ArmNN
 24.02
Public Member Functions | List of all members
NeonUnidirectionalSequenceLstmFloatWorkload Class Reference

#include <NeonUnidirectionalSequenceLstmFloatWorkload.hpp>

Inheritance diagram for NeonUnidirectionalSequenceLstmFloatWorkload:
[legend]
Collaboration diagram for NeonUnidirectionalSequenceLstmFloatWorkload:
[legend]

Public Member Functions

 NeonUnidirectionalSequenceLstmFloatWorkload (const UnidirectionalSequenceLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
virtual void Execute () const override
 
- Public Member Functions inherited from TypedWorkload< QueueDescriptor, DataTypes >
 TypedWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
 
- Public Member Functions inherited from BaseWorkload< QueueDescriptor >
 BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
 
virtual const std::string & GetName () const override
 
void ExecuteAsync (ExecutionData &executionData) override
 
void PostAllocationConfigure () override
 
const QueueDescriptor & GetData () const
 
arm::pipe::ProfilingGuid GetGuid () const final
 
virtual bool SupportsTensorHandleReplacement () const override
 
void ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
 
void ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
 
- Public Member Functions inherited from IWorkload
virtual ~IWorkload ()
 
virtual void RegisterDebugCallback (const DebugCallbackFunction &)
 
virtual armnn::Optional< armnn::MemoryRequirements > GetMemoryRequirements ()
 

Additional Inherited Members

- Protected Attributes inherited from BaseWorkload< QueueDescriptor >
QueueDescriptor m_Data
 
const arm::pipe::ProfilingGuid m_Guid
 
const std::string m_Name
 

Detailed Description

Definition at line 21 of file NeonUnidirectionalSequenceLstmFloatWorkload.hpp.

Constructor & Destructor Documentation

◆ NeonUnidirectionalSequenceLstmFloatWorkload()

NeonUnidirectionalSequenceLstmFloatWorkload ( const UnidirectionalSequenceLstmQueueDescriptor &  descriptor,
const WorkloadInfo &  info 
)

Definition at line 32 of file NeonUnidirectionalSequenceLstmFloatWorkload.cpp.

33  : FloatWorkload<UnidirectionalSequenceLstmQueueDescriptor>(descriptor, info)
34 {
35  // Report Profiling Details
36  ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonUnidirectionalSequenceLstmFloatWorkload_Construct",
37  descriptor.m_Parameters,
38  info,
39  GetGuid());
40 
41  const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
42  arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
43 
44  TensorInfo inputInfo = info.m_InputTensorInfos[0];
45  TensorInfo outputInfo = info.m_OutputTensorInfos[0];
46 
47  arm_compute::DataType armComputeDataType = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetDataType();
48  armnn::DataType armnnDataType = GetArmNNDataType(armComputeDataType);
49 
50  TensorShape inputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetShape();
51  TensorShape cellStateLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetShape();
52  TensorShape outputLayerShape = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetShape();
53 
54  unsigned int maxTime = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
55  unsigned int batchSize = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
56  unsigned int inputSize = inputLayerShape[2];
57  unsigned int outputSize = outputLayerShape[2];
58  unsigned int numUnits = cellStateLayerShape[1];
59 
60  const TensorShape timeMajorShapeInput({maxTime, batchSize, inputSize});
61  const TensorShape timeMajorShapeOutput({maxTime, batchSize, outputSize});
62 
63  //
64  // Permute: performed if Unidirectional Sequence Layer inputs/outputs are in batch major format.
65  //
66  if (!m_Data.m_Parameters.m_TimeMajor)
67  {
68  std::unique_ptr<arm_compute::NEPermute> layer(new arm_compute::NEPermute());
69 
70  TensorInfo permuteOutInfo = inputInfo;
71  permuteOutInfo.SetShape(timeMajorShapeInput);
72  BuildArmComputeTensor(m_PermuteFirstOut, permuteOutInfo);
73  armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermuteFirstOut);
74 
75  // Permute to time major format.
76  layer->configure(&input, &m_PermuteFirstOut, arm_compute::PermutationVector(0U,2U,1U));
77  m_Permute1.reset(layer.release());
78  }
79 
80  //
81  // Split and Concat Tensors
82  //
83  for (unsigned int i = 0; i < maxTime; ++i)
84  {
85  arm_compute::Tensor splitter_out;
86  arm_compute::Tensor concat_in;
87 
88  auto splitterTensorInfo = inputInfo;
89  auto concatTensorInfo = outputInfo;
90  splitterTensorInfo.SetShape({batchSize, inputSize});
91  concatTensorInfo.SetShape({batchSize, outputSize});
92  BuildArmComputeTensor(splitter_out, splitterTensorInfo);
93  BuildArmComputeTensor(concat_in, concatTensorInfo);
94 
95  armcomputetensorutils::InitialiseArmComputeTensorEmpty(splitter_out);
96  armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_in);
97 
98  // append to std::vector<arm_compute::Tensor>
99  m_SplitterOutputsTensors.push_back(std::move(splitter_out));
100  m_ConcatInputsTensors.push_back(std::move(concat_in));
101  }
102 
103  for (unsigned int i = 0; i < maxTime; ++i)
104  {
105  // append to std::vector<arm_compute::ITensor*>
106  m_SplitterOutputs.push_back(&m_SplitterOutputsTensors[i]);
107  m_ConcatInputs.push_back(&m_ConcatInputsTensors[i]);
108  }
109 
110  //
111  // Split
112  //
113  unsigned int numberDimensions = 3;
114  unsigned int dimension = 0; // splitting on 0-dimension (i.e. maxTime dimension)
115 
116  if (maxTime != 1) // ACL split does not work with only one element to split.
117  {
118  ViewsDescriptor splitterDesc(maxTime, numberDimensions);
119  unsigned int splitterDimSizes[3] = {1, batchSize, inputSize};
120  for (unsigned int outputIdx = 0u; outputIdx < maxTime; ++outputIdx)
121  {
122  splitterDesc.SetViewOriginCoord(outputIdx, dimension, splitterDimSizes[dimension] * outputIdx);
123  for (unsigned int dimIdx = 0u; dimIdx < numberDimensions; ++dimIdx)
124  {
125  splitterDesc.SetViewSize(outputIdx, dimIdx, splitterDimSizes[dimIdx]);
126  }
127  }
128 
129  std::set<unsigned int> splitAxis = ComputeSplitAxis(splitterDesc, timeMajorShapeInput);
130 
131  std::unique_ptr<arm_compute::NESplit> split_layer(new arm_compute::NESplit());
132  unsigned int aclAxisSplit = CalcAclAxis(splitterDesc.GetNumDimensions(),
133  *splitAxis.begin());
134  if (!m_Data.m_Parameters.m_TimeMajor)
135  {
136  split_layer->configure(&m_PermuteFirstOut, m_SplitterOutputs, aclAxisSplit);
137  } else
138  {
139  split_layer->configure(&input, m_SplitterOutputs, aclAxisSplit);
140  }
141 
142  split_layer->prepare();
143  m_Splitter.reset(split_layer.release());
144  }
145 
146  //
147  // Lstm
148  //
149  arm_compute::LSTMParams<arm_compute::ITensor> lstm_param;
150 
151  m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
152  BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
153 
154  m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
155  BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
156 
157  m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
158  BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
159 
160  m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
161  BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
162 
163  m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
164  BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
165 
166  m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
167  BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
168 
169  m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
170  BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
171 
172  m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
173  BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
174 
175  m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
176  BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
177 
178  // for future reference: check the AndroidNN API for the logic here
179  if (!m_Data.m_Parameters.m_CifgEnabled)
180  {
181  m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
182  BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
183 
184  m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
185  BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
186 
187  m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
188  if (m_Data.m_CellToInputWeights != nullptr)
189  {
190  BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
191  }
192 
193  m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
194  BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
195 
196  lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(),
197  m_RecurrentToInputWeightsTensor.get(),
198  m_Data.m_CellToInputWeights ? m_CellToInputWeightsTensor.get() : nullptr,
199  m_InputGateBiasTensor.get());
200  }
201 
202  if (m_Data.m_Parameters.m_ProjectionEnabled)
203  {
204  m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
205  BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
206 
207  m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
208  if (m_Data.m_ProjectionBias != nullptr)
209  {
210  BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
211  }
212 
213  lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(),
214  m_Data.m_ProjectionBias ? m_ProjectionBiasTensor.get() : nullptr);
215  }
216 
217  if (m_Data.m_Parameters.m_PeepholeEnabled)
218  {
219  m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
220  BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
221 
222  m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
223  BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
224 
225  lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get());
226  }
227 
228  if (m_Data.m_Parameters.m_LayerNormEnabled)
229  {
230  m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
231  if (!m_Data.m_Parameters.m_CifgEnabled)
232  {
233  BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
234  }
235 
236  m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
237  BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
238 
239  m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
240  BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
241 
242  m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
243  BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
244 
245  auto inputNormWeightTensor = m_Data.m_Parameters.m_CifgEnabled ? nullptr : m_InputLayerNormWeightsTensor.get();
246  lstm_param.set_layer_normalization_params(inputNormWeightTensor,
247  m_ForgetLayerNormWeightsTensor.get(),
248  m_CellLayerNormWeightsTensor.get(),
249  m_OutputLayerNormWeightsTensor.get());
250  }
251 
252  arm_compute::ITensor& output_state_in = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
253  arm_compute::ITensor& cell_state_in = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
254 
255  arm_compute::ITensor& output_state_out = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
256  arm_compute::ITensor& cell_state_out = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
257 
258  m_ScratchBuffer = std::make_unique<arm_compute::Tensor>();
259  if (m_Data.m_Parameters.m_CifgEnabled)
260  {
261  // scratch_buffer [num_units * 3, batch_size] with CIFG
262  BuildArmComputeTensor(*m_ScratchBuffer, TensorInfo({batchSize, numUnits * 3}, armnnDataType));
263  }
264  else
265  {
266  // scratch_buffer [num_units * 4, batch_size] without CIFG
267  BuildArmComputeTensor(*m_ScratchBuffer, TensorInfo({batchSize, numUnits * 4}, armnnDataType));
268  }
269 
270  // Need to be set at negative threshold to be compatible for ACL
271  float cell_threshold = m_Data.m_Parameters.m_ClippingThresCell;
272  float projection_threshold = m_Data.m_Parameters.m_ClippingThresProj;
273 
274  // For preparing the object for the class ActivationLayerInfo, consider 5 situations
275  arm_compute::ActivationLayerInfo activationLayerInfo =
276  ConvertLstmActivationFuncToAclLayerInfo(m_Data.m_Parameters.m_ActivationFunc);
277 
278  for (unsigned int i = 0; i != maxTime; ++i)
279  {
280  // Set LSTM input and output ITensors depending on:
281  // input format (timeMajor) & number of LSTM batches (maxTime).
282  arm_compute::ITensor* outputLSTM;
283  arm_compute::ITensor* inputLSTM;
284 
285  // If there is only one LSTM time major batch, we will not concat OR permute.
286  // Set input of LSTM to be first input ITensor.
287  // Set output of LSTM to be final output ITensor.
288  // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
289  if (maxTime == 1 && m_Data.m_Parameters.m_TimeMajor)
290  {
291  TensorShape inputShape = GetTensorShape(input.info()->tensor_shape(), 1U);
292  TensorShape outputShape = GetTensorShape((&output)->info()->tensor_shape(), 1U);
293 
294  TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
295  TensorShape outputShapeShrink({outputShape[1], outputShape[2]});
296 
297  auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
298  auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
299 
300  input.info()->set_tensor_shape(acl_input_shape_shrink);
301  inputLSTM = const_cast<arm_compute::ITensor*>(&input);
302 
303  output.info()->set_tensor_shape(acl_output_shape_shrink);
304  outputLSTM = &output;
305  }
306  // If there is only one LSTM batch major batch, we will not concat, only permute.
307  // Set input of LSTM to be output of initial permute.
308  // Set output of LSTM to be first element of m_ConcatInputs & use that value later in permute.
309  // LSTM output cannot be > 2 dimensions so need to resize its TensorInfo.
310  else if (maxTime == 1 && !m_Data.m_Parameters.m_TimeMajor)
311  {
312  TensorShape inputShape = GetTensorShape(m_PermuteFirstOut.info()->tensor_shape(), 1U);
313  TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
314  auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
315  m_PermuteFirstOut.info()->set_tensor_shape(acl_input_shape_shrink);
316  inputLSTM = &m_PermuteFirstOut;
317 
318  outputLSTM = const_cast<arm_compute::ITensor*>(m_ConcatInputs[i]);
319  }
320  // Batch major AND/OR 2+ LSTM batches so will use concat AND/OR permute later on.
321  else
322  {
323  inputLSTM = m_SplitterOutputs[i];
324  outputLSTM = const_cast<arm_compute::ITensor*>(m_ConcatInputs[i]);
325  }
326 
327  std::unique_ptr<arm_compute::NELSTMLayer> lstm_layer(new arm_compute::NELSTMLayer());
328  lstm_layer->configure(inputLSTM,
329  m_InputToForgetWeightsTensor.get(),
330  m_InputToCellWeightsTensor.get(),
331  m_InputToOutputWeightsTensor.get(),
332  m_RecurrentToForgetWeightsTensor.get(),
333  m_RecurrentToCellWeightsTensor.get(),
334  m_RecurrentToOutputWeightsTensor.get(),
335  m_ForgetGateBiasTensor.get(),
336  m_CellBiasTensor.get(),
337  m_OutputGateBiasTensor.get(),
338  &output_state_in,
339  &cell_state_in,
340  m_ScratchBuffer.get(),
341  &output_state_out,
342  &cell_state_out,
343  outputLSTM,
344  lstm_param,
345  activationLayerInfo,
346  cell_threshold,
347  projection_threshold);
348 
349  m_Layers.emplace_back(std::move(lstm_layer));
350  }
351 
352  armcomputetensorutils::InitialiseArmComputeTensorEmpty(*m_ScratchBuffer);
353 
354  InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
355  InitializeArmComputeTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
356  InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
357  InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
358  InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
359  InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
360  InitializeArmComputeTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
361  InitializeArmComputeTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
362  InitializeArmComputeTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
363 
364  if (!m_Data.m_Parameters.m_CifgEnabled)
365  {
366  InitializeArmComputeTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
367  InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
368  if (m_Data.m_CellToInputWeights != nullptr)
369  {
370  InitializeArmComputeTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
371  }
372  InitializeArmComputeTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
373  }
374 
375  if (m_Data.m_Parameters.m_ProjectionEnabled)
376  {
377  InitializeArmComputeTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
378  if (m_Data.m_ProjectionBias != nullptr)
379  {
380  InitializeArmComputeTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
381  }
382  }
383 
384  if (m_Data.m_Parameters.m_PeepholeEnabled)
385  {
386  InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
387  InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
388  }
389 
390  if (m_Data.m_Parameters.m_LayerNormEnabled)
391  {
392  if (!m_Data.m_Parameters.m_CifgEnabled)
393  {
394  InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
395  }
396  InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
397  InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
398  InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
399  }
400 
401  // Force Compute Library to perform the necessary copying and reshaping.
402  // After which delete all the input tensors that will no longer be needed.
403  for (uint32_t i = 0; i < m_Layers.size(); ++i)
404  {
405  m_Layers[i]->prepare();
406  }
407 
408  //
409  // Concat
410  //
411 
412  // Expand dimensions of LSTM outputs adding one empty dimension to fit concatenate inputs.
413  TensorShape shape = GetTensorShape(m_ConcatInputs[0]->info()->tensor_shape(), 1U);
414  TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
415  TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
416 
417  if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
418  {
419  for (unsigned int i = 0; i < maxTime; ++i)
420  {
421  m_ConcatInputs[i]->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
422  }
423 
424  ConcatDescriptor concatDescriptor(maxTime, numberDimensions); // maxTime = num inputs (aka. number of views).
425  for (unsigned int inputIdx = 0u; inputIdx < maxTime; ++inputIdx)
426  {
427  concatDescriptor.SetViewOriginCoord(inputIdx, dimension, inputIdx);
428  concatDescriptor.SetConcatAxis(dimension);
429  }
430 
431  m_Concat.reset(new arm_compute::NEConcatenateLayer());
432  unsigned int aclAxisConcat = CalcAclAxis(concatDescriptor.GetNumDimensions(), concatDescriptor.GetConcatAxis());
433  if (!m_Data.m_Parameters.m_TimeMajor)
434  {
435  TensorInfo concatOutputTensorInfo = outputInfo;
436  concatOutputTensorInfo.SetShape(timeMajorShapeOutput);
437  BuildArmComputeTensor(concat_out, concatOutputTensorInfo);
438  armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_out);
439 
440  m_Concat->configure(m_ConcatInputs, &concat_out, aclAxisConcat);
441  }
442  else
443  {
444  m_Concat->configure(m_ConcatInputs, &output, aclAxisConcat);
445  }
446 
447  m_Concat->prepare();
448  }
449  // If only one LSTM batch, we do not concat and/or permute.
450  // Must ensure final output info is expanded to correct batch major dimensions.
451  else
452  {
453  if (!m_Data.m_Parameters.m_TimeMajor)
454  {
455  output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandBatchMajor));
456  }
457  else
458  {
459  output.info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
460  }
461  }
462 
463  //
464  // Permute: only done if input/output are in batch major format.
465  //
466  if (!m_Data.m_Parameters.m_TimeMajor)
467  {
468  // Output now time major. Permute output back to batch major.
469  std::unique_ptr<arm_compute::NEPermute> layer(new arm_compute::NEPermute());
470  if (maxTime != 1)
471  {
472  layer->configure(&concat_out, &output, arm_compute::PermutationVector(0U, 2U, 1U));
473  }
474  else
475  {
476  layer->configure(m_ConcatInputs[0], &output, arm_compute::PermutationVector(0U, 2U, 1U));
477  }
478  m_Permute2.reset(layer.release());
479  }
480 
481  FreeUnusedTensors();
482 }

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, armnn::info, and QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters.

Member Function Documentation

◆ Execute()

void Execute ( ) const
overridevirtual

Implements IWorkload.

Definition at line 484 of file NeonUnidirectionalSequenceLstmFloatWorkload.cpp.

485 {
486  ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID("NeonUnidirectionalSequenceLstmFloatWorkload_Execute");
487  if (m_Permute1)
488  {
489  m_Permute1->run();
490  }
491  if (m_Splitter)
492  {
493  m_Splitter->run();
494  }
495  for (uint32_t i = 0; i < m_Layers.size(); ++i)
496  {
497  m_Layers[i]->run();
498  }
499  if (m_Concat)
500  {
501  m_Concat->run();
502  }
503  if (m_Permute2)
504  {
505  m_Permute2->run();
506  }
507 }

References ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID.

The documentation for this class was generated from the following files:
armnn::ConcatDescriptor
OriginsDescriptor ConcatDescriptor
Definition: DescriptorsFwd.hpp:59
armnn::ComputeSplitAxis
std::set< unsigned int > ComputeSplitAxis(const armnn::SplitterDescriptor &desc, const TensorShape &input)
Definition: ArmComputeUtils.hpp:246
armnn::InitializeArmComputeTensorData
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, TensorInfo tensorInfo, const ITensorHandle *handle)
Definition: NeonWorkloadUtils.hpp:68
armnn::DataType
DataType
Definition: Types.hpp:48
armnn::ConvertLstmActivationFuncToAclLayerInfo
arm_compute::ActivationLayerInfo ConvertLstmActivationFuncToAclLayerInfo(uint32_t activationFunction)
Definition: ArmComputeUtils.hpp:118
armnn::BoostLogSeverityMapping::info
@ info
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:27
ARMNN_REPORT_PROFILING_WORKLOAD_DESC
#define ARMNN_REPORT_PROFILING_WORKLOAD_DESC(name, desc, infos, guid)
Definition: Profiling.hpp:227
armnn::BaseWorkload::GetGuid
arm::pipe::ProfilingGuid GetGuid() const final
Definition: Workload.hpp:67
armnn::BaseWorkload::m_Data
QueueDescriptor m_Data
Definition: Workload.hpp:89
ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID
#define ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID(label)
Creates a profiling event that uses GetGuid() and GetName() from the calling class.
Definition: NeonWorkloadUtils.hpp:32
armnnUtils::GetTensorShape
armnn::TensorShape GetTensorShape(unsigned int numberOfBatches, unsigned int numberOfChannels, unsigned int height, unsigned int width, const armnn::DataLayout dataLayout)
Definition: TensorUtils.cpp:21
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:26