Context creation on C2050 is 1-2 seconds? Why so long?

Accelerated Computing CUDA CUDA Programming and Performance
1

On a 64-bit RedHat Linux machine with a single C2050 card I’m seeing a 1-2 second overhead for creating a context. Is this sort of long delay expected?

2

So surprising,
On Win Vista with a GTX 295, it takes 0.124s for the following block :

DWORD tmp1 = GetTickCount();
// OpenCL Objects
cl_device_id device;
cl_context context;
cl_command_queue queue;
cl_program program;
cl_kernel kernel;
cl_mem buffer;	

// Initialisation OpenCL
cl_platform_id cpPlatform;
clGetPlatformIDs(1, &cpPlatform, NULL);
clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &device, NULL); 
context = clCreateContext(NULL, 1, &device, NULL, NULL, NULL);
queue = clCreateCommandQueue(context, device, (cl_command_queue_properties)0, NULL);	

DWORD tmp2 = GetTickCount();
float delta =(tmp2-tmp1)*1.0/1000 ;
std::cout << "Creation contexte :"<<delta<<std::endl;

Regards, Emmanuel.

3

So surprising,
On Win Vista with a GTX 295, it takes 0.124s for the following block :

DWORD tmp1 = GetTickCount();
// OpenCL Objects
cl_device_id device;
cl_context context;
cl_command_queue queue;
cl_program program;
cl_kernel kernel;
cl_mem buffer;	

// Initialisation OpenCL
cl_platform_id cpPlatform;
clGetPlatformIDs(1, &cpPlatform, NULL);
clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU, 1, &device, NULL); 
context = clCreateContext(NULL, 1, &device, NULL, NULL, NULL);
queue = clCreateCommandQueue(context, device, (cl_command_queue_properties)0, NULL);	

DWORD tmp2 = GetTickCount();
float delta =(tmp2-tmp1)*1.0/1000 ;
std::cout << "Creation contexte :"<<delta<<std::endl;

Regards, Emmanuel.

4

Just upgraded to the OpenCL 1.1 beta driver (devdriver_3.1_linux_64_258.19_opencl1.1.run) and see the same 1-2 second overhead. (The overhead is usually 1.3 seconds to be more precise.)

Anyone seen something similar, or have any wisdom to figure out why I’m seeing such a large overhead?

Thanks,
Brian

5

Just upgraded to the OpenCL 1.1 beta driver (devdriver_3.1_linux_64_258.19_opencl1.1.run) and see the same 1-2 second overhead. (The overhead is usually 1.3 seconds to be more precise.)

Anyone seen something similar, or have any wisdom to figure out why I’m seeing such a large overhead?

Thanks,
Brian