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
|
//
// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#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 VisitDequantizeOperator(DelegateData& delegateData,
TfLiteContext* tfLiteContext,
TfLiteNode* tfLiteNode,
int nodeIndex,
int32_t tfLiteDequantizeOperatorCode)
{
TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]];
if (IsDynamicTensor(tfLiteInputTensor))
{
TF_LITE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
tfLiteDequantizeOperatorCode, nodeIndex);
return kTfLiteError;
}
const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
if (IsDynamicTensor(tfLiteOutputTensor))
{
TF_LITE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnDelegate: Dynamic output tensors are not supported in operator #%d node #%d: ",
tfLiteDequantizeOperatorCode, nodeIndex);
return kTfLiteError;
}
const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
armnn::TensorInfo outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor, true);
UpdateConstantTensorOutputs(inputTensorInfo, outputTensorInfo);
bool isSupported = false;
armnn::BackendId setBackend;
auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported)
{
// If this is a Dequantize with a Constant input then will be replaced by a Constant layer that contains the
// dequantized values during optimization so there's no need to check if it can be supported by the backend
if (tflite::IsConstantTensor(&tfLiteInputTensor))
{
isSupported = true;
}
else
{
FORWARD_LAYER_SUPPORT_FUNC("DEQUANTIZE",
tfLiteContext,
IsDequantizeSupported,
delegateData.m_Backends,
isSupported,
setBackend,
inputTensorInfo,
outputTensorInfo);
}
};
if (!delegateData.m_Network)
{
validateFunc(outputTensorInfo, isSupported);
return isSupported ? kTfLiteOk : kTfLiteError;
}
auto layerName = GetLayerName(armnn::LayerType::Dequantize, nodeIndex);
armnn::IConnectableLayer* dequantizeLayer = delegateData.m_Network->AddDequantizeLayer(layerName.c_str());
dequantizeLayer->SetBackendId(setBackend);
ARMNN_ASSERT(dequantizeLayer != nullptr);
armnn::IOutputSlot& outputSlot = dequantizeLayer->GetOutputSlot(0);
outputSlot.SetTensorInfo(outputTensorInfo);
auto inputsTensorsProcess = ProcessInputs(dequantizeLayer,
delegateData,
tfLiteContext,
tfLiteNode,
nodeIndex);
if (inputsTensorsProcess == kTfLiteError)
{
return inputsTensorsProcess;
}
return Connect(dequantizeLayer, tfLiteNode, delegateData);
}
TfLiteStatus VisitQuantizeOperator(DelegateData& delegateData,
TfLiteContext* tfLiteContext,
TfLiteNode* tfLiteNode,
int nodeIndex,
int32_t tfLiteQuantizeOperatorCode)
{
TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
const TfLiteTensor* tfLiteTensors = tfLiteContext->tensors;
const TfLiteTensor& tfLiteInputTensor = tfLiteTensors[tfLiteNode->inputs->data[0]];
if (IsDynamicTensor(tfLiteInputTensor))
{
TF_LITE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
tfLiteQuantizeOperatorCode, nodeIndex);
return kTfLiteError;
}
const TfLiteTensor& tfLiteOutputTensor = tfLiteTensors[tfLiteNode->outputs->data[0]];
if (IsDynamicTensor(tfLiteOutputTensor))
{
TF_LITE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnDelegate: Dynamic output tensors are not supported in operator #%d node #%d: ",
tfLiteQuantizeOperatorCode, nodeIndex);
return kTfLiteError;
}
// Only affine per-layer quantization is supported.
if (!IsAffineQuantization(tfLiteOutputTensor))
{
TF_LITE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnDelegate: Only affine per-layer quantization is supported in operator #%d node #%d: ",
tfLiteQuantizeOperatorCode, nodeIndex);
return kTfLiteError;
}
const armnn::TensorInfo& inputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteInputTensor);
const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteTensor(tfLiteOutputTensor, true);
bool isSupported = false;
armnn::BackendId setBackend;
auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported)
{
FORWARD_LAYER_SUPPORT_FUNC("QUANTIZE",
tfLiteContext,
IsQuantizeSupported,
delegateData.m_Backends,
isSupported,
setBackend,
inputTensorInfo,
outputTensorInfo);
};
if (!delegateData.m_Network)
{
validateFunc(outputTensorInfo, isSupported);
return isSupported ? kTfLiteOk : kTfLiteError;
}
auto layerName = GetLayerName(armnn::LayerType::Quantize, nodeIndex);
armnn::IConnectableLayer* quantizeLayer = delegateData.m_Network->AddQuantizeLayer(layerName.c_str());
quantizeLayer->SetBackendId(setBackend);
ARMNN_ASSERT(quantizeLayer != nullptr);
armnn::IOutputSlot& outputSlot = quantizeLayer->GetOutputSlot(0);
outputSlot.SetTensorInfo(outputTensorInfo);
// try to connect the Constant Inputs if there are any
if (ProcessInputs(quantizeLayer, delegateData, tfLiteContext, tfLiteNode, nodeIndex) != kTfLiteOk)
{
return kTfLiteError;
}
return Connect(quantizeLayer, tfLiteNode, delegateData);
}
} // namespace armnnDelegate
|
