From fd627ffaec8fd8801d980b4c91ee7c0607ab6aaf Mon Sep 17 00:00:00 2001 From: Jan Eilers Date: Thu, 25 Feb 2021 17:44:00 +0000 Subject: IVGCVSW-5687 Update Doxygen Docu * Update Doxygen Documentation for 21.02 release Signed-off-by: Jan Eilers Change-Id: I9ed2f9caab038836ea99d7b378d7899fe431a4e5 --- ..._convert_fp32_network_to_bf16_8hpp_source.xhtml | 139 +++++++++++++++++++++ 1 file changed, 139 insertions(+) create mode 100644 21.02/_convert_fp32_network_to_bf16_8hpp_source.xhtml (limited to '21.02/_convert_fp32_network_to_bf16_8hpp_source.xhtml') diff --git a/21.02/_convert_fp32_network_to_bf16_8hpp_source.xhtml b/21.02/_convert_fp32_network_to_bf16_8hpp_source.xhtml new file mode 100644 index 0000000000..59e5596f1d --- /dev/null +++ b/21.02/_convert_fp32_network_to_bf16_8hpp_source.xhtml @@ -0,0 +1,139 @@ + + + + + + + + + + + + + +ArmNN: src/armnn/optimizations/ConvertFp32NetworkToBf16.hpp Source File + + + + + + + + + + + + + + + + +
+
+ + + + ArmNN + + + +
+
+  21.02 +
+
+
+ + + + + + + +
+
+ +
+
+
+ +
+ +
+
+ + +
+ +
+ +
+
+
ConvertFp32NetworkToBf16.hpp
+
+
+Go to the documentation of this file.
1 //
2 // Copyright © 2020 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #pragma once
6 
7 #include "NetworkUtils.hpp"
8 #include "Optimization.hpp"
9 
10 #include <armnn/utility/PolymorphicDowncast.hpp>
11 
12 namespace armnn
13 {
14 namespace optimizations
15 {
16 
17 template <typename LayerT>
18 inline LayerT* ConvertWeight(Layer* l)
19 {
20  LayerT* layer = PolymorphicDowncast<LayerT*>(l);
21  if ((layer->GetType() == LayerType::Convolution2d || layer->GetType() == LayerType::FullyConnected)
22  && layer->m_Weight)
23  {
24  const TensorInfo& info = layer->m_Weight->GetTensorInfo();
25 
26  if (info.GetDataType() == DataType::Float32)
27  {
28  std::vector<BFloat16> newValues(info.GetNumElements());
29 
30  armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16(layer->m_Weight->template GetTensor<float>(),
31  info.GetNumElements(),
32  newValues.data());
33 
34  TensorInfo newInfo(info);
35  newInfo.SetDataType(DataType::BFloat16);
36  ConstTensor newInput(newInfo, newValues);
37  layer->m_Weight.reset(new ScopedCpuTensorHandle(newInput));
38  }
39  }
40  return layer;
41 }
42 
43 class ConvertFp32NetworkToBf16Impl
44 {
45 public:
46 
47  void Run(Graph& graph, Layer& layer) const
48  {
49  // Only convert Float32 To BFloat16 for the Input of Convolution2d layer and FullyConnected layer.
50  // And also convert weight data type from Float32 to Bfloat16.
51  // Do not convert bias data type.
52  if (layer.GetType() == LayerType::Convolution2d)
53  {
54  if (layer.GetDataType() == DataType::Float32)
55  {
56  InsertConvertFp32ToBf16LayersBefore(graph,layer);
57  ConvertWeight<Convolution2dLayer>(&layer);
58  }
59  }
60  else if (layer.GetType() == LayerType::FullyConnected)
61  {
62  if (layer.GetDataType() == DataType::Float32)
63  {
64  InsertConvertFp32ToBf16LayersBefore(graph,layer);
65  ConvertWeight<FullyConnectedLayer>(&layer);
66  }
67  }
68  }
69 
70 protected:
71  ConvertFp32NetworkToBf16Impl() = default;
72  ~ConvertFp32NetworkToBf16Impl() = default;
73 };
74 
75 using Fp32NetworkToBf16Converter = OptimizeForType<Layer, ConvertFp32NetworkToBf16Impl>;
76 
77 } // namespace optimizations
78 } // namespace armnn
+ + + +
armnn
Copyright (c) 2021 ARM Limited and Contributors.
Definition: 01_00_software_tools.dox:6
+
PolymorphicDowncast.hpp
+
armnn::optimizations::ConvertWeight
LayerT * ConvertWeight(Layer *l)
Definition: ConvertFp32NetworkToBf16.hpp:18
+ +
armnn::TensorInfo::GetDataType
DataType GetDataType() const
Definition: Tensor.hpp:194
+
armnn::ConstTensor
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
Definition: Tensor.hpp:314
+
armnn::Layer::GetType
LayerType GetType() const override
Returns the armnn::LayerType of this layer.
Definition: Layer.hpp:265
+
NetworkUtils.hpp
+ + + + +
armnnUtils::FloatingPointConverter::ConvertFloat32ToBFloat16
static void ConvertFloat32ToBFloat16(const float *srcFloat32Buffer, size_t numElements, void *dstBFloat16Buffer)
Definition: FloatingPointConverter.cpp:46
+ +
armnn::InsertConvertFp32ToBf16LayersBefore
std::vector< ConvertFp32ToBf16Layer * > InsertConvertFp32ToBf16LayersBefore(Graph &graph, Layer &layer, bool expectCorrectInputType)
Definition: NetworkUtils.cpp:90
+ +
armnn::Layer::GetDataType
DataType GetDataType() const
Definition: Layer.cpp:283
+ +
Optimization.hpp
+ + +
armnn::TensorInfo::GetNumElements
unsigned int GetNumElements() const
Definition: Tensor.hpp:192
+
+
+ + + + -- cgit v1.2.1