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
|
/*
* Copyright (c) 2022-2023 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifndef ACL_SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUKERNELCOMPONENTGRAPH
#define ACL_SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUKERNELCOMPONENTGRAPH
#include "src/dynamic_fusion/sketch/ArgumentPack.h"
#include "src/dynamic_fusion/sketch/gpu/GpuComponentServices.h"
#include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentStream.h"
#include "src/dynamic_fusion/sketch/utils/DependencyGraph.h"
#include <vector>
namespace arm_compute
{
namespace experimental
{
namespace dynamic_fusion
{
class IGpuKernelComponent;
/** A multi-input (tensors), multi-output (tensors) acyclic directed graph of gpu kernel components
* Its main purposes are:
* - Perform "graph-level" optimizations like fusion of kernel components (not the fusion of operators)
* - Automatically assign memory descriptions @ref MemoryDescriptor of all tensors based on graph topology
*/
class GpuKernelComponentGraph
{
public:
/** Constructor
*
* @param[in] context @ref GpuWorkloadContext to be used by the graph
* @param[in] services @ref GpuComponentServices to be used by the graph
*/
GpuKernelComponentGraph(GpuWorkloadContext *context, GpuComponentServices *services);
/** Prevent instances of this class from being copy constructed */
GpuKernelComponentGraph(const GpuKernelComponentGraph &graph) = delete;
/** Prevent instances of this class from being copied */
GpuKernelComponentGraph &operator=(const GpuKernelComponentGraph &graph) = delete;
/** Allow instances of this class to be move constructed */
GpuKernelComponentGraph(GpuKernelComponentGraph &&graph) = default;
/** Allow instances of this class to be moved */
GpuKernelComponentGraph &operator=(GpuKernelComponentGraph &&graph) = default;
/** Create a new component and add it to the component graph
* Component id is automatically allocated
*
* @tparam T Component type
* @tparam Args Component argument types
*
* @param[in] args Component arguments except for component id, which is auto-allocated
*/
template <typename T, typename... Args>
void add_new_component(Args &&...args)
{
auto comp = _services->component_factory().create<T>(std::forward<Args>(args)...);
ArgumentPack<ITensorInfo> tensors = comp->tensors();
const auto src_tensor_ids = get_tensor_ids(tensors.get_const_src_tensors());
const auto dst_tensor_ids = get_tensor_ids(tensors.get_const_dst_tensors());
bool success = _dependency_graph.add_operator(comp->id(), src_tensor_ids, dst_tensor_ids);
ARM_COMPUTE_UNUSED(success);
ARM_COMPUTE_ERROR_ON(!success);
_components[comp->id()] = std::move(comp);
for (auto t : tensors.get_const_src_tensors())
{
_tensors[t->id()] = t;
}
for (auto t : tensors.get_const_dst_tensors())
{
_tensors[t->id()] = t;
}
}
/** Perform component fusion and serialize the graph into a stream of component groups
*
* @param[in] mem_map MemoryDescriptorMap for all the tensors in the component graph
*
* @return GpuKernelComponentStream
*/
GpuKernelComponentStream fuse(const MemoryDescriptorMap &mem_map) const;
private:
static std::vector<DependencyGraph::TensorId> get_tensor_ids(const std::vector<const ITensorInfo *> tensors);
GpuWorkloadContext *_context;
GpuComponentServices *_services;
std::map<ComponentId, std::unique_ptr<IGpuKernelComponent>> _components;
std::map<ITensorInfo::Id, const ITensorInfo *> _tensors;
DependencyGraph _dependency_graph{};
};
} // namespace dynamic_fusion
} // namespace experimental
} // namespace arm_compute
#endif /* ACL_SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUKERNELCOMPONENTGRAPH */
|
