
276  // Reshape weights if needed
277  if(_needs_weights_reshape)
278  {
279  // Reshape the weights
280  _transpose_weights = std::make_unique<kernels::CpuTransposeKernel>();
281  _transpose_weights->configure(weights, &_reshaped_weights);
282  weights_to_use = &_reshaped_weights;
283  _trans_weights_idx = AuxTensorIdx::TransposedWeights;
284  }
285 
286  // Convert weights if needed
287  if(_is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout))
288  {
289  // Convert weights
290  _convert_weights = std::make_unique<CpuConvertFullyConnectedWeights>();
291  _convert_weights->configure(weights_to_use,
292  &_converted_weights,
293  src->tensor_shape(),
294  fc_info.weights_trained_layout);
295 
296  weights_to_use = &_converted_weights;
297  _needs_weights_conversion = true;
298  _trans_weights_idx = AuxTensorIdx::ConvertedWeights;
299  }
300 
301  if(_is_fc_after_conv)
302  {
303  // Fully Connected layer after a Convolution Layer without batches
304  configure_conv_fc(src, weights_to_use, biases, dst, fc_info.activation_info);
305  }
306  else
307  {
308  // Fully Connected layer after a Fully Connected Layer without batches
309  configure_fc_fc(src, weights_to_use, biases, dst, fc_info.activation_info);
310  }
311 
312  // Retain the tensorinfo with the weights to use
313  if(_needs_weights_reshape || _needs_weights_conversion)
314  {
315  _trans_weights = *weights_to_use;
316  }
317 
318  // Set auxiliary memory requirements
319  auto gemm_mem_req = (_is_quantized_asymmetric) ? _mm_gemmlowp->workspace() : _mm_gemm->workspace();
320  for(unsigned int i = 0; i < gemm_mem_req.size(); ++i)
321  {
322  _aux_mem[i] = gemm_mem_req[i];
323  }
324 
325  if(_aux_mem[Pretranspose].size > 0)
326  {
327  // Release permuted weights at the end of prepare as they are further transposed by the assembly dispatch
328  // Do not release them if biases are dynamic and data type is quantized, since the weights tensor will be used for biases offset calculation
329  _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), (_is_quantized_asymmetric && biases
330  && !(biases->are_values_constant())) ? MemoryLifetime::Persistent : MemoryLifetime::Prepare,
331  _reshaped_weights.total_size());
332  _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Prepare, _converted_weights.total_size());
333  }
334  else
335  {
336  _aux_mem[TransposedWeights] = MemoryInfo(offset_int_vec(TransposedWeights), _needs_weights_conversion ? MemoryLifetime::Prepare : MemoryLifetime::Persistent, _reshaped_weights.total_size());
337  _aux_mem[ConvertedWeights] = MemoryInfo(offset_int_vec(ConvertedWeights), MemoryLifetime::Persistent, _converted_weights.total_size());
338  }
339  _aux_mem[FlattenedSrc] = MemoryInfo(offset_int_vec(FlattenedSrc), MemoryLifetime::Temporary, _flattened_src.total_size());
340 }
341 
342 Status CpuFullyConnected::has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *src, const ITensorInfo *weights,
343  const ITensorInfo *biases, const ITensorInfo *dst, FullyConnectedLayerInfo fc_info, WeightsInfo weights_info)
344 {
345  GEMMInfo gemm_info(false, false, true /* Reshape weights only for the first run */);
346  gemm_info.set_activation_info(fc_info.activation_info);
347  gemm_info.set_fast_math(fc_info.enable_fast_math);
348  gemm_info.set_fixed_format(weights_info.weight_format() != WeightFormat::UNSPECIFIED);
349  gemm_info.set_weight_format(weights_info.weight_format());
350 
351  return CpuGemm::has_opt_impl(expected_weight_format, src, weights, biases, dst, gemm_info);
352 }
353 
354 Status CpuFullyConnected::validate(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst,
355  FullyConnectedLayerInfo fc_info)
356 {
358  ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(src, weights, dst);
365 
366  bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
367  bool is_fc_after_conv = true;
368 
369  const ITensorInfo &flatten_src = TensorInfo(src->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(src)));
370  const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
371  const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
372 
373  // With the Fully Connected layer we can have 4 different cases:
374  // 1) Convolution layer -> Fully Connected layer without batches
375  // 2) Fully Connected layer -> Fully Connected layer without batches
376  // 3) Convolution layer -> Fully Connected layer with batches
377  // 4) Fully Connected layer -> Fully Connected layer with batches
378 
379  const ITensorInfo *src_to_use = src;
380  const ITensorInfo *weights_to_use = weights;
381 
382  // Check if we have a fully connected layer with batches
383  const bool is_batched_fc_layer = dst->dimension(1) > 1;
384 
385  if(biases != nullptr)
386  {
389  {
391  }
392  else
393  {
395  }
396  }
397 
398  if(is_batched_fc_layer)
399  {
400  is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(src->tensor_shape().cbegin() + 3, src->tensor_shape().cend(), dst->tensor_shape().cbegin() + 1));
401  }
402  else
403  {
404  is_fc_after_conv = src->num_dimensions() > 1;
405  }
406 
407  if(!weights_reshaped)
408  {
409  // Validate reshape weights kernel
411  weights_to_use = &reshaped_weights;
412  }
413 
414  if(is_fc_after_conv && (src->data_layout() != fc_info.weights_trained_layout))
415  {
416  // Validate convert weights kernel
418  &converted_weights,
419  src->tensor_shape(),
420  fc_info.weights_trained_layout));
421  weights_to_use = &converted_weights;
422  }
423 
424  if(is_fc_after_conv)
425  {
426  // Fully Connected layer after a Convolution Layer without batches
427  ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (src->dimension(0) * src->dimension(1) * src->dimension(2))));
428 
429  // Validate flatten kernel
431  src_to_use = &flatten_src;
432  }
433  else
434  {
435  // Fully Connected layer after a Fully Connected Layer without batches
436  ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != weights_to_use->dimension(1));
437  }
438  // Validate matrix multiply kernel
439  ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(src_to_use, weights_to_use, biases, dst, fc_info.activation_info, fc_info.enable_fast_math));
440 
441  return Status{};
442 }
443 
445 {
446  prepare(tensors);
447 
448  auto src = tensors.get_const_tensor(ACL_SRC_0);
449 
450  CpuAuxTensorHandler flattened_src(offset_int_vec(FlattenedSrc), _flattened_src, tensors, false);
451  CpuAuxTensorHandler transformed_wei(offset_int_vec(_trans_weights_idx), _trans_weights, tensors, false);
452 
453  // Linearize src if it comes from a convolutional layer
454  if(_is_fc_after_conv)
455  {
456  ITensorPack flatten_pack{ { ACL_SRC, src }, { ACL_DST, flattened_src.get() } };
457  _flatten->run(flatten_pack);
458  }
459 
460  ITensorPack gemm_pack = tensors;
461  gemm_pack.add_const_tensor(ACL_SRC_0, (_is_fc_after_conv) ? flattened_src.get() : src);
462  if(_needs_weights_reshape || _needs_weights_conversion)
463  {
464  gemm_pack.add_const_tensor(ACL_SRC_1, transformed_wei.get());
465  }
466 
467  // Run matrix multiply
468  if(_is_quantized_asymmetric)
469  {
470  _mm_gemmlowp->run(gemm_pack);
471  }
472  else
473  {
474  _mm_gemm->run(gemm_pack);
475  }
476 }
477 
479 {
480  if(!_is_prepared)
481  {
482  auto weights = tensors.get_const_tensor(ACL_SRC_1);
483 
484  CpuAuxTensorHandler reshaped_weights(offset_int_vec(TransposedWeights), _reshaped_weights, tensors, false);
485  CpuAuxTensorHandler converted_weights(offset_int_vec(ConvertedWeights), _converted_weights, tensors, false);
486 
487  // Pointer to current weights
488  const ITensor *cur_weights = weights;
489 
490  // Reshape of the weights (happens only once)
491  if(_needs_weights_reshape)
492  {
493  // Run reshape weights kernel and mark weights as unused
494  ITensorPack transpose_pack{ { ACL_SRC, weights }, { ACL_DST, reshaped_weights.get() } };
495  NEScheduler::get().schedule_op(_transpose_weights.get(), Window::DimY, _transpose_weights->window(), transpose_pack);
496 
497  cur_weights->mark_as_unused();
498  cur_weights = reshaped_weights.get();
499  }
500 
501  // Convert weights if needed (happens only once)
502  if(_needs_weights_conversion)
503  {
504  ITensorPack convert_pack{ { ACL_SRC, cur_weights }, { ACL_DST, converted_weights.get() } };
505  _convert_weights->run(convert_pack);
506 
507  cur_weights->mark_as_unused();
508  cur_weights = converted_weights.get();
509  }
510 
511  ITensorPack gemm_pack = tensors;
512  gemm_pack.add_const_tensor(ACL_SRC_1, cur_weights);
513 
514  // Prepare GEMM prepare and release unused weights
515  if(!_is_quantized_asymmetric)
516  {
517  _mm_gemm->prepare(gemm_pack);
518  }
519  else
520  {
521  _mm_gemmlowp->prepare(gemm_pack);
522  }
523 
524  _is_prepared = true;
525  }
526 }
527 
529 {
530  return _aux_mem;
531 }
532 } // namespace cpu
533 } // namespace arm_compute
bool is_data_type_quantized(DataType dt)
Check if a given data type is of quantized type.
Definition: Utils.h:1030
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_fixed_format(bool fixed_format)
Set fixed-format flag.
Definition: Types.h:2578
void set_activation_info(const ActivationLayerInfo &activation_info)
Set activation layer info.
Definition: Types.h:2545
std::unique_ptr< ITensorInfo > clone() const override
Provide a clone of the current object of class T.
Definition: TensorInfo.cpp:302
void run(ITensorPack &tensors) override
Run the kernels contained in the function.
static Status has_opt_impl(arm_compute::WeightFormat &weight_format, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *d, const GEMMInfo &gemm_info=GEMMInfo())
Indicates whether or not there is an optimal assembly implementation that can be used to process the ...
Definition: CpuGemm.cpp:371
bool enabled() const
Check if initialised.
Definition: Types.h:1694
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:1817
#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 has_opt_impl(arm_compute::WeightFormat &expected_weight_format, const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *dst, FullyConnectedLayerInfo fc_info, WeightsInfo weights_info)
Static function that queries whether there exists fixed-format kernel and if it exists it will return...
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:153
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:1809
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
WeightFormat
Memory layouts for the weights tensor.
Definition: Types.h:2015
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:1639
void configure(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *biases, ITensorInfo *dst, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo(), const WeightsInfo &weights_info=WeightsInfo())
Set the input and output tensors.
Interface for CPU tensor.
Definition: ITensor.h:36
bool enable_fast_math
Enable fast math computation.
Definition: Types.h:1818
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2022 Arm Limited.
std::vector< MemoryInfo > MemoryRequirements
Definition: Types.h:134
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.
Convolution Layer Weights Information class.
Definition: Types.h:2073
#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 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:1816
size_t total_size() const override
Returns the total size of the tensor in bytes.
Definition: TensorInfo.h:250
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:2287
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:1812
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:1052
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:1814
#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:1815
arm_compute::WeightFormat weight_format() const
Definition: Types.h:2123
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:2339
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1679
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:2489
static IScheduler & get()
Access the scheduler singleton.
Definition: Scheduler.cpp:94
void set_weight_format(arm_compute::WeightFormat weight_format)
Set weight format to be used.
Definition: Types.h:2592
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.