ArmNN
 22.08
Public Member Functions | List of all members
EthosnRefConvolution2dWorkload Class Reference

#include <EthosnRefConvolution2dWorkload.hpp>

Inheritance diagram for EthosnRefConvolution2dWorkload:
BaseWorkload< Convolution2dQueueDescriptor > IWorkload

Public Member Functions

 EthosnRefConvolution2dWorkload (const Convolution2dQueueDescriptor &descriptor, const WorkloadInfo &info, bool winograd=false, std::string BlockConfigs="")
 
virtual void Execute () const override
 
- Public Member Functions inherited from BaseWorkload< Convolution2dQueueDescriptor >
 BaseWorkload (const Convolution2dQueueDescriptor &descriptor, const WorkloadInfo &info)
 
void ExecuteAsync (ExecutionData &executionData) override
 
void PostAllocationConfigure () override
 
const Convolution2dQueueDescriptor & GetData () const
 
arm::pipe::ProfilingGuid GetGuid () const final
 
virtual bool SupportsTensorHandleReplacement () const override
 
void ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
 
void ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
 
- Public Member Functions inherited from IWorkload
virtual ~IWorkload ()
 
virtual void RegisterDebugCallback (const DebugCallbackFunction &)
 
virtual armnn::Optional< armnn::MemoryRequirements > GetMemoryRequirements ()
 

Additional Inherited Members

- Protected Attributes inherited from BaseWorkload< Convolution2dQueueDescriptor >
Convolution2dQueueDescriptor m_Data
 
const arm::pipe::ProfilingGuid m_Guid
 

Detailed Description

Definition at line 15 of file EthosnRefConvolution2dWorkload.hpp.

Constructor & Destructor Documentation

◆ EthosnRefConvolution2dWorkload()

EthosnRefConvolution2dWorkload ( const Convolution2dQueueDescriptor &  descriptor,
const WorkloadInfo &  info,
bool  winograd = false,
std::string  BlockConfigs = "" 
)
explicit

Definition at line 34 of file EthosnRefConvolution2dWorkload.cpp.

References Convolution2dQueueDescriptor::m_Bias, Convolution2dDescriptor::m_BiasEnabled, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, and Convolution2dQueueDescriptor::m_Weight.

36  : BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
37 {
38  m_Weight = std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Weight));
39 
40  if (descriptor.m_Parameters.m_BiasEnabled)
41  {
42  m_Bias = std::make_unique<ScopedCpuTensorHandle>(*(descriptor.m_Bias));
43  }
44 
45  m_Winograd = winograd;
46 
47  m_IsBlockConfigsWinograd2dCompatible = false;
48  std::string part;
49  std::size_t pos = 0;
50  while (!(part = Split(BlockConfigs, ",", pos)).empty())
51  {
52  if (part == "16x8")
53  {
54  m_IsBlockConfigsWinograd2dCompatible = true;
55  }
56  else if (part == "8x16")
57  {
58  m_IsBlockConfigsWinograd2dCompatible = true;
59  }
60  else if (part == "8x8")
61  {
62  m_IsBlockConfigsWinograd2dCompatible = true;
63  }
64  }
65  if (BlockConfigs == "")
66  {
67  m_IsBlockConfigsWinograd2dCompatible = true;
68  }
69 }

Member Function Documentation

◆ Execute()

void Execute ( ) const
overridevirtual

Implements IWorkload.

Definition at line 72 of file EthosnRefConvolution2dWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_ETHOSN, armnn::ethosnref::CheckDataType(), armnn::ethosnref::DivideRoundUp(), TensorInfo::GetDataType(), DataLayoutIndexed::GetHeightIndex(), TensorInfo::GetQuantizationOffset(), TensorInfo::GetQuantizationScale(), TensorInfo::GetShape(), armnn::GetTensorInfo(), DataLayoutIndexed::GetWidthIndex(), Convolution2dDescriptor::m_BiasEnabled, BaseWorkload< Convolution2dQueueDescriptor >::m_Data, Convolution2dDescriptor::m_DataLayout, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, Convolution2dDescriptor::m_StrideX, Convolution2dDescriptor::m_StrideY, armnn::QAsymmU8, armnn::QSymmS16, and armnn::QSymmS8.

72  {
73 
74  ARMNN_SCOPED_PROFILING_EVENT_ETHOSN("EthosnRefConvolution2dWorkload_Execute");
75 
76  const TensorInfo& inputInfo = GetTensorInfo(m_Data.m_Inputs[0]);
77  const TensorInfo& weightsInfo = m_Weight->GetTensorInfo();
78 
79  bool useWinograd = false;
80 
81  if (m_Winograd == true) {
82  const armnnUtils::DataLayoutIndexed dataLayoutIndexed(m_Data.m_Parameters.m_DataLayout);
83  unsigned int weightsHeight = weightsInfo.GetShape()[dataLayoutIndexed.GetHeightIndex()];
84  unsigned int weightsWidth = weightsInfo.GetShape()[dataLayoutIndexed.GetWidthIndex()];
85  unsigned int xStride = m_Data.m_Parameters.m_StrideX;
86  unsigned int yStride = m_Data.m_Parameters.m_StrideY;
87 
88  bool isWinograd1D = weightsWidth == 1 || weightsHeight == 1;
89  bool isBlockSizeOk = isWinograd1D? true : m_IsBlockConfigsWinograd2dCompatible;
90 
91  uint32_t numMultsDirect, numMultsWinograd;
92 
93  if ( weightsWidth == 1 || weightsHeight == 1)
94  {
95  // 1D convolution kernel dim w x 1 or 1 x h
96  // numOfMultiplications = 2 * w or 2 * h DIRECT
97  // = 4 * CEIL(W/3) or 4 * CEIL(H/3) WINOGRAD
98  numMultsDirect = weightsWidth * weightsHeight * 2;
99  numMultsWinograd = 4 * DivideRoundUp(weightsWidth * weightsHeight, 3);
100  }
101  else
102  {
103  // 2D convolution kernel dim w x h
104  // numOfMultiplications = 4 * w * h DIRECT
105  // = 16 * CEIL(W/3) * CEIL(H/3) WINOGRAD
106  numMultsDirect = weightsWidth * weightsHeight * 4;
107  numMultsWinograd = 16 * DivideRoundUp(weightsWidth, 3) * DivideRoundUp(weightsHeight, 3);
108  }
109  if (numMultsWinograd < numMultsDirect &&
110  isBlockSizeOk &&
111  //TODO: FENCHSW-922 Need to revisit then when
112  // SW add support for Winogard + Strided
113  xStride == 1 && yStride == 1)
114  useWinograd = true;
115  }
116 
117  if (useWinograd == true)
118  {
119  std::cout << "[EthosNRef] Runnning in Winograd Mode.\n";
120  }
121  else
122  {
123  std::cout << "[EthosNRef] Runnning in Direct Mode.\n";
124  }
125 
126  if (CheckDataType(DataType::QSymmS16, inputInfo.GetDataType(), weightsInfo.GetDataType())) {
127  const int16_t* inputData = GetInputTensorData<int16_t>(0, m_Data);
128  const int16_t* weightsData = m_Weight->template GetConstTensor<int16_t>();
129  const int32_t* biasData = m_Data.m_Parameters.m_BiasEnabled ? m_Bias->template GetConstTensor<int32_t>() : nullptr;
130  const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);
131 
132  if(useWinograd == false)
133  {
134  EthosnRefConvImpl<armnn::Convolution2dQueueDescriptor, int16_t, int16_t, int32_t, int64_t>(
135  m_Data,
136  inputData, inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(),
137  weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
138  biasData,
139  outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
140  }
141  else
142  {
143  EthosnRefWinogradConvImpl<armnn::Convolution2dQueueDescriptor, int16_t, int16_t, int32_t>(
144  m_Data,
145  inputData, inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(),
146  weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
147  biasData,
148  outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
149  }
150  }
151  else if (CheckDataType(DataType::QSymmS8, inputInfo.GetDataType(), weightsInfo.GetDataType())) {
152  const int8_t* inputData = GetInputTensorData<int8_t>(0, m_Data);
153  const int8_t* weightsData = m_Weight->template GetConstTensor<int8_t>();
154  const int32_t* biasData = m_Data.m_Parameters.m_BiasEnabled ? m_Bias->template GetConstTensor<int32_t>() : nullptr;
155  const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);
156 
157  if(useWinograd == false)
158  {
159  EthosnRefConvImpl<armnn::Convolution2dQueueDescriptor, int8_t, int8_t, int32_t, int32_t>(
160  m_Data,
161  inputData, inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(),
162  weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
163  biasData,
164  outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
165  }
166  else
167  {
168  EthosnRefWinogradConvImpl<armnn::Convolution2dQueueDescriptor, int8_t, int8_t, int32_t>(
169  m_Data,
170  inputData, inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(),
171  weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
172  biasData,
173  outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
174  }
175  }
176  else { // QAsymmU8
177  assert(CheckDataType(DataType::QAsymmU8, inputInfo.GetDataType(), weightsInfo.GetDataType()));
178 
179  const uint8_t* inputData = GetInputTensorData<uint8_t>(0, m_Data);
180  const uint8_t* weightsData = m_Weight->template GetConstTensor<uint8_t>();
181  const int32_t* biasData = m_Data.m_Parameters.m_BiasEnabled ? m_Bias->template GetConstTensor<int32_t>() : nullptr;
182  const TensorInfo& outputInfo = GetTensorInfo(m_Data.m_Outputs[0]);
183 
184  if(useWinograd == false)
185  {
186  EthosnRefConvImpl<armnn::Convolution2dQueueDescriptor, uint8_t, uint8_t, int32_t, int32_t>(
187  m_Data,
188  inputData, inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(),
189  weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
190  biasData,
191  outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
192  }
193  else
194  {
195  EthosnRefWinogradConvImpl<armnn::Convolution2dQueueDescriptor, uint8_t, uint8_t, int32_t>(
196  m_Data,
197  inputData, inputInfo.GetQuantizationScale(), inputInfo.GetQuantizationOffset(),
198  weightsData, weightsInfo.GetQuantizationScale(), weightsInfo.GetQuantizationOffset(),
199  biasData,
200  outputInfo.GetQuantizationScale(), outputInfo.GetQuantizationOffset(), weightsInfo);
201  }
202  }
203 }
armnn::Convolution2dDescriptor::m_BiasEnabled
bool m_BiasEnabled
Enable/disable bias.
Definition: Descriptors.hpp:543
armnn::Convolution2dDescriptor::m_DataLayout
DataLayout m_DataLayout
The data layout to be used (NCHW, NHWC).
Definition: Descriptors.hpp:545
armnn::ethosnref::DivideRoundUp
uint32_t DivideRoundUp(uint32_t numerator, uint32_t denominator)
Definition: EthosnRefWorkloadUtils.hpp:51
armnn::ethosnref::CheckDataType
bool CheckDataType(DataType type, DataType inputType, DataType weightsType)
Definition: EthosnRefWorkloadUtils.hpp:56
armnn::Convolution2dDescriptor::m_StrideX
uint32_t m_StrideX
Stride value when proceeding through input for the width dimension.
Definition: Descriptors.hpp:535
armnnUtils::DataLayoutIndexed
Provides access to the appropriate indexes for Channels, Height and Width based on DataLayout...
Definition: DataLayoutIndexed.hpp:17
armnn::Convolution2dDescriptor::m_StrideY
uint32_t m_StrideY
Stride value when proceeding through input for the height dimension.
Definition: Descriptors.hpp:537
ARMNN_SCOPED_PROFILING_EVENT_ETHOSN
#define ARMNN_SCOPED_PROFILING_EVENT_ETHOSN(name)
Definition: EthosnRefWorkloadUtils.hpp:21
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:27
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:26
armnn::GetTensorInfo
const TensorInfo & GetTensorInfo(const ITensorHandle *tensorHandle)
float32 helpers
Definition: RefWorkloadUtils.hpp:26
The documentation for this class was generated from the following files: