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
117
118
119
120
121
122
123
124
125
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClDepthwiseConvolutionBaseWorkload.hpp"
#include "TypeUtils.hpp"
#include <backends/aclCommon/ArmComputeUtils.hpp>
#include <backends/aclCommon/ArmComputeTensorUtils.hpp>
#include <backends/cl/ClTensorHandle.hpp>
#include <backends/CpuTensorHandle.hpp>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status ClDepthwiseConvolutionWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const DepthwiseConvolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases)
{
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
arm_compute::TensorInfo aclBiasesInfo;
arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
if (descriptor.m_BiasEnabled)
{
BOOST_ASSERT(biases.has_value());
aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
optionalAclBiasesInfo = &aclBiasesInfo;
}
const arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor);
const unsigned int aclDepthMultiplier = weights.GetShape()[0];
return arm_compute::CLDepthwiseConvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
aclPadStrideInfo,
aclDepthMultiplier);
}
template<armnn::DataType... dataTypes>
ClDepthwiseConvolutionBaseWorkload<dataTypes...>::ClDepthwiseConvolutionBaseWorkload(
const DepthwiseConvolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<DepthwiseConvolution2dQueueDescriptor, dataTypes...>(descriptor, info)
{
auto& weightInfo = m_Data.m_Weight->GetTensorInfo();
m_KernelTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_KernelTensor, weightInfo);
if (m_Data.m_Parameters.m_BiasEnabled)
{
m_BiasTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo());
}
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);
std::string name = std::string("ClDepthwiseConvolution") +
GetDataTypeName(m_Data.m_Weight->GetTensorInfo().GetDataType()) + "Workload";
m_Data.ValidateInputsOutputs(name, 1, 1);
arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
const unsigned int depthMultiplier = weightInfo.GetShape()[0];
//Check for optimisation opportunities.
bool use3x3Optimisation = (weightInfo.GetShape()[3] == 3) && (weightInfo.GetShape()[2] == 3);
if (use3x3Optimisation)
{
m_DepthwiseConvolutionLayer = std::make_unique<arm_compute::CLDepthwiseConvolutionLayer3x3>();
static_cast<arm_compute::CLDepthwiseConvolutionLayer3x3*>(m_DepthwiseConvolutionLayer.get())->configure(
&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo,
depthMultiplier);
}
else
{
m_DepthwiseConvolutionLayer = std::make_unique<arm_compute::CLDepthwiseConvolutionLayer>();
static_cast<arm_compute::CLDepthwiseConvolutionLayer*>(m_DepthwiseConvolutionLayer.get())->configure(
&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo,
depthMultiplier);
}
BOOST_ASSERT(m_DepthwiseConvolutionLayer);
}
template<armnn::DataType... dataTypes>
void ClDepthwiseConvolutionBaseWorkload<dataTypes...>::FreeUnusedTensors()
{
FreeTensorIfUnused(m_KernelTensor);
FreeTensorIfUnused(m_BiasTensor);
}
// Generate known implementations for linker
template class ClDepthwiseConvolutionBaseWorkload<DataType::Float16, DataType::Float32>;
template class ClDepthwiseConvolutionBaseWorkload<DataType::QuantisedAsymm8>;
} // namespace armnn
|
