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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "NeonDepthwiseConvolutionUint8Workload.hpp"
#include <backends/neon/NeonLayerSupport.hpp>
#include <backends/CpuTensorHandle.hpp>
#include <backends/aclCommon/ArmComputeTensorUtils.hpp>
namespace armnn
{
using namespace armcomputetensorutils;
NeonDepthwiseConvolutionUint8Workload::NeonDepthwiseConvolutionUint8Workload(
const DepthwiseConvolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info)
: Uint8Workload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info)
{
const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
m_KernelTensor = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
if (m_Data.m_Parameters.m_BiasEnabled)
{
m_BiasTensor = std::make_unique<arm_compute::Tensor>();
BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
}
arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX,
m_Data.m_Parameters.m_StrideY,
m_Data.m_Parameters.m_PadLeft,
m_Data.m_Parameters.m_PadRight,
m_Data.m_Parameters.m_PadTop,
m_Data.m_Parameters.m_PadBottom,
arm_compute::DimensionRoundingType::FLOOR);
m_Data.ValidateInputsOutputs("NeonDepthwiseConvolutionUint8Workload", 1, 1);
arm_compute::ITensor& input = static_cast<INeonTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ITensor& output = static_cast<INeonTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
input.info()->set_data_layout(aclDataLayout);
output.info()->set_data_layout(aclDataLayout);
bool use3x3Optimisation = weightInfo.GetShape()[3] == 3 && weightInfo.GetShape()[2] == 3;
if (use3x3Optimisation)
{
m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer3x3>();
static_cast<arm_compute::NEDepthwiseConvolutionLayer3x3*>(
m_pDepthwiseConvolutionLayer.get())->configure(&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo);
}
else
{
m_pDepthwiseConvolutionLayer = std::make_unique<arm_compute::NEDepthwiseConvolutionLayer>();
static_cast<arm_compute::NEDepthwiseConvolutionLayer*>(
m_pDepthwiseConvolutionLayer.get())->configure(&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo);
}
BOOST_ASSERT(m_pDepthwiseConvolutionLayer);
InitializeArmComputeTensorData(*m_KernelTensor, m_Data.m_Weight);
if (m_BiasTensor)
{
InitializeArmComputeTensorData(*m_BiasTensor, m_Data.m_Bias);
}
m_pDepthwiseConvolutionLayer->prepare();
FreeUnusedTensors();
}
void NeonDepthwiseConvolutionUint8Workload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDepthwiseConvolutionUint8Workload_Execute");
BOOST_ASSERT(m_pDepthwiseConvolutionLayer);
m_pDepthwiseConvolutionLayer->run();
}
void NeonDepthwiseConvolutionUint8Workload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_KernelTensor);
FreeTensorIfUnused(m_BiasTensor);
}
} //namespace armnn
|
