1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include <armnnUtils/FloatingPointConverter.hpp>
7
8 #include "BFloat16.hpp"
9 #include "Half.hpp"
10
11 #include <armnn/utility/Assert.hpp>
12
13 namespace armnnUtils
14 {
15
ConvertFloat32To16(const float * srcFloat32Buffer,size_t numElements,void * dstFloat16Buffer)16 void FloatingPointConverter::ConvertFloat32To16(const float* srcFloat32Buffer,
17 size_t numElements,
18 void* dstFloat16Buffer)
19 {
20 ARMNN_ASSERT(srcFloat32Buffer != nullptr);
21 ARMNN_ASSERT(dstFloat16Buffer != nullptr);
22
23 armnn::Half* pHalf = static_cast<armnn::Half*>(dstFloat16Buffer);
24
25 for (size_t i = 0; i < numElements; i++)
26 {
27 pHalf[i] = armnn::Half(srcFloat32Buffer[i]);
28 }
29 }
30
ConvertFloat16To32(const void * srcFloat16Buffer,size_t numElements,float * dstFloat32Buffer)31 void FloatingPointConverter::ConvertFloat16To32(const void* srcFloat16Buffer,
32 size_t numElements,
33 float* dstFloat32Buffer)
34 {
35 ARMNN_ASSERT(srcFloat16Buffer != nullptr);
36 ARMNN_ASSERT(dstFloat32Buffer != nullptr);
37
38 const armnn::Half* pHalf = static_cast<const armnn::Half*>(srcFloat16Buffer);
39
40 for (size_t i = 0; i < numElements; i++)
41 {
42 dstFloat32Buffer[i] = pHalf[i];
43 }
44 }
45
ConvertFloat32ToBFloat16(const float * srcFloat32Buffer,size_t numElements,void * dstBFloat16Buffer)46 void FloatingPointConverter::ConvertFloat32ToBFloat16(const float* srcFloat32Buffer,
47 size_t numElements,
48 void* dstBFloat16Buffer)
49 {
50 ARMNN_ASSERT(srcFloat32Buffer != nullptr);
51 ARMNN_ASSERT(dstBFloat16Buffer != nullptr);
52
53 armnn::BFloat16* bf16 = static_cast<armnn::BFloat16*>(dstBFloat16Buffer);
54
55 for (size_t i = 0; i < numElements; i++)
56 {
57 bf16[i] = armnn::BFloat16(srcFloat32Buffer[i]);
58 }
59 }
60
ConvertBFloat16ToFloat32(const void * srcBFloat16Buffer,size_t numElements,float * dstFloat32Buffer)61 void FloatingPointConverter::ConvertBFloat16ToFloat32(const void* srcBFloat16Buffer,
62 size_t numElements,
63 float* dstFloat32Buffer)
64 {
65 ARMNN_ASSERT(srcBFloat16Buffer != nullptr);
66 ARMNN_ASSERT(dstFloat32Buffer != nullptr);
67
68 const armnn::BFloat16* bf16 = static_cast<const armnn::BFloat16*>(srcBFloat16Buffer);
69
70 for (size_t i = 0; i < numElements; i++)
71 {
72 dstFloat32Buffer[i] = bf16[i].ToFloat32();
73 }
74 }
75
76 } //namespace armnnUtils
77