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
|
//
// Copyright © 2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include <OpaqueDelegateUtils.hpp>
namespace armnnOpaqueDelegate
{
TfLiteStatus VisitBatchMatMulOperator(DelegateData& delegateData,
TfLiteOpaqueContext* tfLiteContext,
TfLiteOpaqueNode* tfLiteNode,
int nodeIndex,
int32_t operatorCode)
{
TF_LITE_ENSURE_STATUS(ValidateNumInputs(tfLiteContext, tfLiteNode, 2, nodeIndex));
TF_LITE_ENSURE_STATUS(ValidateNumOutputs(tfLiteContext, tfLiteNode, 1, nodeIndex));
// Gather input indices and use to get input tensor.
auto numInputs = TfLiteOpaqueNodeNumberOfInputs(tfLiteNode);
const int* inputTensors;
if (TfLiteOpaqueNodeInputs(tfLiteNode, &inputTensors, &numInputs) != kTfLiteOk)
{
TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnOpaqueDelegate: Unable to gather input tensor indices from node #%d: ",
nodeIndex);
return kTfLiteError;
}
const TfLiteOpaqueTensor* kTfLiteLHSInputTensor =
TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[0]);
const TfLiteOpaqueTensor* kTfLiteRHSInputTensor =
TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, inputTensors[1]);
if (!IsValid(tfLiteContext, kTfLiteLHSInputTensor, operatorCode, nodeIndex))
{
return kTfLiteError;
}
if (!IsValid(tfLiteContext, kTfLiteRHSInputTensor, operatorCode, nodeIndex))
{
return kTfLiteError;
}
if (IsDynamicTensor(kTfLiteLHSInputTensor) || IsDynamicTensor(kTfLiteRHSInputTensor))
{
TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnOpaqueDelegate: Dynamic input tensors are not supported in operator #%d node #%d: ",
operatorCode, nodeIndex);
return kTfLiteError;
}
// Gather output indices and use to get output tensors.
int numOutputs = 0;
const int* outputTensors;
if (TfLiteOpaqueNodeOutputs(tfLiteNode, &outputTensors, &numOutputs) != kTfLiteOk)
{
TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnOpaqueDelegate: Unable to gather output tensor indices from node #%d: ",
nodeIndex);
return kTfLiteError;
}
const TfLiteOpaqueTensor* kTfLiteOutputTensor =
TfLiteOpaqueContextGetOpaqueTensor(tfLiteContext, outputTensors[0]);
if (IsDynamicTensor(kTfLiteOutputTensor))
{
TF_LITE_OPAQUE_MAYBE_KERNEL_LOG(
tfLiteContext,
"TfLiteArmnnOpaqueDelegate: Dynamic output tensors are not supported in operator #%d node #%d: ",
operatorCode, nodeIndex);
return kTfLiteError;
}
const armnn::TensorInfo& armnnLHSInputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(kTfLiteLHSInputTensor);
const armnn::TensorInfo& armnnRHSInputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(kTfLiteRHSInputTensor);
const armnn::TensorInfo& outputTensorInfo = GetTensorInfoForTfLiteOpaqueTensor(kTfLiteOutputTensor, true);
armnn::BatchMatMulDescriptor descriptor;
auto* params = reinterpret_cast<TfLiteBatchMatMulParams *>(TfLiteOpaqueNodeGetBuiltinData(tfLiteNode));
// Tensorflow params are called adjoint, however they are actually just transposes behind the scene. They do
// not perform ajoint.
descriptor.m_TransposeX = params->adj_x;
descriptor.m_TransposeY = params->adj_y;
// Check if supported
bool isSupported = false;
armnn::BackendId setBackend;
auto validateFunc = [&](const armnn::TensorInfo& outputTensorInfo, bool& isSupported)
{
FORWARD_LAYER_OPAQUE_SUPPORT_FUNC("BATCH_MATMUL",
tfLiteContext,
IsBatchMatMulSupported,
delegateData.m_Backends,
isSupported,
setBackend,
armnnLHSInputTensorInfo,
armnnRHSInputTensorInfo,
outputTensorInfo,
descriptor);
};
if (!delegateData.m_Network)
{
validateFunc(outputTensorInfo, isSupported);
return isSupported ? kTfLiteOk : kTfLiteError;
}
armnn::IConnectableLayer* layer = delegateData.m_Network->AddBatchMatMulLayer(descriptor);
layer->SetBackendId(setBackend);
ARMNN_ASSERT(layer != nullptr);
armnn::IOutputSlot& outputSlot = layer->GetOutputSlot(0);
outputSlot.SetTensorInfo(outputTensorInfo);
// try to connect the Constant Inputs if there are any
if(ProcessInputs(layer,delegateData, tfLiteContext, tfLiteNode) != kTfLiteOk )
{
return kTfLiteError;
}
return Connect(layer, tfLiteContext, tfLiteNode, delegateData);
}
} // namespace armnnOpaqueDelegate
|
