1 //
2 // Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonQLstmWorkload.hpp"
7 #include "NeonWorkloadUtils.hpp"
8
9 #include "aclCommon/ArmComputeTensorUtils.hpp"
10
11 #include "neon/NeonTensorHandle.hpp"
12
13 namespace armnn
14 {
15 using namespace armcomputetensorutils;
16
NeonQLstmWorkload(const QLstmQueueDescriptor & descriptor,const WorkloadInfo & info)17 NeonQLstmWorkload::NeonQLstmWorkload(const QLstmQueueDescriptor& descriptor, const WorkloadInfo& info)
18 : NeonBaseWorkload<QLstmQueueDescriptor>(descriptor, info)
19 {
20 // Report Profiling Details
21 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonQLstmWorkload_Construct",
22 descriptor.m_Parameters,
23 info,
24 this->GetGuid());
25
26 arm_compute::LSTMParams<arm_compute::ITensor> qLstmParams;
27
28 // Mandatory params
29 m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
30 BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
31
32 m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
33 BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
34
35 m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
36 BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
37
38 m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
39 BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
40
41 m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
42 BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
43
44 m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
45 BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
46
47 m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
48 BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
49
50 m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
51 BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
52
53 m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
54 BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
55
56 // Create tensors for optional params if they are enabled
57 if (m_Data.m_Parameters.m_PeepholeEnabled)
58 {
59 m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
60
61 if (!m_Data.m_Parameters.m_CifgEnabled)
62 {
63 // In ACL this is categorised as a CIFG param and not a Peephole param
64 BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
65 }
66
67 m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
68 BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
69
70 m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
71 BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
72
73 // Set Peephole params
74 qLstmParams.set_peephole_params(m_CellToForgetWeightsTensor.get(),
75 m_CellToOutputWeightsTensor.get());
76 }
77
78 if (m_Data.m_Parameters.m_ProjectionEnabled)
79 {
80 m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
81 BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
82
83 m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
84 if (m_Data.m_ProjectionBias != nullptr)
85 {
86 BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
87 }
88
89 // Set projection params
90 qLstmParams.set_projection_params(
91 m_ProjectionWeightsTensor.get(),
92 m_Data.m_ProjectionBias != nullptr ? m_ProjectionBiasTensor.get() : nullptr);
93 }
94
95 if (m_Data.m_Parameters.m_LayerNormEnabled)
96 {
97 m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
98
99 if (!m_Data.m_Parameters.m_CifgEnabled)
100 {
101 BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
102 }
103
104 m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
105 BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
106
107 m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
108 BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
109
110 m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
111 BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
112
113 // Set layer norm params
114 qLstmParams.set_layer_normalization_params(
115 m_Data.m_InputLayerNormWeights != nullptr ? m_InputLayerNormWeightsTensor.get() : nullptr,
116 m_ForgetLayerNormWeightsTensor.get(),
117 m_CellLayerNormWeightsTensor.get(),
118 m_OutputLayerNormWeightsTensor.get());
119 }
120
121 if (!m_Data.m_Parameters.m_CifgEnabled)
122 {
123 m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
124 BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
125
126 m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
127 BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
128
129 m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
130 BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
131
132 // Set CIFG params
133 qLstmParams.set_cifg_params(
134 m_InputToInputWeightsTensor.get(),
135 m_RecurrentToInputWeightsTensor.get(),
136 m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr,
137 m_InputGateBiasTensor.get());
138 }
139
140 // Input/Output tensors
141 const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
142 arm_compute::ITensor& outputStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
143 const arm_compute::ITensor& cellStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
144
145 arm_compute::ITensor& outputStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
146 arm_compute::ITensor& cellStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
147 arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
148
149 // Set scalar descriptor params
150 qLstmParams.set_cell_clip_params(m_Data.m_Parameters.m_CellClip);
151 qLstmParams.set_projection_clip_params(m_Data.m_Parameters.m_ProjectionClip);
152 qLstmParams.set_hidden_state_params(m_Data.m_Parameters.m_HiddenStateZeroPoint,
153 m_Data.m_Parameters.m_HiddenStateScale);
154 qLstmParams.set_matmul_scale_params(m_Data.m_Parameters.m_InputIntermediateScale,
155 m_Data.m_Parameters.m_ForgetIntermediateScale,
156 m_Data.m_Parameters.m_CellIntermediateScale,
157 m_Data.m_Parameters.m_OutputIntermediateScale);
158
159 // QLSTM NEON configure
160 m_QLstmLayer.configure(&input,
161 m_InputToForgetWeightsTensor.get(),
162 m_InputToCellWeightsTensor.get(),
163 m_InputToOutputWeightsTensor.get(),
164 m_RecurrentToForgetWeightsTensor.get(),
165 m_RecurrentToCellWeightsTensor.get(),
166 m_RecurrentToOutputWeightsTensor.get(),
167 m_ForgetGateBiasTensor.get(),
168 m_CellBiasTensor.get(),
169 m_OutputGateBiasTensor.get(),
170 &cellStateIn,
171 &outputStateIn,
172 &cellStateOut,
173 &outputStateOut,
174 &output,
175 qLstmParams);
176
177 // Initialise ACL tensor data for mandatory params
178 InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
179 InitializeArmComputeTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
180 InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
181
182 InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
183 InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
184 InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
185
186 InitializeArmComputeTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
187 InitializeArmComputeTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
188 InitializeArmComputeTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
189
190 // Initialise ACL tensor data for optional params
191 if (!m_Data.m_Parameters.m_CifgEnabled)
192 {
193 InitializeArmComputeTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
194 InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
195 InitializeArmComputeTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
196 }
197
198 if (m_Data.m_Parameters.m_ProjectionEnabled)
199 {
200 InitializeArmComputeTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
201
202 if (m_Data.m_ProjectionBias != nullptr)
203 {
204 InitializeArmComputeTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
205 }
206 }
207
208 if (m_Data.m_Parameters.m_PeepholeEnabled)
209 {
210 if (!m_Data.m_Parameters.m_CifgEnabled)
211 {
212 InitializeArmComputeTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
213 }
214
215 InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
216 InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
217 }
218
219 if (m_Data.m_Parameters.m_LayerNormEnabled)
220 {
221 if (!m_Data.m_Parameters.m_CifgEnabled)
222 {
223 InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
224 }
225
226 InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
227 InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
228 InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
229 }
230
231 // QLSTM NEON prepare
232 m_QLstmLayer.prepare();
233
234 FreeUnusedTensors();
235 }
236
Execute() const237 void NeonQLstmWorkload::Execute() const
238 {
239 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonQuantizedLstmWorkload_Execute", this->GetGuid());
240 m_QLstmLayer.run();
241 }
242
NeonQLstmWorkloadValidate(const TensorInfo & input,const TensorInfo & cellStateIn,const TensorInfo & outputStateIn,const TensorInfo & cellStateOut,const TensorInfo & outputStateOut,const TensorInfo & output,const QLstmDescriptor & descriptor,const LstmInputParamsInfo & paramsInfo)243 arm_compute::Status NeonQLstmWorkloadValidate(const TensorInfo& input,
244 const TensorInfo& cellStateIn,
245 const TensorInfo& outputStateIn,
246 const TensorInfo& cellStateOut,
247 const TensorInfo& outputStateOut,
248 const TensorInfo& output,
249 const QLstmDescriptor& descriptor,
250 const LstmInputParamsInfo& paramsInfo)
251 {
252 arm_compute::LSTMParams<arm_compute::ITensorInfo> aclParamsInfo;
253
254 // Input/Output tensor info
255 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
256 const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn);
257 const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn);
258
259 const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut);
260 const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut);
261 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
262
263 // Mandatory tensor info
264 const arm_compute::TensorInfo aclInputToForgetWeightsInfo
265 = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights());
266 const arm_compute::TensorInfo aclInputToCellWeightsInfo
267 = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights());
268 const arm_compute::TensorInfo aclInputToOutputWeightsInfo
269 = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights());
270 const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo
271 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights());
272 const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo
273 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights());
274 const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo
275 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights());
276 const arm_compute::TensorInfo aclForgetGateBiasInfo
277 = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias());
278 const arm_compute::TensorInfo aclCellBiasInfo
279 = BuildArmComputeTensorInfo(paramsInfo.GetCellBias());
280 const arm_compute::TensorInfo aclOutputGateBiasInfo
281 = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias());
282
283 // Optional tensor info
284 arm_compute::TensorInfo aclInputToInputWeightsInfo;
285 arm_compute::TensorInfo aclRecurrentToInputWeightsInfo;
286
287 arm_compute::TensorInfo aclCellToInputWeightsInfo;
288 arm_compute::TensorInfo aclCellToForgetWeightsInfo;
289 arm_compute::TensorInfo aclCellToOutputWeightsInfo;
290
291 arm_compute::TensorInfo aclInputGateBiasInfo;
292
293 arm_compute::TensorInfo aclProjectionWeightsInfo;
294 arm_compute::TensorInfo aclProjectionBiasInfo;
295
296 arm_compute::TensorInfo aclInputLayerNormWeightsInfo;
297 arm_compute::TensorInfo aclForgetLayerNormWeightsInfo;
298 arm_compute::TensorInfo aclCellLayerNormWeightsInfo;
299 arm_compute::TensorInfo aclOutputLayerNormWeightsInfo;
300
301 // Create tensor info for optional params if they are enabled
302 if (descriptor.m_PeepholeEnabled)
303 {
304 if (!descriptor.m_CifgEnabled)
305 {
306 aclCellToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToInputWeights());
307 }
308
309 aclCellToForgetWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToForgetWeights());
310 aclCellToOutputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToOutputWeights());
311
312 // Set peephole params info
313 aclParamsInfo.set_peephole_params(&aclCellToForgetWeightsInfo,
314 &aclCellToOutputWeightsInfo);
315 }
316
317 if (descriptor.m_ProjectionEnabled)
318 {
319 aclProjectionWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionWeights());
320
321 if (paramsInfo.m_ProjectionBias != nullptr)
322 {
323 aclProjectionBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionBias());
324 }
325
326 // Set projection params info
327 aclParamsInfo.set_projection_params(
328 &aclProjectionWeightsInfo,
329 paramsInfo.m_ProjectionBias != nullptr ? &aclProjectionBiasInfo : nullptr);
330 }
331
332 if (descriptor.m_LayerNormEnabled)
333 {
334 if (!descriptor.m_CifgEnabled)
335 {
336 aclInputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputLayerNormWeights());
337 }
338
339 aclForgetLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetLayerNormWeights());
340 aclCellLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellLayerNormWeights());
341 aclOutputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputLayerNormWeights());
342
343 // Set layer norm params info
344 aclParamsInfo.set_layer_normalization_params(
345 paramsInfo.m_InputLayerNormWeights != nullptr ? &aclInputLayerNormWeightsInfo : nullptr,
346 &aclForgetLayerNormWeightsInfo,
347 &aclCellLayerNormWeightsInfo,
348 &aclOutputLayerNormWeightsInfo);
349 }
350
351 if (!descriptor.m_CifgEnabled)
352 {
353 aclInputToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights());
354 aclRecurrentToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights());
355 aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias());
356
357 // Set CIFG params info
358 aclParamsInfo.set_cifg_params(
359 &aclInputToInputWeightsInfo,
360 &aclRecurrentToInputWeightsInfo,
361 paramsInfo.m_CellToInputWeights != nullptr ? &aclCellToInputWeightsInfo : nullptr,
362 &aclInputGateBiasInfo);
363 }
364
365 // Set scalar descriptor params
366 aclParamsInfo.set_cell_clip_params(descriptor.m_CellClip);
367 aclParamsInfo.set_projection_clip_params(descriptor.m_ProjectionClip);
368 aclParamsInfo.set_hidden_state_params(descriptor.m_HiddenStateZeroPoint, descriptor.m_HiddenStateScale);
369 aclParamsInfo.set_matmul_scale_params(descriptor.m_InputIntermediateScale,
370 descriptor.m_ForgetIntermediateScale,
371 descriptor.m_CellIntermediateScale,
372 descriptor.m_OutputIntermediateScale);
373
374 // QLSTM NEON validate
375 return arm_compute::NEQLSTMLayer::validate(&aclInputInfo,
376 &aclInputToForgetWeightsInfo,
377 &aclInputToCellWeightsInfo,
378 &aclInputToOutputWeightsInfo,
379 &aclRecurrentToForgetWeightsInfo,
380 &aclRecurrentToCellWeightsInfo,
381 &aclRecurrentToOutputWeightsInfo,
382 &aclForgetGateBiasInfo,
383 &aclCellBiasInfo,
384 &aclOutputGateBiasInfo,
385 &aclCellStateInInfo,
386 &aclOutputStateInInfo,
387 &aclCellStateOutInfo,
388 &aclOutputStateOutInfo,
389 &aclOutputInfo,
390 aclParamsInfo);
391 }
392
FreeUnusedTensors()393 void NeonQLstmWorkload::FreeUnusedTensors()
394 {
395 FreeTensorIfUnused(m_InputToInputWeightsTensor);
396 FreeTensorIfUnused(m_InputToForgetWeightsTensor);
397 FreeTensorIfUnused(m_InputToCellWeightsTensor);
398 FreeTensorIfUnused(m_InputToOutputWeightsTensor);
399
400 FreeTensorIfUnused(m_RecurrentToInputWeightsTensor);
401 FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor);
402 FreeTensorIfUnused(m_RecurrentToCellWeightsTensor);
403 FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor);
404
405 FreeTensorIfUnused(m_CellToInputWeightsTensor);
406 FreeTensorIfUnused(m_CellToForgetWeightsTensor);
407 FreeTensorIfUnused(m_CellToOutputWeightsTensor);
408
409 FreeTensorIfUnused(m_InputGateBiasTensor);
410 FreeTensorIfUnused(m_ForgetGateBiasTensor);
411 FreeTensorIfUnused(m_CellBiasTensor);
412 FreeTensorIfUnused(m_OutputGateBiasTensor);
413
414 FreeTensorIfUnused(m_ProjectionWeightsTensor);
415 FreeTensorIfUnused(m_ProjectionBiasTensor);
416
417 FreeTensorIfUnused(m_InputLayerNormWeightsTensor);
418 FreeTensorIfUnused(m_ForgetLayerNormWeightsTensor);
419 FreeTensorIfUnused(m_CellLayerNormWeightsTensor);
420 FreeTensorIfUnused(m_OutputLayerNormWeightsTensor);
421 }
422
423 } //namespace armnn