1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
|
//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "DelegateUtils.hpp"
#include <armnn/LstmParams.hpp>
#include <armnn/Tensor.hpp>
#include <armnn/utility/IgnoreUnused.hpp>
#include <tensorflow/lite/builtin_ops.h>
#include <tensorflow/lite/c/builtin_op_data.h>
#include <tensorflow/lite/c/common.h>
#include <tensorflow/lite/minimal_logging.h>
namespace armnnDelegate
{
TfLiteStatus VisitLstmOperator(DelegateData& delegateData,
TfLiteContext* tfLiteContext,
TfLiteNode* tfLiteNode,
int nodeIndex,
int32_t operatorCode)
{
auto numInputs = tfLiteNode->inputs->size;
if (numInputs < 2)
{
TF_LITE_MAYBE_KERNEL_LOG(
tfLiteContext, "TfLiteArmnnDelegate: Minimum number of inputs (%d != %d) in node #%d",
2, numInputs, nodeIndex);
return kTfLiteError;
}
const auto nodeParams = reinterpret_cast<TfLiteLSTMParams*>(tfLiteNode->builtin_data);
const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]];
if (!IsValid(tfLiteContext, tfLiteInputTensor, operatorCode, nodeIndex))
{
return kTfLiteError;
}
const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
if (!IsValid(tfLiteContext, tfLiteOutputTensor, operatorCode, nodeIndex))
{
return kTfLiteError;
}
// Set the params structure for the AddLstmLayer call
armnn::LstmInputParams params;
if (!IsOptionalOperandPresent(tfLiteNode, 1))
{
params.m_InputToInputWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 1);
}
params.m_InputToForgetWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 2);
params.m_InputToCellWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 3);
params.m_InputToOutputWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 4);
// Recurrent weight tensors of size {n_cell, n_output}
if (!IsOptionalOperandPresent(tfLiteNode, 5))
{
params.m_RecurrentToInputWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 5);
}
params.m_RecurrentToForgetWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 6);
params.m_RecurrentToCellWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 7);
params.m_RecurrentToOutputWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 8);
// Peephole weights tensors of size {n_cell}, representing a diagonal matrix.
if (!IsOptionalOperandPresent(tfLiteNode, 9))
{
params.m_CellToInputWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 9);
}
if (!IsOptionalOperandPresent(tfLiteNode, 10))
{
params.m_CellToForgetWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 10);
}
if (!IsOptionalOperandPresent(tfLiteNode, 11))
{
params.m_CellToOutputWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 11);
}
// Gates bias tensors of size {n_cell}
if (!IsOptionalOperandPresent(tfLiteNode, 12))
{
params.m_InputGateBias = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 12);
}
params.m_ForgetGateBias = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 13);
params.m_CellBias = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 14);
params.m_OutputGateBias = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 15);
// Projection weight tensor of size {n_output, n_cell}
if (!IsOptionalOperandPresent(tfLiteNode, 16))
{
params.m_ProjectionWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 16);
}
// Projection bias tensor of size {n_output}
if (!IsOptionalOperandPresent(tfLiteNode, 17))
{
params.m_ProjectionBias = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 17);
}
// These state tensors are defined as variable tensors, and will be modified by this op.
armnn::TensorInfo outputStateInInfo = GetTensorInfoForTfLiteTensor(tfLiteTensors[tfLiteNode->inputs->data[18]]);
armnn::TensorInfo cellStateInInfo = GetTensorInfoForTfLiteTensor(tfLiteTensors[tfLiteNode->inputs->data[19]]);
// Layer norm coefficient tensors of size {n_cell}, representing a diagonal matrix.
if (tfLiteNode->inputs->size >= 21 && !IsOptionalOperandPresent(tfLiteNode, 20))
{
params.m_InputLayerNormWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 20);
}
if (tfLiteNode->inputs->size >= 22 && !IsOptionalOperandPresent(tfLiteNode, 21))
{
params.m_ForgetLayerNormWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 21);
}
if (tfLiteNode->inputs->size >= 23 && !IsOptionalOperandPresent(tfLiteNode, 22))
{
params.m_CellLayerNormWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 22);
}
if (tfLiteNode->inputs->size >= 24 && !IsOptionalOperandPresent(tfLiteNode, 23))
{
params.m_OutputLayerNormWeights = GetConstTensorForTfLiteTensor(tfLiteTensors, tfLiteNode, 23);
}
// set the layer descriptor
armnn::LstmDescriptor desc;
desc.m_ActivationFunc = NonNegative(nodeParams->activation, nodeIndex);
desc.m_ClippingThresCell = nodeParams->cell_clip;
desc.m_ClippingThresProj = nodeParams->proj_clip;
desc.m_CifgEnabled = (params.m_InputToInputWeights == nullptr
|| params.m_RecurrentToInputWeights == nullptr
|| params.m_InputGateBias == nullptr);
desc.m_PeepholeEnabled = (params.m_CellToForgetWeights != nullptr || params.m_CellToOutputWeights != nullptr);
desc.m_ProjectionEnabled = (params.m_ProjectionWeights != nullptr);
desc.m_LayerNormEnabled = (params.m_InputLayerNormWeights != nullptr
|| params.m_ForgetLayerNormWeights != nullptr
|| params.m_CellLayerNormWeights != nullptr
|| params.m_OutputLayerNormWeights != nullptr);
const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor);
unsigned int batchSize = inputTensorInfo.GetShape()[0];
unsigned int outputSize = outputTensorInfo.GetShape()[1];
unsigned int numUnits = cellStateInInfo.GetShape()[1];
armnn::DataType dataType = inputTensorInfo.GetDataType();
float qScale = inputTensorInfo.GetQuantizationScale();
float qOffset = inputTensorInfo.GetQuantizationOffset();
armnn::TensorInfo scratchBufferTensorInfo({batchSize, numUnits * 3}, dataType, qScale, qOffset);
if (!desc.m_CifgEnabled)
{
scratchBufferTensorInfo = armnn::TensorInfo({batchSize, numUnits * 4}, dataType, qScale, qOffset);
}
armnn::TensorInfo cellStateOutTensorInfo({batchSize, numUnits}, dataType, qScale, qOffset);
armnn::TensorInfo outputStateOutTensorInfo({batchSize, outputSize}, dataType, qScale, qOffset);
armnn::LstmInputParamsInfo paramsInfo;
paramsInfo.m_InputToForgetWeights = &(params.m_InputToForgetWeights->GetInfo());
paramsInfo.m_InputToCellWeights = &(params.m_InputToCellWeights->GetInfo());
paramsInfo.m_InputToOutputWeights = &(params.m_InputToOutputWeights->GetInfo());
paramsInfo.m_RecurrentToForgetWeights = &(params.m_RecurrentToForgetWeights->GetInfo());
paramsInfo.m_RecurrentToCellWeights = &(params.m_RecurrentToCellWeights->GetInfo());
paramsInfo.m_RecurrentToOutputWeights = &(params.m_RecurrentToOutputWeights->GetInfo());
paramsInfo.m_ForgetGateBias = &(params.m_ForgetGateBias->GetInfo());
paramsInfo.m_CellBias = &(params.m_CellBias->GetInfo());
paramsInfo.m_OutputGateBias = &(params.m_OutputGateBias->GetInfo());
if (!desc.m_CifgEnabled)
{
paramsInfo.m_InputToInputWeights = &(params.m_InputToInputWeights->GetInfo());
paramsInfo.m_RecurrentToInputWeights = &(params.m_RecurrentToInputWeights->GetInfo());
if (params.m_CellToInputWeights != nullptr)
{
paramsInfo.m_CellToInputWeights = &(params.m_CellToInputWeights->GetInfo());
}
paramsInfo.m_InputGateBias = &(params.m_InputGateBias->GetInfo());
}
if (desc.m_ProjectionEnabled)
{
paramsInfo.m_ProjectionWeights = &(params.m_ProjectionWeights->GetInfo());
if (params.m_ProjectionBias != nullptr)
{
paramsInfo.m_ProjectionBias = &(params.m_ProjectionBias->GetInfo());
}
}
if (desc.m_PeepholeEnabled)
{
paramsInfo.m_CellToForgetWeights = &(params.m_CellToForgetWeights->GetInfo());
paramsInfo.m_CellToOutputWeights = &(params.m_CellToOutputWeights->GetInfo());
}
if (desc.m_LayerNormEnabled)
{
if(!desc.m_CifgEnabled)
{
paramsInfo.m_InputLayerNormWeights = &(params.m_InputLayerNormWeights->GetInfo());
}
paramsInfo.m_ForgetLayerNormWeights = &(params.m_ForgetLayerNormWeights->GetInfo());
paramsInfo.m_CellLayerNormWeights = &(params.m_CellLayerNormWeights->GetInfo());
paramsInfo.m_OutputLayerNormWeights = &(params.m_OutputLayerNormWeights->GetInfo());
}
bool isSupported = false;
auto validateFunc = [&](const armnn::TensorInfo& outputInfo, bool& isSupported)
{
FORWARD_LAYER_SUPPORT_FUNC("LSTM",
tfLiteContext,
IsLstmSupported,
delegateData.m_Backends,
isSupported,
inputTensorInfo,
outputStateInInfo,
cellStateInInfo,
scratchBufferTensorInfo,
outputStateOutTensorInfo,
cellStateOutTensorInfo,
outputInfo,
desc,
paramsInfo);
};
if (!delegateData.m_Network)
{
validateFunc(outputTensorInfo, isSupported);
return isSupported ? kTfLiteOk : kTfLiteError;
}
armnn::IConnectableLayer* layer = delegateData.m_Network->AddLstmLayer(desc, params);
ARMNN_ASSERT(layer != nullptr);
layer->GetOutputSlot(0).SetTensorInfo(scratchBufferTensorInfo);
layer->GetOutputSlot(1).SetTensorInfo(outputStateOutTensorInfo);
layer->GetOutputSlot(2).SetTensorInfo(cellStateOutTensorInfo);
layer->GetOutputSlot(3).SetTensorInfo(outputTensorInfo);
// Connect the inputs
// input_layer
delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[0]]->Connect(layer->GetInputSlot(0));
// cellStateIn
delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[18]]->Connect(layer->GetInputSlot(1));
//outputStateIn
delegateData.m_OutputSlotForNode[tfLiteNode->inputs->data[19]]->Connect(layer->GetInputSlot(2));
// In the test_model there is only 1 Output
armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(1);
delegateData.m_OutputSlotForNode[static_cast<unsigned long>(tfLiteNode->outputs->data[0])] = &outputSlot;
return kTfLiteOk;
}
} // namespace armnnDelegate
|
