1 /** 2 * Copyright 2019 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_DEVICE_GPU_CUDA_DRIVER_H_ 18 #define MINDSPORE_CCSRC_RUNTIME_DEVICE_GPU_CUDA_DRIVER_H_ 19 20 #include <cuda_runtime_api.h> 21 22 namespace mindspore { 23 namespace device { 24 namespace gpu { 25 typedef void *CudaDeviceStream; 26 typedef void *CudaDeviceEvent; 27 typedef void *HostMemPtr; 28 typedef void *DeviceMemPtr; 29 30 class CudaDriver { 31 public: 32 // Encapsulate the cuda APIs associated with memory operations 33 // such as malloc/free and memory copy from host to device and reverse. 34 static size_t AllocDeviceMem(size_t size, DeviceMemPtr *addr); 35 static bool FreeDeviceMem(const DeviceMemPtr &addr); 36 static size_t AllocHostPinnedMem(size_t size, void **addr); 37 static void FreeHostPinnedMem(void *addr); 38 39 static void CudaHostRegister(void *addr, size_t alloc_size); 40 41 static void CudaHostUnregister(void *addr); 42 43 static bool CopyHostMemToDevice(const DeviceMemPtr &dst, const void *src, size_t size); 44 static bool CopyDeviceMemToHost(const HostMemPtr &dst, const DeviceMemPtr &src, size_t size); 45 static bool CopyHostMemToHost(const DeviceMemPtr &dst, const void *src, size_t size); 46 47 static bool CopyHostMemToDeviceAsync(const DeviceMemPtr &dst, const void *src, size_t size, 48 CudaDeviceStream stream = 0); 49 static bool CopyDeviceMemToHostAsync(const HostMemPtr &dst, const void *src, size_t size, 50 CudaDeviceStream stream = 0); 51 static bool CopyDeviceMemToDeviceAsync(const DeviceMemPtr &dst, const void *src, size_t size, 52 CudaDeviceStream stream = 0); 53 54 static size_t total_mem_size(); 55 static size_t free_mem_size(); 56 57 // Encapsulate the cuda APIs associated with device resource 58 // such as Stream and Event. 59 static bool CreateStream(CudaDeviceStream *stream); 60 static bool CreateStreamWithPriority(CudaDeviceStream *stream, int priority); 61 static bool DestroyStream(const CudaDeviceStream &stream); 62 static bool SyncStream(const CudaDeviceStream &stream); 63 static bool QueryStream(const CudaDeviceStream &stream); 64 65 static bool ConstructEvent(CudaDeviceEvent *event, unsigned int flag = cudaEventDefault); 66 static bool DestroyEvent(const CudaDeviceEvent &event); 67 static bool RecordEvent(CudaDeviceEvent event, CudaDeviceStream stream = 0); 68 static bool SyncEvent(const CudaDeviceEvent &event); 69 static bool QueryEvent(const CudaDeviceEvent &event); 70 static bool ElapsedTime(float *cost_time, const CudaDeviceEvent &start, const CudaDeviceEvent &end); 71 72 // Encapsulate the cuda APIs associated with device management. 73 static int device_count(); 74 static bool SetDevice(int index); 75 76 private: 77 CudaDriver() = delete; 78 ~CudaDriver() = delete; 79 CudaDriver(const CudaDriver &) = delete; 80 CudaDriver &operator=(const CudaDriver &) = delete; 81 82 static constexpr float mem_malloc_retry_rate_{0.99}; 83 static constexpr size_t mem_malloc_retry_conut_max_{10}; 84 static constexpr size_t mem_malloc_align_size_{4}; 85 }; 86 } // namespace gpu 87 } // namespace device 88 } // namespace mindspore 89 90 #endif // MINDSPORE_CCSRC_RUNTIME_DEVICE_GPU_CUDA_DRIVER_H_ 91