CUDA driver API学习

之前一直都是用runtime API写CUDA代码，对context等不是很了解，打算简单看一下driver API。

context management:
感觉对性能调优可能会有帮助的一些API：
*

1. CUresult cuCtxGetSharedMemConfig (CUsharedconfig *pConfig):*
   获取当前context的shared memory bank width(4 bytes/ 8 bytes)；可用于检查是否存在bank conflict。

2. CUresult cuCtxPopCurrent (CUcontext *pctx):
   将当前CPU thread的context从当前CPU thread弹出；对应的context的句柄存放在pctx中，之后可被其他CPU thread设置为其current context(通过cuCtxPushCurrent())。

   1. CUresult cuCtxPushCurrent (CUcontext ctx):
      将ctx指定的context push到当前CPU thread对应的stack中，使其成为该thread的current context；上一个current context可以通过调用cuCtxDestroy()或者cuCtxPopCurrent()来获取；

   2. CUresult cuCtxSetCurrent (CUcontext ctx)：
      将ctx指定的context与当前CPU thread绑定；其效果为：若当前CPU thread存在一个context stack，则ctx会取代当前stack的top context成为新的top；若ctx为空，实际上相当于对stack的top进行pop操作。

   3. CUresult cuCtxSynchronize (void)；

module management
1. *CUresult cuModuleGetFunction (CUfunction *hfunc,
CUmodule hmod, const char name)
Returns in hfunc the handle of the function of name name located in module *hmod

1. CUresult cuModuleLoad (CUmodule *module, const char
   *fname);
   Takes a filename fname and loads the corresponding module module into the current context. The file should be a cubin file as output by nvcc, or a PTX file either as output by nvcc or handwritten.

memory management
1. *CUresult cuArrayCreate (CUarray *pHandle, const
CUDA_ARRAY_DESCRIPTOR pAllocateArray);
用于创建一个cuda array，pHandle为返回的array handle，pAllocateArray为指定的array的信息(包括width, height, format等)；

1. CUresult cuIpcGetMemHandle (CUipcMemHandle
   *pHandle, CUdeviceptr dptr);
   Takes a pointer to the base of an existing device memory allocation created with cuMemAlloc and exports it for use in another process. This is a lightweight operation and may be called multiple times on an allocation without adverse effects.
   用于IPC；(restricted to devices with support for unified addressing on Linux operating systems.)

2. CUresult cuIpcOpenMemHandle (CUdeviceptr *pdptr,
   CUipcMemHandle handle, unsigned int Flags);
   和上面一个结合使用：Maps memory exported from another process with cuIpcGetMemHandle into the current device address space，将handle对应的memory映射到自己的地址空间中并返回在本进程中的指针。

3. CUresult cuMemAlloc (CUdeviceptr *dptr, size_t bytesize);
   Allocates bytesize bytes of linear memory on the device and returns in *dptr a pointer to the allocated memory。

occupancy
1. CUresult cuOccupancyMaxActiveBlocksPerMultiprocessor
(int *numBlocks, CUfunction func, int blockSize, size_t
dynamicSMemSize);
用于计算指定function的occupancy(也即每个sm上可以容纳的最大的active block数目)；其中numBlocks为返回值；

1. CUresult cuOccupancyMaxPotentialBlockSize (int
   *minGridSize, int *blockSize, CUfunction func,
   CUoccupancyB2DSize blockSizeToDynamicSMemSize,
   size_t dynamicSMemSize, int blockSizeLimit)；
   用于suggest a launch configuration with reasonable occupancy.
   Returns in *blockSize a reasonable block size that can achieve the maximum occupancy.
   返回值包括：minGridSize(达到maximum occupancy所需的最小的gridSize); blockSize(最大的block size);
   输入参数包括：func(所需要计算的funciton的handle); dynamicSMemSize(dynamic shared memory usage,当每个block所需的dynamic shared memory size大小与block大小无关时，通过该值指定)； blockSizeToDynamicSMemSize(为一个自定义的函数，若每个block所需的dynamic shared memory的大小与block大小有关，需要通过该函数指出其关系)；

profile control

1. CUresult cuProfilerInitialize (const char *configFile,
   const char *outputFile, CUoutput_mode outputMode);
   用于初始化profile，其中configFile指定的文件中列出了需要进行profile的选项，outputFile指定了输出文件名，outputMode为输出模式；
   profile过程：
   cuProfilerInitialize();
   cuProfilerStart();
   // code to be profiled
   cuProfilerStop();
   …
   cuProfilerStart();
   // code to be profiled
   cuProfilerStop();

   1. CUresult cuProfilerStart (void)；
      CUresult cuProfilerStop (void);
      这两个函数的作用在于：可以对指定的piece of code进行profile；