xref: /aosp_15_r20/external/ComputeLibrary/arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2022 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 ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_GPUWORKLOADSKETCH
25 #define ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_GPUWORKLOADSKETCH
26 
27 #include "arm_compute/core/TensorInfo.h"
28 #include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h"
29 
30 #include <memory>
31 
32 namespace arm_compute
33 {
34 namespace experimental
35 {
36 namespace dynamic_fusion
37 {
38 /** A descriptor of a workload of operators
39  *
40  * A "workload" is a basic unit of computation to schedule and perform. It contains one or more operators that can be "fused" together.
41  * Note that a workload may still contain multiple kernels.
42  */
43 class GpuWorkloadSketch
44 {
45 public:
46     /** Global context used for the creation of a workload */
47     using Context = GpuWorkloadContext;
48     /** Internal opaque implementation */
49     class Implementation;
50 
51 public:
52     /** Constructor
53      *
54      * @param[in] context Gpu context for the creation of a workload
55      */
56     explicit GpuWorkloadSketch(GpuWorkloadContext *context);
57     /** Destructor */
58     ~GpuWorkloadSketch();
59     /** Get the implementation */
60     Implementation &implementation();
61     /** Get the implementation */
62     const Implementation &implementation() const;
63     /** Get the gpu workload context of this sketch */
64     const GpuWorkloadContext *gpu_context() const;
65     /** Create a @ref TensorInfo associated with the workload sketch.
66      *
67      * @return TensorInfo   Newly created tensor info
68      */
69     template <typename... Args>
create_tensor_info(Args &&...args)70     TensorInfo create_tensor_info(Args &&... args)
71     {
72         auto tensor_info = TensorInfo(std::forward<Args>(args)...);
73         register_new_tensor(tensor_info);
74         return tensor_info;
75     }
76     /** Create a default @ref TensorInfo associated with the workload sketch
77      * It is usually used by user input or output tensors
78      *
79      * @return TensorInfo   Newly created tensor info
80      */
81     TensorInfo create_tensor_info();
82 
83 private:
84     /** Register a new tensor by setting a new id to it and register its memory descriptor in the sketch
85      *
86      * @param[in,out] tensor_info @ref ITensorInfo that will be registered
87      */
88     void register_new_tensor(ITensorInfo &tensor_info);
89     std::unique_ptr<Implementation> _impl; /**< Internal opaque implementation*/
90 };
91 
92 } // namespace dynamic_fusion
93 } // namespace experimental
94 } // namespace arm_compute
95 #endif /* ARM_COMPUTE_DYNAMIC_FUSION_SKETCH_GPU_GPUWORKLOADSKETCH */
96