C++ cuInit函数代码示例

void CudaModule::staticInit(void)
{
  if (s_inited) {
    return;
  }
  
  s_inited = true;
  s_available = false;

  checkError("cuInit", cuInit(0));
  s_available = true;
  
  s_device = selectDevice();
  printDeviceInfo(s_device);

  U32 flags = 0;
  flags |= CU_CTX_SCHED_SPIN; // use sync() if you want to yield
  
#if (CUDA_VERSION >= 2030)
  if (getDriverVersion() >= 23) 
  {
    // reduce launch overhead with large localmem
    flags |= CU_CTX_LMEM_RESIZE_TO_MAX; 
  }
#endif

  // OpenGL & window context must have been initialized !
  checkError("cuGLCtxCreate", cuGLCtxCreate( &s_context, flags, s_device));

  checkError("cuEventCreate", cuEventCreate(&s_startEvent, 0));
  checkError("cuEventCreate", cuEventCreate(&s_endEvent, 0));
}

int main(int argc, char* argv[])
{
    cuInit(0);
    int devs = 0;
    cuDeviceGetCount(&devs);
    assert(devs > 0);
    CUdevice dev;
    CUresult status;
    CUcontext ctx = 0;
    cuDeviceGet(&dev, 0);
    cuCtxCreate(&ctx, 0, dev);
    {
        size_t f = 0, t = 0;
        CUresult r = cuMemGetInfo(&f, &t);
        fprintf( stderr, "Do cuMemGetInfo: %d, %zu/%zu\n", r, f, t );
    }
    
    __init("\n");
 
    printf("\nPress any key to exit...");
    char c;
    scanf("%c", &c);
 
    return 0;
}

static bool
nvptx_init (void)
{
  CUresult r;
  int ndevs;

  if (instantiated_devices != 0)
    return true;

  r = cuInit (0);
  if (r != CUDA_SUCCESS)
    GOMP_PLUGIN_fatal ("cuInit error: %s", cuda_error (r));

  ptx_events = NULL;

  pthread_mutex_init (&ptx_event_lock, NULL);

  r = cuDeviceGetCount (&ndevs);
  if (r != CUDA_SUCCESS)
    GOMP_PLUGIN_fatal ("cuDeviceGetCount error: %s", cuda_error (r));

  ptx_devices = GOMP_PLUGIN_malloc_cleared (sizeof (struct ptx_device *)
					    * ndevs);

  return true;
}

	CUDADevice(DeviceInfo& info, Stats &stats, bool background_) : Device(stats)
	{
		background = background_;

		cuDevId = info.num;
		cuDevice = 0;
		cuContext = 0;

		/* intialize */
		if(cuda_error(cuInit(0)))
			return;

		/* setup device and context */
		if(cuda_error(cuDeviceGet(&cuDevice, cuDevId)))
			return;

		CUresult result;

		if(background) {
			result = cuCtxCreate(&cuContext, 0, cuDevice);
		}
		else {
			result = cuGLCtxCreate(&cuContext, 0, cuDevice);

			if(result != CUDA_SUCCESS) {
				result = cuCtxCreate(&cuContext, 0, cuDevice);
				background = true;
			}
		}

		if(cuda_error_(result, "cuCtxCreate"))
			return;

		cuda_pop_context();
	}

int mcopy_gpu_init(struct device_info *device_info)
{
	char fname[256];
	CUresult res;

	/* printf("madd_gpu_init called.\n"); */

	/* Initialization */
	if ((res = cuInit(0)) != CUDA_SUCCESS) {
		printf("cuInit failed: res = %lu\n", (unsigned long)res);
		return -1;
	}

	if ((res = cuDeviceGet(&device_info->dev, 0)) != CUDA_SUCCESS) {
		printf("cuDeviceGet failed: res = %lu\n", (unsigned long)res);
		return -1;
	}

	if ((res = cuCtxCreate(&device_info->context, 0, device_info->dev)) !=
	 CUDA_SUCCESS) {
		printf("cuCtxCreate failed: res = %lu\n", (unsigned long)res);
		return -1;
	}

	return 0;
}

	void CudaVideoRender::setVideoFile(LPCWSTR a_pVideoFilename)
	{
		m_sFileName=a_pVideoFilename;

		//char* video_file=new char[MAX_PATH];
		//WideCharToMultiByte( CP_ACP, 0, a_pVideoFilename, -1, video_file, MAX_PATH, NULL, NULL );

		// Find out the video size 
		m_bIsProgressive = loadVideoSource(a_pVideoFilename, 
			m_nVideoWidth, m_nVideoHeight, 
			m_nWindowWidth, m_nWindowHeight );
		//delete video_file;

		// Initialize CUDA
		//TODO: dont init cuda twice??
		cuInit(0);

		int bTCC = 0;
		// If we are using TCC driver, then always turn off interop
		if (bTCC) m_bInterop = false;

		// Initialize CUDA/D3D9 context and other video memory resources
		initCudaResources(m_bInterop, bTCC);

		m_pVideoSource->start();
	}

static int
nvptx_get_num_devices (void)
{
  int n;
  CUresult r;

  /* PR libgomp/65099: Currently, we only support offloading in 64-bit
     configurations.  */
  if (sizeof (void *) != 8)
    return 0;

  /* This function will be called before the plugin has been initialized in
     order to enumerate available devices, but CUDA API routines can't be used
     until cuInit has been called.  Just call it now (but don't yet do any
     further initialization).  */
  if (instantiated_devices == 0)
    {
      r = cuInit (0);
      /* This is not an error: e.g. we may have CUDA libraries installed but
         no devices available.  */
      if (r != CUDA_SUCCESS)
        return 0;
    }

  r = cuDeviceGetCount (&n);
  if (r!= CUDA_SUCCESS)
    GOMP_PLUGIN_fatal ("cuDeviceGetCount error: %s", cuda_error (r));

  return n;
}

CUresult cuda_driver_api_init(CUcontext *pctx, CUmodule *pmod, const char *f)
{
	CUresult res;
	CUdevice dev;

	res = cuInit(0);
	if (res != CUDA_SUCCESS) {
		printf("cuInit failed: res = %lu\n", (unsigned long)res);
		return res;
	}

	res = cuDeviceGet(&dev, 0);
	if (res != CUDA_SUCCESS) {
		printf("cuDeviceGet failed: res = %lu\n", (unsigned long)res);
		return res;
	}

	res = cuCtxCreate(pctx, 0, dev);
	if (res != CUDA_SUCCESS) {
		printf("cuCtxCreate failed: res = %lu\n", (unsigned long)res);
		return res;
	}
	
	res = cuModuleLoad(pmod, f);
	if (res != CUDA_SUCCESS) {
		printf("cuModuleLoad() failed\n");
		cuCtxDestroy(*pctx);
		return res;
	}

	return CUDA_SUCCESS;
}

int
init_cuda_context (void)
{
#ifdef _ENABLE_CUDA_
    CUresult curesult = CUDA_SUCCESS;
    CUdevice cuDevice;
    int local_rank, dev_count;
    int dev_id = 0;
    char * str;

    if ((str = getenv("LOCAL_RANK")) != NULL) {
        cudaGetDeviceCount(&dev_count);
        local_rank = atoi(str);
        dev_id = local_rank % dev_count;
    }

    curesult = cuInit(0);
    if (curesult != CUDA_SUCCESS) {
        return 1;
    }

    curesult = cuDeviceGet(&cuDevice, dev_id);
    if (curesult != CUDA_SUCCESS) {
        return 1;
    }

    curesult = cuCtxCreate(&cuContext, 0, cuDevice);
    if (curesult != CUDA_SUCCESS) {
        return 1;
    }
#endif
    return 0;
}

extern "C" void sarafft_init() {
  printf( "Cuda is about to be initialized!\n" );
  fflush ( stdout );
  char *OMPI_COMM_WORLD_LOCAL_RANK = getenv( "OMPI_COMM_WORLD_LOCAL_RANK" );
  if ( NULL == OMPI_COMM_WORLD_LOCAL_RANK ) {
    printf( "OMPI_COMM_WORLD_LOCAL_RANK not set!\n" );
    fflush ( stdout );
    exit( 80 );
  }
  int localRank = atoi( OMPI_COMM_WORLD_LOCAL_RANK );
  printf( "Local rank is %d\n", localRank );
  fflush ( stdout );
  if ( CUDA_SUCCESS != cuInit( 0 ) ) {
    printf( "cuInit failed!\n" );
    fflush ( stdout );
    exit( 81 );
  }
  CUdevice device;
  if ( CUDA_SUCCESS != cuDeviceGet( &device, localRank ) ) {
    printf( "cuDeviceGet failed!\n" );
    fflush ( stdout );
    exit( 82 );
  }
  if ( CUDA_SUCCESS != cuCtxCreate( &cuda_context, CU_CTX_SCHED_YIELD, device ) ) {
    printf( "cuCtxCreate failed!\n" );
    fflush ( stdout );
    exit( 83 );
  }
  printf( "Cuda was initialized successfully!\n" );
  fflush ( stdout );
}

main()
{
  /* initialize CUDA */
  CUresult res;
  res = cuInit(0);
  MY_CUDA_CHECK(res, "cuInit()");

  /* check GPU is setted or not */
  int device_num;
  res = cuDeviceGetCount(&device_num);
  MY_CUDA_CHECK(res, "cuDeviceGetCount()");

  if (device_num == 0) {        // no GPU is detected
    fprintf(stderr, "no CUDA capable GPU is detected...\n");
    exit(1);
  }

  printf("%d GPUs are detected\n", device_num);

  for (int i=0; i<device_num; i++)
    {
      /* get device handle of GPU No.i */
      CUdevice dev;
      res = cuDeviceGet(&dev, i);
      MY_CUDA_CHECK(res, "cuDeviceGet()");
      
      /* search compute capability of GPU No.i */
      int major=0, minor=0;
      res = cuDeviceComputeCapability(&major, &minor, dev);
      MY_CUDA_CHECK(res, "cuDeviceComputeCapability()");
     
      printf("GPU[%d] : actual compute capability is : %d%d\n", i, major, minor);
    }
}

int main() {

	int ngpu;
	CUdevice cuDevice;
	CUcontext cuContext;
	cuInit(0);
	cuDeviceGetCount(&ngpu);
	//printf("ngpu = %d\n", ngpu);

	size_t *totals, *frees ;
	totals = (size_t *) calloc (ngpu, sizeof(size_t));
	frees = (size_t *) calloc (ngpu, sizeof(size_t));

	int tid;
	omp_set_num_threads(ngpu);
	#pragma omp parallel private(tid, cuDevice, cuContext) shared(frees, totals)
	{
		tid = omp_get_thread_num();
		//printf("nthreads = %d, tid = %d\n", omp_get_num_threads(), tid);
		cuDeviceGet(&cuDevice, tid);
		cuCtxCreate(&cuContext, tid, cuDevice);
		cuMemGetInfo((size_t*)&frees[tid], (size_t*)&totals[tid]);
	}

	printf ("\ttotal\t\tfree\t\tused\n");
	for(int i=0; i<ngpu; i++) {
		printf("GPU %d\t%lu\t%lu\t%lu\n", i, (size_t)totals[i], (size_t)frees[i], (size_t)totals[i]-(size_t)frees[i]);
	}

	return 0;
}

Object cuda_over_map(Object self, int nparts, int *argcv,
        Object *argv, int flags) {
    CUresult error;
    cuInit(0);
    int deviceCount = 0;
    error = cuDeviceGetCount(&deviceCount);
    if (deviceCount == 0) {
        raiseError("No CUDA devices found");
    }
    CUdevice cuDevice;
    CUcontext cuContext;
    CUmodule cuModule;
    CUfunction cuFunc;
    error = cuDeviceGet(&cuDevice, 0);
    error = cuCtxCreate(&cuContext, 0, cuDevice);
    CUdeviceptr d_A;
    CUdeviceptr d_B;
    CUdeviceptr d_res;
    errcheck(cuModuleLoad(&cuModule, grcstring(argv[argcv[0]])));
    CUdeviceptr dps[argcv[0]];
    void *args[argcv[0]+2];
    int size = INT_MAX;
    for (int i=0; i<argcv[0]; i++) {
        struct CudaFloatArray *a = (struct CudaFloatArray *)argv[i];
        if (a->size < size)
            size = a->size;
        errcheck(cuMemAlloc(&dps[i], size * sizeof(float)));
        errcheck(cuMemcpyHtoD(dps[i], &a->data, size * sizeof(float)));
        args[i+1] = &dps[i];
    }
    struct CudaFloatArray *r =
        (struct CudaFloatArray *)(alloc_CudaFloatArray(size));
    int fsize = sizeof(float) * size;
    errcheck(cuMemAlloc(&d_res, fsize));
    errcheck(cuMemcpyHtoD(d_res, &r->data, fsize));
    args[0] = &d_res;
    args[argcv[0]+1] = &size;

    int threadsPerBlock = 256;
    int blocksPerGrid = (size + threadsPerBlock - 1) / threadsPerBlock;
    char name[256];
    strcpy(name, "block");
    strcat(name, grcstring(argv[argcv[0]]) + strlen("_cuda/"));
    for (int i=0; name[i] != 0; i++)
        if (name[i] == '.') {
            name[i] = 0;
            break;
        }
    errcheck(cuModuleGetFunction(&cuFunc, cuModule, name));
    errcheck(cuLaunchKernel(cuFunc, blocksPerGrid, 1, 1,
        threadsPerBlock, 1, 1,
        0,
        NULL, args, NULL));
    errcheck(cuMemcpyDtoH(&r->data, d_res, fsize));
    cuMemFree(d_res);
    for (int i=0; i<argcv[0]; i++)
        cuMemFree(dps[i]);
    return (Object)r;
}

device::device() {
	cuInit(0);
	cuDeviceGet(&cu_device, 0);
	checkCudaError("device::device Init");
	//cuCtxCreate(&cu_context, 0, cu_device);
	//checkCudaError("device::device Create context");
	device_name = props.name;
}

    void init(){
      if(!isNotInitialized)
        return;

      cuInit(0);

      isNotInitialized = false;
    }