• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**
2  * Copyright 2019-2022 Huawei Technologies Co., Ltd
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16 #ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_GPUKERNELFACTORY_H_
17 #define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_GPUKERNELFACTORY_H_
18 
19 #include <functional>
20 #include <map>
21 #include <string>
22 #include <vector>
23 #include <utility>
24 #include <memory>
25 #include "plugin/device/gpu/kernel/gpu_kernel.h"
26 #include "plugin/device/gpu/hal/device/kernel_info_setter.h"
27 #include "kernel/kernel_build_info.h"
28 #include "kernel/common_utils.h"
29 
30 namespace mindspore {
31 namespace kernel {
32 using NativeGpuKernelModCreater = std::function<NativeGpuKernelMod *()>;
33 class NativeGpuKernelModFactory {
34  public:
35   ~NativeGpuKernelModFactory() = default;
36 
37   static NativeGpuKernelModFactory &GetInstance();
38 
39   void Register(const std::string &kernel_name, const KernelAttr &kernel_attr, NativeGpuKernelModCreater &&creator);
40 
41   NativeGpuKernelMod *Create(const std::string &kernel_name, const CNodePtr &apply_kernel);
42 
43   bool SearchRegistered(const std::string &kernel_name, const KernelBuildInfoPtr &kernel_info);
44 
45   std::string SupportedTypeList(const std::string &kernel_name);
46 
47   std::vector<KernelAttr> GetGpuSupportedList(const std::string &kernel_name);
48 
49   // Judge whether is registered kernel.
50   bool IsRegistered(const std::string &kernel_name);
51 
52   bool ReducePrecision(const std::string &kernel_name,
53                        std::shared_ptr<mindspore::kernel::KernelBuildInfo::KernelBuildInfoBuilder> builder);
54 
55   std::pair<std::vector<size_t>, TypeId> reduce_flag_{{}, kNumberTypeInt64};
56 
57  private:
58   NativeGpuKernelModFactory() = default;
59 
60   NativeGpuKernelModFactory(NativeGpuKernelModFactory const &);
61 
62   NativeGpuKernelModFactory &operator=(const NativeGpuKernelModFactory &);
63 
64   std::pair<bool, size_t> GpuKernelAttrCheck(const std::string &kernel_name, const KernelBuildInfo *kernel_info);
65   void CheckSM(const KernelBuildInfo *kernel_info, const size_t &input_index);
66   bool CheckIOParam(const std::string &kernel_name, const KernelBuildInfo *kernel_info,
67                     std::vector<std::pair<KernelAttr, NativeGpuKernelModCreater>> *iter_second, size_t attr_index);
68 
69   // Set output and input ref map to kernel info which will be used by graph compiler.
70   void SetRefMapToKernelInfo(const std::string &kernel_name, size_t index, device::KernelInfo *kernel_info);
71 
72   // map to maintain kernel and creator, KernelAttr object and creator must be registered as a pair.
73   std::map<std::string, std::vector<std::pair<KernelAttr, NativeGpuKernelModCreater>>> map_kernel_name_to_creater_;
74 };
75 
76 class GpuKernelRegister {
77  public:
GpuKernelRegister(const std::string & kernel_name,const KernelAttr & kernel_attr,NativeGpuKernelModCreater && creator)78   GpuKernelRegister(const std::string &kernel_name, const KernelAttr &kernel_attr,
79                     NativeGpuKernelModCreater &&creator) {
80     NativeGpuKernelModFactory::GetInstance().Register(kernel_name, kernel_attr, std::move(creator));
81   }
82   ~GpuKernelRegister() = default;
83 };
84 
85 #define UNIQUE_KERNEL_NAME(kernel) KERNEL_NAME(g_##kernel##_gpu_kernel_reg, __COUNTER__)
86 #define KERNEL_NAME(kernel, cnt) MERGE(kernel, cnt)
87 #define MERGE(kernel, cnt) kernel##cnt
88 
89 #define MS_REG_GPU_KERNEL(OPNAME, OPCLASS)                                                                   \
90   static_assert(std::is_base_of<NativeGpuKernelMod, OPCLASS>::value, " must be base of NativeGpuKernelMod"); \
91   static const GpuKernelRegister UNIQUE_KERNEL_NAME(OPNAME)(#OPNAME, KernelAttr(), []() { return new OPCLASS(); });
92 
93 // regular register of fixed accuracy kernels
94 #define MS_REG_GPU_KERNEL_REGULAR(OPNAME, ATTR, OPCLASS)                                                     \
95   static_assert(std::is_base_of<NativeGpuKernelMod, OPCLASS>::value, " must be base of NativeGpuKernelMod"); \
96   static const GpuKernelRegister UNIQUE_KERNEL_NAME(OPNAME)(#OPNAME, ATTR, []() { return new OPCLASS(); });
97 
98 // register of mixed accuracy kernels which use template and maintain one typename, ignore input num
99 #define MS_REG_GPU_KERNEL_SAME(OPNAME, ATTR, OPCLASS, T)                                                        \
100   static_assert(std::is_base_of<NativeGpuKernelMod, OPCLASS<T>>::value, " must be base of NativeGpuKernelMod"); \
101   static const GpuKernelRegister UNIQUE_KERNEL_NAME(OPNAME)(#OPNAME, ATTR, []() { return new OPCLASS<T>(); });
102 
103 // register of mixed accuracy kernels which use template and maintain one typename
104 #define MS_REG_GPU_KERNEL_ONE(OPNAME, ATTR, OPCLASS, T)                                                         \
105   static_assert(std::is_base_of<NativeGpuKernelMod, OPCLASS<T>>::value, " must be base of NativeGpuKernelMod"); \
106   static const GpuKernelRegister UNIQUE_KERNEL_NAME(OPNAME)(#OPNAME, ATTR, []() { return new OPCLASS<T>(); });
107 
108 // register of mixed accuracy kernels which use template and maintain two typename
109 #define MS_REG_GPU_KERNEL_TWO(OPNAME, ATTR, OPCLASS, T, S)                                                         \
110   static_assert(std::is_base_of<NativeGpuKernelMod, OPCLASS<T, S>>::value, " must be base of NativeGpuKernelMod"); \
111   static const GpuKernelRegister UNIQUE_KERNEL_NAME(OPNAME)(#OPNAME, ATTR, []() { return new OPCLASS<T, S>(); });
112 
113 // register of mixed accuracy kernels which use template and maintain three typename
114 #define MS_REG_GPU_KERNEL_THREE(OPNAME, ATTR, OPCLASS, T, S, G)                                                       \
115   static_assert(std::is_base_of<NativeGpuKernelMod, OPCLASS<T, S, G>>::value, " must be base of NativeGpuKernelMod"); \
116   static const GpuKernelRegister UNIQUE_KERNEL_NAME(OPNAME)(#OPNAME, ATTR, []() { return new OPCLASS<T, S, G>(); });
117 }  // namespace kernel
118 }  // namespace mindspore
119 #endif  // MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_GPU_GPUKERNELFACTORY_H_
120