ArmNN
 22.08
Public Member Functions | List of all members
ClUnidirectionalSequenceLstmFloatWorkload Class Reference

#include <ClUnidirectionalSequenceLstmFloatWorkload.hpp>

Inheritance diagram for ClUnidirectionalSequenceLstmFloatWorkload:
TypedWorkload< QueueDescriptor, DataTypes > BaseWorkload< QueueDescriptor > IWorkload

Public Member Functions

 ClUnidirectionalSequenceLstmFloatWorkload (const UnidirectionalSequenceLstmQueueDescriptor &descriptor, const WorkloadInfo &info, const arm_compute::CLCompileContext &clCompileContext)
 
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)
 
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
 

Detailed Description

Definition at line 22 of file ClUnidirectionalSequenceLstmFloatWorkload.hpp.

Constructor & Destructor Documentation

◆ ClUnidirectionalSequenceLstmFloatWorkload()

ClUnidirectionalSequenceLstmFloatWorkload ( const UnidirectionalSequenceLstmQueueDescriptor &  descriptor,
const WorkloadInfo &  info,
const arm_compute::CLCompileContext &  clCompileContext 
)

Definition at line 32 of file ClUnidirectionalSequenceLstmFloatWorkload.cpp.

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, WorkloadInfo::m_InputTensorInfos, and WorkloadInfo::m_OutputTensorInfos.

35  : FloatWorkload<UnidirectionalSequenceLstmQueueDescriptor>(descriptor, info)
36 {
37  // Report Profiling Details
38  ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClUnidirectionalSequenceLstmFloatWorkload_Construct",
39  descriptor.m_Parameters,
40  info,
41  GetGuid());
42 
43  const arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
44  arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
45 
46  TensorInfo inputInfo = info.m_InputTensorInfos[0];
47  TensorInfo outputInfo = info.m_OutputTensorInfos[2];
48 
49  arm_compute::DataType armComputeDataType = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetDataType();
50  armnn::DataType armnnDataType = GetArmNNDataType(armComputeDataType);
51 
52  TensorShape inputLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetShape();
53  TensorShape cellStateLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetShape();
54  TensorShape outputLayerShape = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetShape();
55 
56  unsigned int maxTime = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[0] : inputLayerShape[1];
57  unsigned int batchSize = m_Data.m_Parameters.m_TimeMajor ? inputLayerShape[1] : inputLayerShape[0];
58  unsigned int inputSize = inputLayerShape[2];
59  unsigned int outputSize = outputLayerShape[2];
60  unsigned int numUnits = cellStateLayerShape[1];
61 
62  const TensorShape timeMajorShapeInput({maxTime, batchSize, inputSize});
63  const TensorShape timeMajorShapeOutput({maxTime, batchSize, outputSize});
64 
65  //
66  // Permute: performed if Unidirectional Sequence Layer inputs/outputs are in batch major format.
67  //
68  if (!m_Data.m_Parameters.m_TimeMajor)
69  {
70  std::unique_ptr<arm_compute::CLPermute> layer(new arm_compute::CLPermute());
71 
72  TensorInfo permuteOutInfo = inputInfo;
73  permuteOutInfo.SetShape(timeMajorShapeInput);
74  BuildArmComputeTensor(m_PermuteFirstOut, permuteOutInfo);
75  armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermuteFirstOut);
76 
77  // Permute to time major format.
78  layer->configure(clCompileContext, &input, &m_PermuteFirstOut, arm_compute::PermutationVector(0U,2U,1U));
79  m_Permute1.reset(layer.release());
80  }
81 
82  //
83  // Split and Concat Tensors
84  //
85  for (unsigned int i = 0; i < maxTime; ++i)
86  {
87  arm_compute::CLTensor splitter_out;
88  arm_compute::CLTensor concat_in;
89 
90  auto splitterTensorInfo = inputInfo;
91  auto concatTensorInfo = outputInfo;
92  splitterTensorInfo.SetShape({batchSize, inputSize});
93  concatTensorInfo.SetShape({batchSize, outputSize});
94  BuildArmComputeTensor(splitter_out, splitterTensorInfo);
95  BuildArmComputeTensor(concat_in, concatTensorInfo);
96 
97  armcomputetensorutils::InitialiseArmComputeTensorEmpty(splitter_out);
98  armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_in);
99 
100  // append to std::vector<arm_compute::CLTensor>
101  m_SplitterOutputsTensors.push_back(std::move(splitter_out));
102  m_ConcatInputsTensors.push_back(std::move(concat_in));
103  }
104 
105  for (unsigned int i = 0; i < maxTime; ++i)
106  {
107  // append to std::vector<arm_compute::ICLTensor*>
108  m_SplitterOutputs.push_back(&m_SplitterOutputsTensors[i]);
109  m_ConcatInputs.push_back(&m_ConcatInputsTensors[i]);
110  }
111 
112  //
113  // Split
114  //
115  unsigned int numberDimensions = 3;
116  unsigned int dimension = 0; // splitting on 0-dimension (i.e. maxTime dimension)
117 
118  if (maxTime != 1) // ACL split does not work with only one element to split.
119  {
120  ViewsDescriptor splitterDesc(maxTime, numberDimensions);
121  unsigned int splitterDimSizes[3] = {1, batchSize, inputSize};
122  for (unsigned int outputIdx = 0u; outputIdx < maxTime; ++outputIdx)
123  {
124  splitterDesc.SetViewOriginCoord(outputIdx, dimension, splitterDimSizes[dimension] * outputIdx);
125  for (unsigned int dimIdx = 0u; dimIdx < numberDimensions; ++dimIdx)
126  {
127  splitterDesc.SetViewSize(outputIdx, dimIdx, splitterDimSizes[dimIdx]);
128  }
129  }
130 
131  std::set<unsigned int> splitAxis = ComputeSplitAxis(splitterDesc, timeMajorShapeInput);
132 
133  std::unique_ptr<arm_compute::CLSplit> split_layer(new arm_compute::CLSplit());
134  unsigned int aclAxisSplit = CalcAclAxis(splitterDesc.GetNumDimensions(), *splitAxis.begin());
135  if (!m_Data.m_Parameters.m_TimeMajor)
136  {
137  split_layer->configure(&m_PermuteFirstOut, m_SplitterOutputs, aclAxisSplit);
138  }
139  else
140  {
141  split_layer->configure(&input, m_SplitterOutputs, aclAxisSplit);
142  }
143 
144  split_layer->prepare();
145  m_Splitter.reset(split_layer.release());
146  }
147 
148  //
149  // Lstm
150  //
151  arm_compute::LSTMParams<arm_compute::ICLTensor> lstm_param;
152 
153  m_InputToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
154  BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
155 
156  m_InputToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>();
157  BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
158 
159  m_InputToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
160  BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
161 
162  m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
163  BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
164 
165  m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>();
166  BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
167 
168  m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
169  BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
170 
171  m_ForgetGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
172  BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
173 
174  m_CellBiasTensor = std::make_unique<arm_compute::CLTensor>();
175  BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
176 
177  m_OutputGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
178  BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
179 
180  // for future reference: check the AndroidNN API for the logic here
181  if (!m_Data.m_Parameters.m_CifgEnabled)
182  {
183  m_InputToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
184  BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
185 
186  m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
187  BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
188 
189  m_CellToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
190  if (m_Data.m_CellToInputWeights != nullptr)
191  {
192  BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
193  }
194 
195  m_InputGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
196  BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
197 
198  lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(),
199  m_RecurrentToInputWeightsTensor.get(),
200  m_Data.m_CellToInputWeights ? m_CellToInputWeightsTensor.get() : nullptr,
201  m_InputGateBiasTensor.get());
202  }
203 
204  if (m_Data.m_Parameters.m_ProjectionEnabled)
205  {
206  m_ProjectionWeightsTensor = std::make_unique<arm_compute::CLTensor>();
207  BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
208 
209  m_ProjectionBiasTensor = std::make_unique<arm_compute::CLTensor>();
210  if (m_Data.m_ProjectionBias != nullptr)
211  {
212  BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
213  }
214 
215  lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(),
216  m_Data.m_ProjectionBias ? m_ProjectionBiasTensor.get() : nullptr);
217  }
218 
219  if (m_Data.m_Parameters.m_PeepholeEnabled)
220  {
221  m_CellToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
222  BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
223 
224  m_CellToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
225  BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
226 
227  lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get());
228  }
229 
230  if (m_Data.m_Parameters.m_LayerNormEnabled)
231  {
232  m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>();
233  if (!m_Data.m_Parameters.m_CifgEnabled)
234  {
235  BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
236  }
237 
238  m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>();
239  BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
240 
241  m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>();
242  BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
243 
244  m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>();
245  BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
246 
247  auto inputNormWeightTensor = m_Data.m_Parameters.m_CifgEnabled ? nullptr : m_InputLayerNormWeightsTensor.get();
248  lstm_param.set_layer_normalization_params(inputNormWeightTensor,
249  m_ForgetLayerNormWeightsTensor.get(),
250  m_CellLayerNormWeightsTensor.get(),
251  m_OutputLayerNormWeightsTensor.get());
252  }
253 
254  arm_compute::ICLTensor& output_state_in = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
255  arm_compute::ICLTensor& cell_state_in = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
256 
257  arm_compute::ICLTensor& output_state_out = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
258  arm_compute::ICLTensor& cell_state_out = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
259 
260  m_ScratchBuffer = std::make_unique<arm_compute::CLTensor>();
261  if (m_Data.m_Parameters.m_CifgEnabled)
262  {
263  // scratch_buffer [num_units * 3, batch_size] with CIFG
264  BuildArmComputeTensor(*m_ScratchBuffer, TensorInfo({batchSize, numUnits * 3}, armnnDataType));
265  }
266  else
267  {
268  // scratch_buffer [num_units * 4, batch_size] without CIFG
269  BuildArmComputeTensor(*m_ScratchBuffer, TensorInfo({batchSize, numUnits * 4}, armnnDataType));
270  }
271 
272  // Need to be set at negative threshold to be compatible for ACL
273  float cell_threshold = m_Data.m_Parameters.m_ClippingThresCell;
274  float projection_threshold = m_Data.m_Parameters.m_ClippingThresProj;
275 
276  // For preparing the object for the class ActivationLayerInfo, consider 5 situations
277  arm_compute::ActivationLayerInfo activationLayerInfo =
278  ConvertLstmActivationFuncToAclLayerInfo(m_Data.m_Parameters.m_ActivationFunc);
279 
280  for (unsigned int i = 0; i != maxTime; ++i)
281  {
282  // Set LSTM input and output ITensors depending on:
283  // input format (timeMajor) & number of LSTM batches (maxTime).
284  arm_compute::ICLTensor* outputLSTM;
285  arm_compute::ICLTensor* inputLSTM;
286  // If there is only one LSTM time major batch, we will not concat OR permute.
287  // Set input of LSTM to be first input ITensor.
288  // Set output of LSTM to be final output ITensor.
289  // LSTM input/output cannot be > 2 dimensions so need to resize its TensorInfo.
290  if (maxTime == 1 && m_Data.m_Parameters.m_TimeMajor)
291  {
292  TensorShape inputShape = GetTensorShape((&input)->info()->tensor_shape(), 1U);
293  TensorShape outputShape = GetTensorShape((&output)->info()->tensor_shape(), 1U);
294  TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
295  TensorShape outputShapeShrink({outputShape[1], outputShape[2]});
296  auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
297  auto acl_output_shape_shrink = BuildArmComputeTensorShape(outputShapeShrink);
298  (&input)->info()->set_tensor_shape(acl_input_shape_shrink);
299  inputLSTM = const_cast<arm_compute::ICLTensor*>(&input);
300  (&output)->info()->set_tensor_shape(acl_output_shape_shrink);
301  outputLSTM = &output;
302  }
303  // If there is only one LSTM batch major batch, we will not concat, only permute.
304  // Set input of LSTM to be output of initial permute.
305  // Set output of LSTM to be first element of m_ConcatInputs & use that value later in permute.
306  // LSTM output cannot be > 2 dimensions so need to resize its TensorInfo.
307  else if (maxTime == 1 && !m_Data.m_Parameters.m_TimeMajor)
308  {
309  TensorShape inputShape = GetTensorShape(m_PermuteFirstOut.info()->tensor_shape(), 1U);
310  TensorShape inputShapeShrink({inputShape[1], inputShape[2]});
311  auto acl_input_shape_shrink = BuildArmComputeTensorShape(inputShapeShrink);
312  m_PermuteFirstOut.info()->set_tensor_shape(acl_input_shape_shrink);
313  inputLSTM = &m_PermuteFirstOut;
314  outputLSTM = const_cast<arm_compute::ICLTensor*>(m_ConcatInputs[i]);
315  }
316  // Batch major AND/OR 2+ LSTM batches so will use concat AND/OR permute later on.
317  else
318  {
319  inputLSTM = m_SplitterOutputs[i];
320  outputLSTM = const_cast<arm_compute::ICLTensor*>(m_ConcatInputs[i]);
321  }
322 
323  std::unique_ptr<arm_compute::CLLSTMLayer> lstm_layer(new arm_compute::CLLSTMLayer());
324  lstm_layer->configure(clCompileContext,
325  inputLSTM,
326  m_InputToForgetWeightsTensor.get(),
327  m_InputToCellWeightsTensor.get(),
328  m_InputToOutputWeightsTensor.get(),
329  m_RecurrentToForgetWeightsTensor.get(),
330  m_RecurrentToCellWeightsTensor.get(),
331  m_RecurrentToOutputWeightsTensor.get(),
332  m_ForgetGateBiasTensor.get(),
333  m_CellBiasTensor.get(),
334  m_OutputGateBiasTensor.get(),
335  &output_state_in,
336  &cell_state_in,
337  m_ScratchBuffer.get(),
338  &output_state_out,
339  &cell_state_out,
340  outputLSTM,
341  lstm_param,
342  activationLayerInfo,
343  cell_threshold,
344  projection_threshold);
345 
346  m_Layers.emplace_back(std::move(lstm_layer));
347  }
348 
349  armcomputetensorutils::InitialiseArmComputeTensorEmpty(*m_ScratchBuffer);
350 
351  InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
352  InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
353  InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
354  InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
355  InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
356  InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
357  InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
358  InitializeArmComputeClTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
359  InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
360 
361  if (!m_Data.m_Parameters.m_CifgEnabled)
362  {
363  InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
364  InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
365  if (m_Data.m_CellToInputWeights != nullptr)
366  {
367  InitializeArmComputeClTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
368  }
369  InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
370  }
371 
372  if (m_Data.m_Parameters.m_ProjectionEnabled)
373  {
374  InitializeArmComputeClTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
375  if (m_Data.m_ProjectionBias != nullptr)
376  {
377  InitializeArmComputeClTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
378  }
379  }
380 
381  if (m_Data.m_Parameters.m_PeepholeEnabled)
382  {
383  InitializeArmComputeClTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
384  InitializeArmComputeClTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
385  }
386 
387  if (m_Data.m_Parameters.m_LayerNormEnabled)
388  {
389  if (!m_Data.m_Parameters.m_CifgEnabled)
390  {
391  InitializeArmComputeClTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
392  }
393  InitializeArmComputeClTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
394  InitializeArmComputeClTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
395  InitializeArmComputeClTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
396  }
397 
398  // Force Compute Library to perform the necessary copying and reshaping.
399  // After which delete all the input tensors that will no longer be needed.
400  for (uint32_t i = 0; i < m_Layers.size(); ++i)
401  {
402  m_Layers[i]->prepare();
403  }
404 
405  //
406  // Concat
407  //
408 
409  // Expand dimensions of LSTM outputs adding one empty dimension to fit concatenate inputs.
410  TensorShape shape = GetTensorShape(m_ConcatInputs[0]->info()->tensor_shape(), 1U);
411  TensorShape shapeExpandTimeMajor({1, shape[0], shape[1]});
412  TensorShape shapeExpandBatchMajor({shape[0], 1, shape[1]});
413 
414  if (maxTime != 1) // ACL concat does not work with only one element to concatenate.
415  {
416  for (unsigned int i = 0; i < maxTime; ++i)
417  {
418  m_ConcatInputs[i]->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
419  }
420 
421  ConcatDescriptor concatDescriptor(maxTime, numberDimensions); // maxTime = num inputs (aka. number of views).
422  for (unsigned int inputIdx = 0u; inputIdx < maxTime; ++inputIdx)
423  {
424  concatDescriptor.SetViewOriginCoord(inputIdx, dimension, inputIdx);
425  concatDescriptor.SetConcatAxis(dimension);
426  }
427 
428  m_Concat.reset(new arm_compute::CLConcatenateLayer());
429  unsigned int aclAxisConcat = CalcAclAxis(concatDescriptor.GetNumDimensions(),
430  concatDescriptor.GetConcatAxis());
431  if (!m_Data.m_Parameters.m_TimeMajor)
432  {
433  TensorInfo concatOuputTensorInfo = outputInfo;
434  concatOuputTensorInfo.SetShape(timeMajorShapeOutput);
435  BuildArmComputeTensor(concat_out, concatOuputTensorInfo);
436  armcomputetensorutils::InitialiseArmComputeTensorEmpty(concat_out);
437 
438  m_Concat->configure(m_ConcatInputs, &concat_out, aclAxisConcat);
439  }
440  else
441  {
442  m_Concat->configure(m_ConcatInputs, &output, aclAxisConcat);
443  }
444 
445  m_Concat->prepare();
446  }
447  // If only one LSTM batch, we do not concat and/or permute.
448  // Must ensure final output info is expanded to correct batch major dimensions.
449  else
450  {
451  if (!m_Data.m_Parameters.m_TimeMajor)
452  {
453  (&output)->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandBatchMajor));
454  }
455  else
456  {
457  (&output)->info()->set_tensor_shape(BuildArmComputeTensorShape(shapeExpandTimeMajor));
458  }
459  }
460 
461  //
462  // Permute: only done if input/output are in batch major format.
463  //
464  if (!m_Data.m_Parameters.m_TimeMajor)
465  {
466  // Output now time major. Permute output back to batch major.
467  std::unique_ptr<arm_compute::CLPermute> layer(new arm_compute::CLPermute());
468  if (maxTime != 1)
469  {
470  layer->configure(clCompileContext, &concat_out, &output, arm_compute::PermutationVector(0U, 2U, 1U));
471  }
472  else
473  {
474  layer->configure(clCompileContext, m_ConcatInputs[0], &output, arm_compute::PermutationVector(0U, 2U, 1U));
475  }
476  m_Permute2.reset(layer.release());
477  }
478 
479  FreeUnusedTensors();
480 }
armnn::BaseWorkload::GetGuid
arm::pipe::ProfilingGuid GetGuid() const final
Definition: Workload.hpp:61
armnn::ConvertLstmActivationFuncToAclLayerInfo
arm_compute::ActivationLayerInfo ConvertLstmActivationFuncToAclLayerInfo(uint32_t activationFunction)
Definition: ArmComputeUtils.hpp:116
armnn::ComputeSplitAxis
std::set< unsigned int > ComputeSplitAxis(const armnn::SplitterDescriptor &desc, const TensorShape &input)
Definition: ArmComputeUtils.hpp:244
armnn::DataType
DataType
Definition: Types.hpp:48
armnn::BaseWorkload::m_Data
QueueDescriptor m_Data
Definition: Workload.hpp:83
armnnUtils::GetTensorShape
armnn::TensorShape GetTensorShape(unsigned int numberOfBatches, unsigned int numberOfChannels, unsigned int height, unsigned int width, const armnn::DataLayout dataLayout)
Definition: TensorUtils.cpp:19
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:27
armnn::ConcatDescriptor
OriginsDescriptor ConcatDescriptor
Definition: DescriptorsFwd.hpp:56
ARMNN_REPORT_PROFILING_WORKLOAD_DESC
#define ARMNN_REPORT_PROFILING_WORKLOAD_DESC(name, desc, infos, guid)
Definition: Profiling.hpp:227
armnn::InitializeArmComputeClTensorData
void InitializeArmComputeClTensorData(arm_compute::CLTensor &clTensor, const ConstTensorHandle *handle)
Definition: ClWorkloadUtils.hpp:116
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:26

Member Function Documentation

◆ Execute()

void Execute ( ) const
overridevirtual

Implements IWorkload.

Definition at line 482 of file ClUnidirectionalSequenceLstmFloatWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_CL_GUID.

483 {
484  ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClUnidirectionalSequenceLstmFloatWorkload_Execute", GetGuid());
485  if (m_Permute1)
486  {
487  m_Permute1->run();
488  }
489  if (m_Splitter)
490  {
491  m_Splitter->run();
492  }
493  for (uint32_t i = 0; i < m_Layers.size(); ++i)
494  {
495  m_Layers[i]->run();
496  }
497  if (m_Concat)
498  {
499  m_Concat->run();
500  }
501  if (m_Permute2)
502  {
503  m_Permute2->run();
504  }
505 }
ARMNN_SCOPED_PROFILING_EVENT_CL_GUID
#define ARMNN_SCOPED_PROFILING_EVENT_CL_GUID(name, guid)
Definition: ClWorkloadUtils.hpp:28
armnn::BaseWorkload::GetGuid
arm::pipe::ProfilingGuid GetGuid() const final
Definition: Workload.hpp:61
The documentation for this class was generated from the following files: