• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /**
2  * Copyright 2021 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 
17 #ifndef MINDSPORE_CCSRC_RUNTIME_HARDWARE_DEVICE_CONTEXT_H_
18 #define MINDSPORE_CCSRC_RUNTIME_HARDWARE_DEVICE_CONTEXT_H_
19 
20 #include <string>
21 #include <vector>
22 #include <memory>
23 #include "runtime/device/device_address.h"
24 #include "runtime/device/bucket.h"
25 #include "backend/session/kernel_graph.h"
26 #include "backend/session/anf_runtime_algorithm.h"
27 
28 namespace mindspore {
29 namespace device {
30 using mindspore::kernel::AddressPtr;
31 using mindspore::kernel::KernelMod;
32 
33 const size_t kDeviceContextsNumOne = 1;
34 const size_t kDeviceContextsNumTwo = 2;
35 
36 struct DeviceContextKey {
37   // device type name, such as 'GPU' 'Ascend' 'CPU'.
38   std::string device_name_;
39   uint32_t device_id_{0};
40 
41   // Use the result of ToString() as key to look up DeviceContext
42   // in cache map which maintains created DeviceContext objects.
ToStringDeviceContextKey43   std::string ToString() const { return device_name_ + "_" + std::to_string(device_id_); }
44 };
45 
46 // DeviceContext is unified interface of interaction with device.
47 class DeviceContext {
48  public:
DeviceContext(const DeviceContextKey & device_context_key)49   explicit DeviceContext(const DeviceContextKey &device_context_key) : device_context_key_(device_context_key) {}
50   virtual ~DeviceContext() = default;
51 
52   // Initialize the device context.
53   virtual void Initialize() = 0;
54 
55   // Destroy device context and release device resource.
Destroy()56   virtual void Destroy() {}
57 
58   // Relevant function to allocate and free device memory.
59   virtual bool AllocateMemory(DeviceAddress *const &address, size_t size) const = 0;
60   virtual void FreeMemory(DeviceAddress *const &address) const = 0;
61 
62   // Allocate continuous device memory end to end into 'addr_list'.
63   // Communication operators may need continuous memory for input and output
64   // to optimize the communication performance.
AllocateContinuousMemory(const std::vector<DeviceAddressPtr> & addr_list,size_t total_size,const std::vector<size_t> & size_list)65   virtual bool AllocateContinuousMemory(const std::vector<DeviceAddressPtr> &addr_list, size_t total_size,
66                                         const std::vector<size_t> &size_list) const {
67     return true;
68   }
69 
70   // Create concrete device address according different device type.
71   virtual DeviceAddressPtr CreateDeviceAddress(void *const device_ptr, size_t device_size, const string &format,
72                                                TypeId type_id) const = 0;
73 
74   // Get device address type according different device type, such GPU, Ascend.
75   virtual DeviceAddressType GetDeviceAddressType() const = 0;
76 
77   // Optimize the kernel graph for graph mode.
OptimizeGraph(const KernelGraphPtr & graph)78   virtual void OptimizeGraph(const KernelGraphPtr &graph) const {}
79 
80   // Optimize the single operator graph for PyNative mode.
OptimizeSingleOpGraph(const KernelGraphPtr & graph)81   virtual void OptimizeSingleOpGraph(const KernelGraphPtr &graph) const {}
82 
83   // Select the matching backend kernels according to the data type and format of input and output for all
84   // execution operators, and set final device data type and format information for backend kernels, device
85   // data type and format which replace original data type and format will use for executing kernels.
86   virtual void SetOperatorInfo(const std::vector<CNodePtr> &nodes) const = 0;
87 
88   // Generate 'KernelMod' for all kernels and set 'KernelMod' into kernel,
89   // 'KernelMod' is real executive object of kernel.
90   virtual void CreateKernel(const std::vector<CNodePtr> &nodes) const = 0;
91 
92   // Adjust kernel graph before run graph, used in Graph Mode.
PreprocessBeforeRunGraph(const KernelGraphPtr & graph)93   virtual void PreprocessBeforeRunGraph(const KernelGraphPtr &graph) const {}
94   // Adjust single op kernel graph before run graph, used in PyNative Mode.
PreprocessBeforeRunSingleOpGraph(const KernelGraphPtr & graph)95   virtual void PreprocessBeforeRunSingleOpGraph(const KernelGraphPtr &graph) const {}
96 
97   // Infer kernel shape and update abstract info for dynamic shape kernel.
UpdateDynamicShape(const CNodePtr & kernel)98   virtual void UpdateDynamicShape(const CNodePtr &kernel) const { AnfAlgo::InferShape(kernel); }
99 
100   // Launch a kernel via 'KernelMod' of the kernel.
101   virtual bool LaunchKernel(const CNodePtr &kernel, const std::vector<AddressPtr> &inputs,
102                             const std::vector<AddressPtr> &workspace, const std::vector<AddressPtr> &outputs,
103                             bool is_dynamic_shape = false) const = 0;
104 
105   // Synchronize stream, device such as GPU and Ascend need stream to launch kernel asynchronously,
106   // using 'SyncStream' to block thread and wait for completing all tasks in stream.
107   // Devices that do not need stream could ignore the implementation of this function.
108   virtual bool SyncStream(size_t stream_id = 0) const { return true; }
109 
110   // Get device_context_key_ to obtain device name and device id.
device_context_key()111   const DeviceContextKey &device_context_key() const { return device_context_key_; }
112 
113   // Get rank id for distributed training.
GetRankID()114   virtual uint32_t GetRankID() const { return 0; }
115 
116   // Create and initialize bucket for every allreduce operator. Bucket is used in PyNative distributed training mode,
117   // one bucket handles all resource to launch and sync allreduce operator.
CreateBucket(uint32_t bucket_id,uint32_t bucket_size)118   virtual std::shared_ptr<Bucket> CreateBucket(uint32_t bucket_id, uint32_t bucket_size) const { return nullptr; }
119 
120  protected:
121   DeviceContextKey device_context_key_;
122 };
123 using DeviceContextPtr = std::shared_ptr<DeviceContext>;
124 }  // namespace device
125 }  // namespace mindspore
126 
127 #endif  // MINDSPORE_CCSRC_RUNTIME_HARDWARE_DEVICE_CONTEXT_H_
128