xref: /aosp_15_r20/external/ComputeLibrary/src/dynamic_fusion/sketch/gpu/GpuKernelComponentGraph.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2022-2023 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #ifndef SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUKERNELCOMPONENTGRAPH
25 #define SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUKERNELCOMPONENTGRAPH
26 
27 #include "src/dynamic_fusion/sketch/ArgumentPack.h"
28 #include "src/dynamic_fusion/sketch/gpu/GpuComponentServices.h"
29 #include "src/dynamic_fusion/sketch/gpu/GpuKernelComponentStream.h"
30 #include "src/dynamic_fusion/sketch/utils/DependencyGraph.h"
31 
32 #include <vector>
33 
34 namespace arm_compute
35 {
36 namespace experimental
37 {
38 namespace dynamic_fusion
39 {
40 class IGpuKernelComponent;
41 
42 /** A multi-input (tensors), multi-output (tensors) acyclic directed graph of gpu kernel components
43  * Its main purposes are:
44  *  - Perform "graph-level" optimizations like fusion of kernel components (not the fusion of operators)
45  *  - Automatically assign memory descriptions @ref MemoryDescriptor of all tensors based on graph topology
46  */
47 class GpuKernelComponentGraph
48 {
49 public:
50     /** Constructor
51      *
52      * @param[in] services @ref GpuComponentServices to be used by the graph
53      */
54     GpuKernelComponentGraph(GpuComponentServices *services);
55     /** Prevent instances of this class from being copy constructed */
56     GpuKernelComponentGraph(const GpuKernelComponentGraph &graph) = delete;
57     /** Prevent instances of this class from being copied */
58     GpuKernelComponentGraph &operator=(const GpuKernelComponentGraph &graph) = delete;
59     /** Allow instances of this class to be move constructed */
60     GpuKernelComponentGraph(GpuKernelComponentGraph &&graph) = default;
61     /** Allow instances of this class to be moved */
62     GpuKernelComponentGraph &operator=(GpuKernelComponentGraph &&graph) = default;
63     /** Create a new component and add it to the component graph
64      * Component id is automatically allocated
65      *
66      * @tparam T    Component type
67      * @tparam Args Component argument types
68      *
69      * @param[in] args Component arguments except for component id, which is auto-allocated
70      */
71     template <typename T, typename... Args>
add_new_component(Args &&...args)72     void add_new_component(Args &&... args)
73     {
74         auto                      comp           = _services->component_factory().create<T>(std::forward<Args>(args)...);
75         ArgumentPack<ITensorInfo> tensors        = comp->tensors();
76         const auto                src_tensor_ids = get_tensor_ids(tensors.get_const_src_tensors());
77         const auto                dst_tensor_ids = get_tensor_ids(tensors.get_const_dst_tensors());
78         bool                      success        = _dependency_graph.add_operator(comp->id(), src_tensor_ids, dst_tensor_ids);
79         ARM_COMPUTE_UNUSED(success);
80         ARM_COMPUTE_ERROR_ON(!success);
81         _components[comp->id()] = std::move(comp);
82         for(auto t : tensors.get_const_src_tensors())
83         {
84             _tensors[t->id()] = t;
85         }
86         for(auto t : tensors.get_const_dst_tensors())
87         {
88             _tensors[t->id()] = t;
89         }
90     }
91     /** Perform component fusion and serialize the graph into a stream of component groups
92      *
93      * @param[in] mem_map MemoryDescriptorMap for all the tensors in the component graph
94      *
95      * @return GpuKernelComponentStream
96      */
97     GpuKernelComponentStream fuse(const MemoryDescriptorMap &mem_map) const;
98 
99 private:
100     static std::vector<DependencyGraph::TensorId> get_tensor_ids(const std::vector<const ITensorInfo *> tensors);
101     GpuComponentServices *_services;
102     std::map<ComponentId, std::unique_ptr<IGpuKernelComponent>> _components;
103     std::map<ITensorInfo::Id, const ITensorInfo *>              _tensors;
104     DependencyGraph _dependency_graph{};
105 };
106 } // namespace dynamic_fusion
107 } // namespace experimental
108 } // namespace arm_compute
109 #endif /* SRC_DYNAMIC_FUSION_SKETCH_GPU_GPUKERNELCOMPONENTGRAPH */
110