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
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include <armnn/backends/IBackendInternal.hpp>
#include <arm_compute/core/Types.h>
#include <arm_compute/runtime/CL/CLBufferAllocator.h>
#include <aclCommon/BaseMemoryManager.hpp>
#include <arm_compute/runtime/CL/CLMemoryRegion.h>
#include <arm_compute/core/CL/CLKernelLibrary.h>
#include <CL/cl_ext.h>
namespace armnn
{
// add new capabilities here..
const BackendCapabilities gpuAccCapabilities("GpuAcc",
{
{"NonConstWeights", false},
{"AsyncExecution", false},
{"ProtectedContentAllocation", true}
});
class ClBackend : public IBackendInternal
{
public:
ClBackend() : m_CustomAllocator(nullptr) {};
ClBackend(std::shared_ptr<ICustomAllocator> allocator)
{
std::string err;
UseCustomMemoryAllocator(allocator, err);
}
~ClBackend() = default;
static const BackendId& GetIdStatic();
const BackendId& GetId() const override { return GetIdStatic(); }
IBackendInternal::IMemoryManagerUniquePtr CreateMemoryManager() const override;
IBackendInternal::IWorkloadFactoryPtr CreateWorkloadFactory(
const IBackendInternal::IMemoryManagerSharedPtr& memoryManager = nullptr) const override;
IBackendInternal::IWorkloadFactoryPtr CreateWorkloadFactory(
TensorHandleFactoryRegistry& registry) const override;
IWorkloadFactoryPtr CreateWorkloadFactory(const IMemoryManagerSharedPtr& memoryManager,
const ModelOptions& modelOptions) const override;
IWorkloadFactoryPtr CreateWorkloadFactory(class TensorHandleFactoryRegistry& tensorHandleFactoryRegistry,
const ModelOptions& modelOptions) const override;
IWorkloadFactoryPtr CreateWorkloadFactory(class TensorHandleFactoryRegistry& tensorHandleFactoryRegistry,
const ModelOptions& modelOptions,
MemorySourceFlags inputFlags,
MemorySourceFlags outputFlags) const override;
std::vector<ITensorHandleFactory::FactoryId> GetHandleFactoryPreferences() const override;
void RegisterTensorHandleFactories(TensorHandleFactoryRegistry& registry) override;
void RegisterTensorHandleFactories(TensorHandleFactoryRegistry& registry,
MemorySourceFlags inputFlags,
MemorySourceFlags outputFlags) override;
IBackendInternal::IBackendContextPtr CreateBackendContext(const IRuntime::CreationOptions&) const override;
IBackendInternal::IBackendProfilingContextPtr CreateBackendProfilingContext(
const IRuntime::CreationOptions&, IBackendProfilingPtr& backendProfiling) override;
IBackendInternal::Optimizations GetOptimizations() const override;
IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
IBackendInternal::ILayerSupportSharedPtr GetLayerSupport(const ModelOptions& modelOptions) const override;
OptimizationViews OptimizeSubgraphView(const SubgraphView& subgraph,
const ModelOptions& modelOptions) const override;
IBackendInternal::IBackendSpecificModelContextPtr CreateBackendSpecificModelContext(
const ModelOptions& modelOptions) const override;
BackendCapabilities GetCapabilities() const override
{
return gpuAccCapabilities;
};
virtual bool UseCustomMemoryAllocator(std::shared_ptr<ICustomAllocator> allocator,
armnn::Optional<std::string&> errMsg) override
{
IgnoreUnused(errMsg);
ARMNN_LOG(info) << "Using Custom Allocator for ClBackend";
// Set flag to signal the backend to use a custom memory allocator
m_CustomAllocator = std::make_shared<ClBackendCustomAllocatorWrapper>(std::move(allocator));
m_UsingCustomAllocator = true;
return m_UsingCustomAllocator;
}
// Cl requires a arm_compute::IAllocator we wrap the Arm NN ICustomAllocator to achieve this
class ClBackendCustomAllocatorWrapper : public arm_compute::IAllocator
{
public:
ClBackendCustomAllocatorWrapper(std::shared_ptr<ICustomAllocator> alloc) : m_CustomAllocator(alloc)
{}
// Inherited methods overridden:
void* allocate(size_t size, size_t alignment) override
{
auto alloc = m_CustomAllocator->allocate(size, alignment);
return MapAllocatedMemory(alloc, size, m_CustomAllocator->GetMemorySourceType());
}
void free(void* ptr) override
{
auto hostMemPtr = m_AllocatedBufferMappings[ptr];
clReleaseMemObject(static_cast<cl_mem>(ptr));
m_CustomAllocator->free(hostMemPtr);
}
std::unique_ptr<arm_compute::IMemoryRegion> make_region(size_t size, size_t alignment) override
{
auto hostMemPtr = m_CustomAllocator->allocate(size, alignment);
cl_mem buffer = MapAllocatedMemory(hostMemPtr, size, m_CustomAllocator->GetMemorySourceType());
return std::make_unique<ClBackendCustomAllocatorMemoryRegion>(cl::Buffer(buffer), hostMemPtr);
}
private:
cl_mem MapAllocatedMemory(void* memory, size_t size, MemorySource source)
{
// Round the size of the buffer to a multiple of the CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE
auto cachelineAlignment =
arm_compute::CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE>();
auto roundedSize = cachelineAlignment + size - (size % cachelineAlignment);
if (source == MemorySource::Malloc)
{
const cl_import_properties_arm importProperties[] =
{
CL_IMPORT_TYPE_ARM,
CL_IMPORT_TYPE_HOST_ARM,
0
};
cl_int error = CL_SUCCESS;
cl_mem buffer = clImportMemoryARM(arm_compute::CLKernelLibrary::get().context().get(),
CL_MEM_READ_WRITE,
importProperties,
memory,
roundedSize,
&error);
if (error == CL_SUCCESS)
{
m_AllocatedBufferMappings.insert(std::make_pair(static_cast<void *>(buffer), memory));
return buffer;
}
throw armnn::Exception(
"Mapping allocated memory from CustomMemoryAllocator failed, errcode: " + std::to_string(error));
}
throw armnn::Exception(
"Attempting to allocate memory with unsupported MemorySource type in CustomAllocator");
}
std::shared_ptr<ICustomAllocator> m_CustomAllocator;
std::map<void*, void*> m_AllocatedBufferMappings;
};
class ClBackendCustomAllocatorMemoryRegion : public arm_compute::ICLMemoryRegion
{
public:
// We need to have a new version of ICLMemoryRegion which holds a hostMemPtr to allow for cpu copy access
ClBackendCustomAllocatorMemoryRegion(const cl::Buffer &buffer, void* hostMemPtr)
: ICLMemoryRegion(buffer.getInfo<CL_MEM_SIZE>())
{
_mem = buffer;
m_HostMemPtr = hostMemPtr;
}
// Inherited methods overridden :
void* ptr() override
{
return nullptr;
}
void* map(cl::CommandQueue &q, bool blocking) override
{
armnn::IgnoreUnused(q, blocking);
if (m_HostMemPtr == nullptr)
{
throw armnn::Exception("ClBackend: Attempting to map memory with an invalid host ptr");
}
_mapping = m_HostMemPtr;
return _mapping;
}
void unmap(cl::CommandQueue &q) override
{
armnn::IgnoreUnused(q);
_mapping = nullptr;
}
void* m_HostMemPtr = nullptr;
};
std::shared_ptr<ClBackendCustomAllocatorWrapper> m_CustomAllocator;
bool m_UsingCustomAllocator = false;
};
} // namespace armnn
|
