Compute Library
 20.08
CLFullyConnectedLayer.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2020 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 "support/MemorySupport.h"
33 
34 #include <algorithm>
35 
36 namespace arm_compute
37 {
39 using namespace arm_compute::utils::cast;
40 
41 namespace
42 {
43 Status construct_gemmlowp_output_stage(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo &output,
44  GEMMLowpOutputStageInfo &gemmlowp_output_stage, ActivationLayerInfo activation_info)
45 {
46  gemmlowp_output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT;
47  gemmlowp_output_stage.gemmlowp_offset = 0;
48  gemmlowp_output_stage.gemmlowp_multiplier = 0;
49  gemmlowp_output_stage.gemmlowp_shift = 0;
50 
51  const auto data_type = input.data_type();
52 
53  // Configure output stage for quantized case
55  {
56  const QuantizationInfo oq_info = output.quantization_info();
57  const UniformQuantizationInfo iq_unif = input.quantization_info().uniform();
58  const UniformQuantizationInfo wq_unif = weights.quantization_info().uniform();
59  const UniformQuantizationInfo oq_unif = oq_info.uniform();
60 
61  const auto output_quant_info = (output.total_size() == 0) ? iq_unif : oq_unif;
62 
63  const float multiplier = (iq_unif.scale * wq_unif.scale) / output_quant_info.scale;
64  int output_multiplier = 0;
65  int output_shift = 0;
66  ARM_COMPUTE_RETURN_ON_ERROR(quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift));
67 
68  PixelValue type_min{};
69  PixelValue type_max{};
70  std::tie(type_min, type_max) = get_min_max(data_type);
71 
72  if(activation_info.enabled())
73  {
74  switch(activation_info.activation())
75  {
77  type_min = PixelValue(oq_unif.offset);
78  break;
80  type_min = PixelValue(oq_unif.offset);
81  type_max = PixelValue(activation_info.a(), data_type, oq_info);
82  break;
84  type_min = PixelValue(activation_info.b(), data_type, oq_info);
85  type_max = PixelValue(activation_info.a(), data_type, oq_info);
86  break;
87  default:
88  ARM_COMPUTE_ERROR("Activation function not supported.");
89  break;
90  }
91  }
92 
93  // Set the GEMMLowp output stage info
94  gemmlowp_output_stage.gemmlowp_offset = output_quant_info.offset;
95  gemmlowp_output_stage.gemmlowp_multiplier = output_multiplier;
96  gemmlowp_output_stage.gemmlowp_shift = output_shift;
97  gemmlowp_output_stage.gemmlowp_multipliers.push_back(output_multiplier);
98  gemmlowp_output_stage.gemmlowp_shifts.push_back(output_shift);
99  type_min.get(gemmlowp_output_stage.gemmlowp_min_bound);
100  type_max.get(gemmlowp_output_stage.gemmlowp_max_bound);
101  }
102 
103  return Status{};
104 }
105 
106 Status validate_mm(const ITensorInfo &input, const ITensorInfo &weights, const ITensorInfo *bias, const ITensorInfo &output, const FullyConnectedLayerInfo &fc_info)
107 {
108  GEMMLowpOutputStageInfo gemmlowp_output_stage;
109  ARM_COMPUTE_RETURN_ON_ERROR(construct_gemmlowp_output_stage(input, weights, output, gemmlowp_output_stage, fc_info.activation_info));
110 
111  const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
112  false, // is_b_reshaped
113  true, // reshape_b_only_on_first_run
114  0, // depth_output_gemm3d
115  false, // reinterpret_input_as_3d
116  fc_info.retain_internal_weights, // retain_internal_weights
117  gemmlowp_output_stage, // gemmlowp_output_stage
118  fc_info.fp_mixed_precision, // fp_mixed_precision
119  true, // broadcast_bias
120  ActivationLayerInfo()); // activation_info
121 
123  {
124  const UniformQuantizationInfo iq_info = input.quantization_info().uniform();
125  const UniformQuantizationInfo wq_info = weights.quantization_info().uniform();
126 
127  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
128  // Extract and negate input and weights offset
129  const QuantizationInfo input_quantization_info(iq_info.scale, -iq_info.offset);
130  const QuantizationInfo weights_quantization_info(wq_info.scale, -wq_info.offset);
131 
132  // Validate gemmlowp function
133  ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyCore::validate(&input.clone()->set_quantization_info(input_quantization_info),
134  &weights.clone()->set_quantization_info(weights_quantization_info),
135  bias,
136  &output,
137  gemm_info));
138  }
139  else
140  {
141  ARM_COMPUTE_RETURN_ON_ERROR(CLGEMM::validate(&input, &weights, bias, &output, 1.f, 1.f, gemm_info));
142  }
143 
144  return Status{};
145 }
146 } // namespace
147 
149 {
150  configure(CLKernelLibrary::get().get_compile_context(), input, output);
151 }
152 
154 {
155  auto k = arm_compute::support::cpp14::make_unique<CLTransposeKernel>();
156  k->configure(compile_context, input, output);
157  _kernel = std::move(k);
158 }
159 
161 {
162  return CLTransposeKernel::validate(input, output);
163 }
164 
165 CLFullyConnectedLayer::CLFullyConnectedLayer(std::shared_ptr<IMemoryManager> memory_manager, IWeightsManager *weights_manager)
166  : _memory_group(memory_manager), _weights_manager(weights_manager), _convert_weights(), _convert_weights_managed(), _reshape_weights_managed_function(), _flatten_layer(), _reshape_weights_function(),
167  _mm_gemm(memory_manager, weights_manager), _mm_gemmlowp(memory_manager), _flatten_output(), _converted_weights_output(), _reshape_weights_output(), _are_weights_converted(true),
168  _are_weights_reshaped(true), _is_fc_after_conv(true), _is_quantized(false), _is_prepared(false), _original_weights(nullptr)
169 {
170 }
171 void CLFullyConnectedLayer::configure_mm(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output,
172  const FullyConnectedLayerInfo &fc_info)
173 {
174  GEMMLowpOutputStageInfo gemmlowp_output_stage;
175  construct_gemmlowp_output_stage(*input->info(), *weights->info(), *output->info(), gemmlowp_output_stage, fc_info.activation_info);
176 
177  const GEMMInfo &gemm_info = GEMMInfo(false, // is_a_reshaped
178  false, // is_b_reshaped
179  true, // reshape_b_only_on_first_run
180  0, // depth_output_gemm3d
181  false, // reinterpret_input_as_3d
182  fc_info.retain_internal_weights, // retain_internal_weights
183  gemmlowp_output_stage, // gemmlowp_output_stage
184  fc_info.fp_mixed_precision, // fp_mixed_precision
185  true, // broadcast_bias
186  fc_info.activation_info); // activation_info
187 
188  if(_is_quantized)
189  {
190  // Since we need negative offsets for computing convolution, we need to change QuantizationInfo()
191  // Extract and negate input and weights offset
192  const QuantizationInfo input_quantization_info = input->info()->quantization_info();
194 
195  input->info()->set_quantization_info(QuantizationInfo(input_quantization_info.uniform().scale, -input_quantization_info.uniform().offset));
197 
198  // Configure gemmlowp function
199  _mm_gemmlowp.configure(compile_context, input, weights, bias, output, gemm_info);
200 
201  // Revert back QuantizatioInfo as input and weights could be used in other fully connected layers
202  input->info()->set_quantization_info(input_quantization_info);
204  }
205  else
206  {
207  // Configure matrix multiply kernel
208  _mm_gemm.configure(compile_context, input, weights, bias, output, 1.f, 1.f, gemm_info);
209  }
210 }
211 
212 void CLFullyConnectedLayer::configure_conv_fc(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output,
213  const FullyConnectedLayerInfo &fc_info)
214 {
215  ARM_COMPUTE_ERROR_ON((weights->info()->dimension(1) != (input->info()->dimension(0) * input->info()->dimension(1) * input->info()->dimension(2))));
216 
217  // If the fully connected layer is called after a convolution layer, the input tensor must be linearized
218 
219  // Initialize output tensor for flatten
220  TensorShape shape_flatten = compute_flatten_shape(input->info());
221  _flatten_output.allocator()->init(input->info()->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(shape_flatten).set_data_layout(DataLayout::NCHW));
222 
223  // Configure flatten kernel
224  _memory_group.manage(&_flatten_output);
225  _flatten_layer.configure(compile_context, input, &_flatten_output);
226 
227  // Configure matrix multiply kernel
228  configure_mm(compile_context, &_flatten_output, weights, bias, output, fc_info);
229 
230  // Allocate the output tensor for flatten once all the configure methods have been called
231  _flatten_output.allocator()->allocate();
232 }
233 
234 void CLFullyConnectedLayer::configure_fc_fc(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *bias, ICLTensor *output,
235  const FullyConnectedLayerInfo &fc_info)
236 {
237  ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != weights->info()->dimension(1));
238 
239  // Configure matrix multiply kernel
240  configure_mm(compile_context, input, weights, bias, output, fc_info);
241 }
242 
244  FullyConnectedLayerInfo fc_info)
245 {
246  configure(CLKernelLibrary::get().get_compile_context(), input, weights, biases, output, fc_info);
247 }
248 
249 void CLFullyConnectedLayer::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output,
250  FullyConnectedLayerInfo fc_info)
251 {
253 
254  // Perform validate step
256  weights->info(),
257  biases != nullptr ? biases->info() : nullptr,
258  output->info(),
259  fc_info));
260 
261  _are_weights_converted = true;
262  _are_weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
263  _is_fc_after_conv = true;
264  _is_quantized = is_data_type_quantized_asymmetric(input->info()->data_type());
265  _is_prepared = fc_info.retain_internal_weights;
266  _original_weights = weights;
267 
268  if(_weights_manager)
269  {
270  _weights_manager->manage(weights);
271  }
272 
273  const ICLTensor *weights_to_use = weights;
274 
275  // With the Fully Connected layer we can have 4 different cases:
276  // 1) Convolution layer -> Fully Connected layer without batches
277  // 2) Fully Connected layer -> Fully Connected layer without batches
278  // 3) Convolution layer -> Fully Connected layer with batches
279  // 4) Fully Connected layer -> Fully Connected layer with batches
280 
281  // Check if we have a fully connected layer with batches
282  const bool is_batched_fc_layer = output->info()->dimension(1) > 1;
283  if(is_batched_fc_layer)
284  {
285  _is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->info()->tensor_shape().cbegin() + 3,
286  input->info()->tensor_shape().cend(),
287  output->info()->tensor_shape().cbegin() + 1));
288  }
289  else
290  {
291  _is_fc_after_conv = input->info()->num_dimensions() > 1;
292  }
293 
294  // Reshape weights if needed
295  if(!_are_weights_reshaped)
296  {
297  if(_weights_manager && _weights_manager->are_weights_managed(weights))
298  {
299  _reshape_weights_managed_function.configure(compile_context, weights);
300  weights_to_use = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->acquire(weights, &_reshape_weights_managed_function));
301  }
302  else
303  {
304  // Reshape the weights
305  _reshape_weights_function.configure(compile_context, weights, &_reshape_weights_output);
306  weights_to_use = &_reshape_weights_output;
307  }
308  }
309 
310  // Convert weights if needed
311  if(_is_fc_after_conv && (input->info()->data_layout() != fc_info.weights_trained_layout))
312  {
313  if(_weights_manager && _weights_manager->are_weights_managed(weights_to_use))
314  {
315  _convert_weights_managed.configure(compile_context, weights_to_use,
316  input->info()->tensor_shape(),
317  fc_info.weights_trained_layout);
318  weights_to_use = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->acquire(weights, &_convert_weights_managed));
319  }
320  else
321  {
322  // Convert weights
323  _convert_weights.configure(compile_context, weights_to_use,
324  &_converted_weights_output,
325  input->info()->tensor_shape(),
326  fc_info.weights_trained_layout);
327 
328  weights_to_use = &_converted_weights_output;
329  }
330  _are_weights_converted = false;
331  }
332 
333  if(_is_fc_after_conv)
334  {
335  // Fully Connected layer after a Convolution Layer without batches
336  configure_conv_fc(compile_context, input, weights_to_use, biases, output, fc_info);
337  }
338  else
339  {
340  // Fully Connected layer after a Fully Connected Layer without batches
341  configure_fc_fc(compile_context, input, weights_to_use, biases, output, fc_info);
342  }
343 }
344 
346  FullyConnectedLayerInfo fc_info)
347 {
351  ARM_COMPUTE_RETURN_ERROR_ON(weights->num_dimensions() > 2);
354 
355  bool weights_reshaped = fc_info.transpose_weights ? fc_info.are_weights_reshaped : true;
356  bool is_fc_after_conv = true;
357 
358  const ITensorInfo &flatten_input = TensorInfo(input->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_flatten_shape(input)).set_data_layout(DataLayout::NCHW));
359  const ITensorInfo &reshaped_weights = TensorInfo(weights->clone()->set_is_resizable(true).reset_padding().set_tensor_shape(compute_transposed_shape(*weights)));
360  const ITensorInfo &converted_weights = weights_reshaped ? TensorInfo(weights->clone()->set_is_resizable(true).reset_padding()) : TensorInfo(*reshaped_weights.clone());
361 
362  // With the Fully Connected layer we can have 4 different cases:
363  // 1) Convolution layer -> Fully Connected layer without batches
364  // 2) Fully Connected layer -> Fully Connected layer without batches
365  // 3) Convolution layer -> Fully Connected layer with batches
366  // 4) Fully Connected layer -> Fully Connected layer with batches
367 
368  const ITensorInfo *input_to_use = input;
369  const ITensorInfo *weights_to_use = weights;
370 
371  // Check if we have a fully connected layer with batches
372  const bool is_batched_fc_layer = output->dimension(1) > 1;
373  if(is_batched_fc_layer)
374  {
375  is_fc_after_conv = (TensorShape::num_max_dimensions >= 4) && (std::equal(input->tensor_shape().cbegin() + 3,
376  input->tensor_shape().cend(),
377  output->tensor_shape().cbegin() + 1));
378  }
379  else
380  {
381  is_fc_after_conv = input->num_dimensions() > 1;
382  }
383 
384  if(!weights_reshaped)
385  {
386  // Validate reshape weights kernel
388  weights_to_use = &reshaped_weights;
389  }
390 
391  if(is_fc_after_conv && (input->data_layout() != fc_info.weights_trained_layout))
392  {
393  // Validate convert weights kernel
395  &converted_weights,
396  input->tensor_shape(),
397  fc_info.weights_trained_layout));
398  weights_to_use = &converted_weights;
399  }
400 
401  if(is_fc_after_conv)
402  {
403  // Fully Connected layer after a Convolution Layer without batches
404  ARM_COMPUTE_RETURN_ERROR_ON((weights_to_use->dimension(1) != (input->dimension(0) * input->dimension(1) * input->dimension(2))));
405 
406  // Validate flatten kernel
408  input_to_use = &flatten_input;
409  }
410  else
411  {
412  // Fully Connected layer after a Fully Connected Layer without batches
413  ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != weights_to_use->dimension(1));
414  }
415 
416  // Validate matrix multiply kernel
417  ARM_COMPUTE_RETURN_ON_ERROR(validate_mm(*input_to_use, *weights_to_use, biases, *output, fc_info));
418 
419  return Status{};
420 }
421 
423 {
424  prepare();
425 
426  MemoryGroupResourceScope scope_mg(_memory_group);
427 
428  // Linearize input if it comes from a convolutional layer
429  if(_is_fc_after_conv)
430  {
431  _flatten_layer.run();
432  }
433 
434  // Run matrix multiply
435  if(_is_quantized)
436  {
437  _mm_gemmlowp.run();
438  }
439  else
440  {
441  _mm_gemm.run();
442  }
443 }
444 
446 {
447  if(!_is_prepared)
448  {
449  if(!_weights_manager)
450  {
451  ARM_COMPUTE_ERROR_ON(!_original_weights->is_used());
452  }
453 
454  auto release_unused = [](CLTensor * w)
455  {
456  if(!w->is_used())
457  {
458  CLScheduler::get().queue().finish();
459  w->allocator()->free();
460  }
461  };
462 
463  // Pointer to current weights
464  const ICLTensor *cur_weights = _original_weights;
465 
466  // Reshape of the weights if needed (happens only once)
467  if(!_are_weights_reshaped)
468  {
469  if(_weights_manager && _weights_manager->are_weights_managed(_original_weights))
470  {
471  cur_weights = utils::cast::polymorphic_downcast<ICLTensor *>(_weights_manager->run(cur_weights, &_reshape_weights_managed_function));
472  }
473  else
474  {
475  // Run reshape weights kernel and mark weights as unused
476  _reshape_weights_output.allocator()->allocate();
477  _reshape_weights_function.run();
478 
479  cur_weights->mark_as_unused();
480  cur_weights = &_reshape_weights_output;
481  }
482  _are_weights_reshaped = true;
483  }
484 
485  // Convert weights if needed (happens only once)
486  if(!_are_weights_converted)
487  {
488  if(_weights_manager && _weights_manager->are_weights_managed(cur_weights))
489  {
490  _weights_manager->run(cur_weights, &_convert_weights_managed);
491  }
492  else
493  {
494  _converted_weights_output.allocator()->allocate();
495  _convert_weights.run();
496  cur_weights->mark_as_unused();
497  }
498 
499  _are_weights_converted = true;
500  }
501 
502  // Release reshaped weights if unused
503  release_unused(&_reshape_weights_output);
504 
505  // Prepare GEMM prepare and release unused weights
506  if(!_is_quantized)
507  {
508  _mm_gemm.prepare();
509  }
510 
511  // Release converted weights if unused
512  release_unused(&_reshape_weights_output);
513  release_unused(&_converted_weights_output);
514 
515  _is_prepared = true;
516  }
517 }
518 } // namespace arm_compute
bool is_data_type_quantized(DataType dt)
Check if a given data type is of quantized type.
Definition: Utils.h:1121
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of CLFlattenLayer.
SimpleTensor< float > w
Definition: DFT.cpp:156
Quantize using a fixed point multiplication.
void prepare() override
Prepare the function for executing.
Definition: CLGEMM.cpp:683
CLFullyConnectedLayer(std::shared_ptr< IMemoryManager > memory_manager=nullptr, IWeightsManager *weights_manager=nullptr)
Constructor.
void run() override
Run the kernels contained in the function.
Definition: CLGEMM.cpp:602
bool enabled() const
Check if initialised.
Definition: Types.h:1567
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
bool retain_internal_weights
Retain internal reshaped weights.
Definition: Types.h:1585
static CLScheduler & get()
Access the scheduler singleton.
Definition: CLScheduler.cpp:99
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
#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
bool is_used() const
Flags if the tensor is used or not.
Definition: ITensor.cpp:163
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
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
void configure(const ICLTensor *input)
Configures the CLFullyConnectedLayerReshapeWeights function.
Fully connected layer info.
Definition: Types.h:1580
static CLKernelLibrary & get()
Access the KernelLibrary singleton.
Store the tensor's metadata.
Definition: ITensorInfo.h:40
CLTensorAllocator * allocator()
Return a pointer to the tensor's allocator.
Definition: CLTensor.cpp:61
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
void configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel's inputs, output.
Status calculate_quantized_multiplier(float multiplier, int32_t *quant_multiplier, int32_t *shift, bool ignore_epsilon=false)
Calculate quantized representation of multiplier.
void manage(const ITensor *weights, ITransformWeights *parent=nullptr)
Start managing a weights tensor.
Status class.
Definition: Error.h:52
void run() override
Run the kernels contained in the function.
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
void prepare() override
Prepare the function for executing.
void init(const TensorInfo &input, size_t alignment=0)
Initialize a tensor based on the passed TensorInfo.
Copyright (c) 2017-2020 Arm Limited.
1 channel, 1 F16 per channel
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
TensorShape compute_transposed_shape(const ITensorInfo &input)
Calculate the transposed shape of a tensor.
void mark_as_unused() const
Marks a tensor as unused.
Definition: ITensor.cpp:168
void manage(IMemoryManageable *obj) override
Sets a object to be managed by the given memory group.
Definition: MemoryGroup.h:79
bool are_weights_managed(const ITensor *weights)
Check if the weights are managed.
TensorShape compute_flatten_shape(const ITensorInfo *input)
Calculate the flattened output shape of a tensor.
void configure(const ICLTensor *input, const ICLTensor *weights, const ICLTensor *biases, ICLTensor *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Set the input and output tensors.
Quantization information.
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of CLGEMMLowpMatrixMultiply...
void run() override final
Run the kernels contained in the function.
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
static Status validate(const ITensorInfo *input, const ITensorInfo *weights, const ITensorInfo *biases, const ITensorInfo *output, FullyConnectedLayerInfo fc_info=FullyConnectedLayerInfo())
Static function to check if given info will lead to a valid configuration of CLFullyConnectedLayer.
quantized, asymmetric fixed-point 8-bit number unsigned
bool are_weights_reshaped
Reshape the weights tensor if false.
Definition: Types.h:1584
UniformQuantizationInfo uniform() const
Return per layer quantization info.
GEMMLowp output stage info.
Definition: Types.h:1881
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
void configure(const ICLTensor *input, ICLTensor *output)
Initialise the kernel's input and output.
virtual ITensorInfo & set_quantization_info(const QuantizationInfo &quantization_info)=0
Set the quantization settings (scale and offset) of the tensor.
void run() override
Run the kernels contained in the function.
ActivationLayerInfo activation_info
Fused activation to apply after the matrix multiplication.
Definition: Types.h:1587
cl::CommandQueue & queue()
Accessor for the associated CL command queue.
Definition: CLScheduler.cpp:41
Weights manager interface to handle weights transformations.
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
Num samples, channels, height, width.
CLCompileContext class.
static Status validate(const ITensorInfo *input, const ITensorInfo *output, const TensorShape &original_input_shape, DataLayout data_layout)
Static function to check if given info will lead to a valid configuration of CLConvertFullyConnectedW...
bool is_data_type_quantized_asymmetric(DataType dt)
Check if a given data type is of asymmetric quantized type.
Definition: Utils.h:1143
__constant DATA_TYPE16 type_min
Definition: minmaxloc.cl:46
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
void allocate() override
Allocate size specified by TensorInfo of OpenCL memory.
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:210
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Memory group resources scope handling class.
Definition: IMemoryGroup.h:82
Interface for OpenCL tensor.
Definition: ICLTensor.h:42
void configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Initialise the kernel's inputs and output.
Definition: CLGEMM.cpp:497
DataLayout weights_trained_layout
Layout that the weights have been trained with.
Definition: Types.h:1582
void configure(const ICLTensor *input, ICLTensor *output, const TensorShape &original_input_shape, DataLayout data_layout)
Initialize the function.
bool fp_mixed_precision
Use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy.
Definition: Types.h:1586
void configure(const ICLTensor *input, ICLTensor *output)
Set the input and output tensors.
const QuantizationInfo weights_quantization_info
static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info=GEMMInfo())
Static function to check if given info will lead to a valid configuration of CLGEMM.
Definition: CLGEMM.cpp:556
void configure(const ICLTensor *input, const TensorShape &original_input_shape, DataLayout data_layout)
Configures the CLConvertFullyConnectedWeights function.
__constant DATA_TYPE16 type_max
Definition: minmaxloc.cl:47
bool transpose_weights
Transpose weights if true.
Definition: Types.h:1583
Store the tensor's metadata.
Definition: TensorInfo.h:45
GEMM information class.
Definition: Types.h:1932
ITensor * run(const ITensor *weights, ITransformWeights *weights_transform)
Run the reshape function.
ActivationFunction activation() const
Get the type of activation function.
Definition: Types.h:1552
quantized, asymmetric fixed-point 8-bit number signed
static constexpr size_t num_max_dimensions
Number of dimensions the tensor has.
Definition: Dimensions.h:45
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of CLFullyConnectedLayerRes...
std::tuple< PixelValue, PixelValue > get_min_max(DataType dt)
Compute the mininum and maximum values a data type can take.
Definition: Utils.h:560
ITensor * acquire(const ITensor *weights, ITransformWeights *weights_transform)
Acquire the requested reshape tensor of the selected weights.
static Status validate(const ITensorInfo *input, const ITensorInfo *output)
Static function to check if given info will lead to a valid configuration of CLTransposeKernel.
Basic implementation of the OpenCL tensor interface.
Definition: CLTensor.h:41