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
|
//
// Copyright © 2017-2019,2021-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonConvertFp16ToFp32Workload.hpp"
#include <armnnUtils/FloatingPointConverter.hpp>
#include <Half.hpp>
#include <backendsCommon/WorkloadUtils.hpp>
static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE;
namespace armnn
{
arm_compute::Status NeonConvertFp16ToFp32WorkloadValidate(const TensorInfo& input, const TensorInfo& output)
{
// Fallback to portable software implementation if Compute Library NECast won't work, so
// this method always returns success
armnn::IgnoreUnused(input);
armnn::IgnoreUnused(output);
return arm_compute::Status();
}
NeonConvertFp16ToFp32Workload::NeonConvertFp16ToFp32Workload(const ConvertFp16ToFp32QueueDescriptor& descriptor,
const WorkloadInfo& info)
: Float16ToFloat32Workload<ConvertFp16ToFp32QueueDescriptor>(descriptor, info)
{
this->m_Data.ValidateInputsOutputs("NeonConvertFp16ToFp32Workload", 1, 1);
arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
if (arm_compute::NECast::validate(input.info(), output.info(), g_AclConvertPolicy))
{
// Use NECast if supported (needs hardware support for FP16)
m_Cast.reset(new arm_compute::NECast());
m_Cast->configure(&input, &output, g_AclConvertPolicy);
}
else
{
// Else use software implementation using Half.hpp
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
}
void NeonConvertFp16ToFp32Workload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID("NeonConvertFp16ToFp32Workload_Execute");
if (m_Cast)
{
// Use NECast if supported and initialised
m_Cast->run();
}
else
{
// Else use softare implementation using Half.hpp
auto convertFunc = [](uint8_t* dst, const uint8_t* src, size_t size)
{
auto input = reinterpret_cast<const Half*>(src);
auto output = reinterpret_cast<float*>(dst);
size_t numElements = size/2; // 2 bytes per fp16
armnnUtils::FloatingPointConverter::ConvertFloat16To32(input, numElements, output);
};
for (const auto& pair : m_TensorHandlePairs)
{
CopyTensorContentsGeneric(pair.first, pair.second, convertFunc);
}
}
}
void NeonConvertFp16ToFp32Workload::ReplaceInputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
{
ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
this->m_Data.m_Inputs[slot] = tensorHandle;
try
{
Reconfigure();
}
catch(armnn::UnimplementedException& e)
{
// Cannot reconfigure, revert the slot back and throw the exception.
this->m_Data.m_Inputs[slot] = backupHandle;
throw e;
}
}
// Replace output tensor handle with the given TensorHandle
void NeonConvertFp16ToFp32Workload::ReplaceOutputTensorHandle(ITensorHandle* tensorHandle, unsigned int slot)
{
ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
this->m_Data.m_Inputs[slot] = tensorHandle;
try
{
Reconfigure();
}
catch(armnn::UnimplementedException& e)
{
// Cannot reconfigure, revert the slot back and throw the exception.
this->m_Data.m_Inputs[slot] = backupHandle;
throw e;
}
}
void NeonConvertFp16ToFp32Workload::Reconfigure()
{
throw armnn::UnimplementedException("Reconfigure not implemented for this workload");
}
} //namespace armnn
|
