aboutsummaryrefslogtreecommitdiff
path: root/src/backends/neon/NeonTensorHandleFactory.cpp
blob: 4e013a37a1d65444451c9693cf4dd37872161c09 (plain)
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
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "NeonTensorHandleFactory.hpp"
#include "NeonTensorHandle.hpp"

#include "Layer.hpp"

#include <armnn/utility/IgnoreUnused.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>

namespace armnn
{

using FactoryId = ITensorHandleFactory::FactoryId;

std::unique_ptr<ITensorHandle> NeonTensorHandleFactory::CreateSubTensorHandle(ITensorHandle& parent,
                                                                              const TensorShape& subTensorShape,
                                                                              const unsigned int* subTensorOrigin)
                                                                              const
{
    const arm_compute::TensorShape shape = armcomputetensorutils::BuildArmComputeTensorShape(subTensorShape);

    arm_compute::Coordinates coords;
    coords.set_num_dimensions(subTensorShape.GetNumDimensions());
    for (unsigned int i = 0; i < subTensorShape.GetNumDimensions(); ++i)
    {
        // Arm compute indexes tensor coords in reverse order.
        unsigned int revertedIndex = subTensorShape.GetNumDimensions() - i - 1;
        coords.set(i, boost::numeric_cast<int>(subTensorOrigin[revertedIndex]));
    }

    const arm_compute::TensorShape parentShape = armcomputetensorutils::BuildArmComputeTensorShape(parent.GetShape());

    // In order for ACL to support subtensors the concat axis cannot be on x or y and the values of x and y
    // must match the parent shapes
    if (coords.x() != 0 || coords.y() != 0)
    {
        return nullptr;
    }
    if ((parentShape.x() != shape.x()) || (parentShape.y() != shape.y()))
    {
        return nullptr;
    }

    if (!::arm_compute::error_on_invalid_subtensor(__func__, __FILE__, __LINE__, parentShape, coords, shape))
    {
        return nullptr;
    }

    return std::make_unique<NeonSubTensorHandle>(
            PolymorphicDowncast<IAclTensorHandle*>(&parent), shape, coords);
}

std::unique_ptr<ITensorHandle> NeonTensorHandleFactory::CreateTensorHandle(const TensorInfo& tensorInfo) const
{
    return NeonTensorHandleFactory::CreateTensorHandle(tensorInfo, true);
}

std::unique_ptr<ITensorHandle> NeonTensorHandleFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
                                                                           DataLayout dataLayout) const
{
    return NeonTensorHandleFactory::CreateTensorHandle(tensorInfo, dataLayout, true);
}

std::unique_ptr<ITensorHandle> NeonTensorHandleFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
                                                                           const bool IsMemoryManaged) const
{
    auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo);
    if (IsMemoryManaged)
    {
        tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
    }
    // If we are not Managing the Memory then we must be importing
    tensorHandle->SetImportEnabledFlag(!IsMemoryManaged);
    tensorHandle->SetImportFlags(GetImportFlags());

    return tensorHandle;
}

std::unique_ptr<ITensorHandle> NeonTensorHandleFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
                                                                           DataLayout dataLayout,
                                                                           const bool IsMemoryManaged) const
{
    auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo, dataLayout);
    if (IsMemoryManaged)
    {
        tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
    }
    // If we are not Managing the Memory then we must be importing
    tensorHandle->SetImportEnabledFlag(!IsMemoryManaged);
    tensorHandle->SetImportFlags(GetImportFlags());

    return tensorHandle;
}

const FactoryId& NeonTensorHandleFactory::GetIdStatic()
{
    static const FactoryId s_Id(NeonTensorHandleFactoryId());
    return s_Id;
}

const FactoryId& NeonTensorHandleFactory::GetId() const
{
    return GetIdStatic();
}

bool NeonTensorHandleFactory::SupportsSubTensors() const
{
    return true;
}

MemorySourceFlags NeonTensorHandleFactory::GetExportFlags() const
{
    return 0;
}

MemorySourceFlags NeonTensorHandleFactory::GetImportFlags() const
{
    return 0;
}

std::vector<Capability> NeonTensorHandleFactory::GetCapabilities(const IConnectableLayer* layer,
                                                                 const IConnectableLayer* connectedLayer,
                                                                 CapabilityClass capabilityClass)

{
    IgnoreUnused(connectedLayer);
    std::vector<Capability> capabilities;
    if (capabilityClass == CapabilityClass::PaddingRequired)
    {
        auto search = paddingRequiredLayers.find((PolymorphicDowncast<const Layer*>(layer))->GetType());
        if ( search != paddingRequiredLayers.end())
        {
            Capability paddingCapability(CapabilityClass::PaddingRequired, true);
            capabilities.push_back(paddingCapability);
        }
    }
    return capabilities;
}

} // namespace armnn