• 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_LITE_SRC_EXTENDRT_DELEGATE_TENSORRT_TENSORRT_ALLOCATOR_H_
18 #define MINDSPORE_LITE_SRC_EXTENDRT_DELEGATE_TENSORRT_TENSORRT_ALLOCATOR_H_
19 #include "src/extendrt/delegate/tensorrt/tensorrt_allocator.h"
20 #include <map>
21 #include <string>
22 #include <NvInfer.h>
23 #include "include/api/types.h"
24 #include "ir/tensor.h"
25 #include "src/extendrt/delegate/tensorrt/tensor_info.h"
26 
27 namespace mindspore::lite {
28 struct CudaTensorParam {
29   void *data = nullptr;
30   bool is_valid_mem = false;
31   size_t size = 0;
32 };
33 class TensorRTAllocator {
34  public:
35   TensorRTAllocator() = default;
36 
37   ~TensorRTAllocator() = default;
38 
39   void *MallocDeviceMem(const TensorInfo &host_tensor, size_t size);
40 
41   void *MallocDeviceMem(const std::string &name, size_t size, nvinfer1::DataType data_type);
42 
43   void *GetDevicePtr(const std::string &tensor_name);
44 
SetCudaStream(cudaStream_t stream)45   void SetCudaStream(cudaStream_t stream) { stream_ = stream; }
46 
47   std::map<std::string, CudaTensorParam> GetAllDevicePtr();
48 
49   int SyncMemInHostAndDevice(tensor::Tensor *host_tensor, const std::string &device_tensor_name, bool is_host2device,
50                              bool sync = true);
51 
52   int SyncMemInHostAndDevice(void *host_data, const std::string &device_tensor_name, size_t data_size,
53                              bool is_host2device, bool sync = true);
54   int SyncMemHostToDevice(const tensor::Tensor &host_tensor, const std::string &device_tensor_name, bool sync = true,
55                           size_t size = 0);
56   int SyncMemDeviceToHost(tensor::Tensor *host_tensor, const std::string &device_tensor_name, bool sync = true);
57   int SyncMemDeviceToHost(void *dst_data, size_t data_size, const std::string &device_tensor_name, bool sync = true);
58 
59   int ClearDeviceMem();
60 
61   void MarkMemValid(const std::string &name, bool isValid);
62 
63   bool GetMemIsValid(const std::string &name);
64 
65  private:
66   std::map<std::string, CudaTensorParam> cuda_tensor_map_;
67   cudaStream_t stream_;
68 };
69 }  // namespace mindspore::lite
70 #endif  // MINDSPORE_LITE_SRC_EXTENDRT_DELEGATE_TENSORRT_TENSORRT_ALLOCATOR_H_
71