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
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#include "MemCopyWorkload.hpp"
#include "backends/CpuTensorHandle.hpp"
#if ARMCOMPUTECL_ENABLED
#include "backends/ClTensorHandle.hpp"
#endif
#if ARMCOMPUTENEON_ENABLED
#include "backends/NeonTensorHandle.hpp"
#endif
#include <cstring>
#include <boost/cast.hpp>
namespace armnn
{
namespace
{
template <typename SrcTensorHandleType, typename DstTensorHandleType>
void GatherTensorHandlePairs(const MemCopyQueueDescriptor& descriptor,
std::vector<std::pair<SrcTensorHandleType*, DstTensorHandleType*>>& tensorHandlePairs)
{
const unsigned int numInputs = boost::numeric_cast<unsigned int>(descriptor.m_Inputs.size());
tensorHandlePairs.reserve(numInputs);
for (unsigned int i = 0; i < numInputs; ++i)
{
SrcTensorHandleType* const srcTensorHandle = boost::polymorphic_downcast<SrcTensorHandleType*>(
descriptor.m_Inputs[i]);
DstTensorHandleType* const dstTensorHandle = boost::polymorphic_downcast<DstTensorHandleType*>(
descriptor.m_Outputs[i]);
tensorHandlePairs.emplace_back(srcTensorHandle, dstTensorHandle);
}
}
void CopyFromCpuToCpu(const ConstCpuTensorHandle& srcHandle, CpuTensorHandle& dstHandle)
{
const unsigned int numBytes = srcHandle.GetTensorInfo().GetNumBytes();
const void* const input = srcHandle.GetConstTensor<void>();
void* const output = dstHandle.GetTensor<void>();
std::memcpy(output, input, numBytes);
}
#if ARMCOMPUTECL_ENABLED || ARMCOMPUTENEON_ENABLED
#include "backends/ArmComputeTensorUtils.hpp"
template <armnn::DataType DataType>
void CopyFromCpuToAclBackend(const ConstCpuTensorHandle& srcHandle, arm_compute::ITensor& dstAclTensor)
{
using T = ResolveType<DataType>;
armnn::armcomputetensorutils::CopyArmComputeITensorData(srcHandle.GetConstTensor<T>(), dstAclTensor);
}
template <armnn::DataType DataType>
void CopyFromAclBackendToCpu(const arm_compute::ITensor& srcAclTensor, CpuTensorHandle& dstHandle)
{
using T = ResolveType<DataType>;
armnn::armcomputetensorutils::CopyArmComputeITensorData(srcAclTensor, dstHandle.GetTensor<T>());
}
#endif // ARMCOMPUTECL_ENABLED || ARMCOMPUTENEON_ENABLED
}
template <armnn::DataType DataType>
CopyFromCpuToCpuWorkload<DataType>::CopyFromCpuToCpuWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromCpuToCpuWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuRef, "CopyFromCpuToCpuWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
CopyFromCpuToCpu(*pair.first, *pair.second);
}
}
template class CopyFromCpuToCpuWorkload<DataType::Float32>;
template class CopyFromCpuToCpuWorkload<DataType::QuantisedAsymm8>;
#if ARMCOMPUTECL_ENABLED
template <armnn::DataType DataType>
CopyFromCpuToClWorkload<DataType>::CopyFromCpuToClWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromCpuToClWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::GpuAcc, "CopyFromCpuToClWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
IClTensorHandle& handle = *pair.second;
handle.Map(true);
CopyFromCpuToAclBackend<DataType>(*pair.first, handle.GetTensor());
handle.UnMap();
}
}
template class CopyFromCpuToClWorkload<DataType::Float32>;
template class CopyFromCpuToClWorkload<DataType::QuantisedAsymm8>;
template <armnn::DataType DataType>
CopyFromClToCpuWorkload<DataType>::CopyFromClToCpuWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromClToCpuWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::GpuAcc, "CopyFromClToCpuWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
IClTensorHandle& handle = *pair.first;
handle.Map(true);
CopyFromAclBackendToCpu<DataType>(handle.GetTensor(), *pair.second);
handle.UnMap();
}
}
template class CopyFromClToCpuWorkload<DataType::Float32>;
template class CopyFromClToCpuWorkload<DataType::QuantisedAsymm8>;
#endif // ARMCOMPUTECL_ENABLED
#if ARMCOMPUTENEON_ENABLED
template <armnn::DataType DataType>
CopyFromCpuToNeonWorkload<DataType>::CopyFromCpuToNeonWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromCpuToNeonWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, "CopyFromCpuToNeonWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
CopyFromCpuToAclBackend<DataType>(*pair.first, pair.second->GetTensor());
}
}
template class CopyFromCpuToNeonWorkload<DataType::Float32>;
template class CopyFromCpuToNeonWorkload<DataType::QuantisedAsymm8>;
template <armnn::DataType DataType>
CopyFromNeonToCpuWorkload<DataType>::CopyFromNeonToCpuWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromNeonToCpuWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, "CopyFromNeonToCpuWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
CopyFromAclBackendToCpu<DataType>(pair.first->GetTensor(), *pair.second);
}
}
template class CopyFromNeonToCpuWorkload<DataType::Float32>;
template class CopyFromNeonToCpuWorkload<DataType::QuantisedAsymm8>;
#endif // ARMCOMPUTENEON_ENABLED
#if ARMCOMPUTECL_ENABLED && ARMCOMPUTENEON_ENABLED
template <armnn::DataType DataType>
CopyFromNeonToClWorkload<DataType>::CopyFromNeonToClWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromNeonToClWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::GpuAcc, "CopyFromNeonToClWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
IClTensorHandle& handle = *pair.second;
handle.Map(true);
handle.GetTensor().copy_from(pair.first->GetTensor());
handle.UnMap();
}
}
template class CopyFromNeonToClWorkload<DataType::Float32>;
template class CopyFromNeonToClWorkload<DataType::QuantisedAsymm8>;
template <armnn::DataType DataType>
CopyFromClToNeonWorkload<DataType>::CopyFromClToNeonWorkload(const MemCopyQueueDescriptor& descriptor,
const WorkloadInfo& info)
: TypedWorkload<MemCopyQueueDescriptor, DataType>(descriptor, info)
{
GatherTensorHandlePairs(descriptor, m_TensorHandlePairs);
}
template <armnn::DataType DataType>
void CopyFromClToNeonWorkload<DataType>::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT(Compute::GpuAcc, "CopyFromClToNeonWorkload_Execute");
for (const auto& pair : m_TensorHandlePairs)
{
IClTensorHandle& handle = *pair.first;
handle.Map(true);
pair.second->GetTensor().copy_from(handle.GetTensor());
handle.UnMap();
}
}
template class CopyFromClToNeonWorkload<DataType::Float32>;
template class CopyFromClToNeonWorkload<DataType::QuantisedAsymm8>;
#endif // ARMCOMPUTECL_ENABLED && ARMCOMPUTENEON_ENABLED
}
|
