Problem with GTX Titan OpenCL Linux library and/or driver

Accelerated Computing GPU-Accelerated Libraries
1

The latest Nvidia OpenCL driver/SDK (313.30) for Linux does not work for programs
that use “fork()”: Once the parent-process ever performed any OpenCL operation,
its son-process fails to create a context.

This problem only appeared with GTX Titan - everything worked well with older devices (i.e. GeForce GTX 480).

The trivial program below fails with the “Device Not Available” error (-2):

I don’t know whether this is the right forum, but I don’t know where else to turn
in order to have this problem fixed.

Thanks.

#include <cl.h>
#include <stdio.h>

int
main(na, argv)
char *argv;
{
cl_int res;
cl_platform_id platform;
cl_context_properties cps[3];
cl_context context;

if(clGetPlatformIDs(1, &platform, NULL) != CL_SUCCESS)
{
	fprintf(stderr, "Parent could not get platform\n");
	exit(1);
}
if(fork())
	exit(0);
sleep(2);
if(clGetPlatformIDs(1, &platform, NULL) != CL_SUCCESS)
{
	fprintf(stderr, "Son could not get platform\n");
	exit(1);
}
cps[0] = CL_CONTEXT_PLATFORM;
cps[1] = (cl_context_properties)platform;
cps[2] = 0;
context = clCreateContextFromType(cps, CL_DEVICE_TYPE_ALL, NULL, NULL, &res);

if(context)
	printf("OK\n");
else
	printf("clCreateContextFromType failed, error=%d\n", res);

}

2

Although this is an OpenCL question, CUDA has a similar issue. If you initialize a CUDA device, then fork, the device context will sometimes work in the child process, but isn’t supported. I suspect that the fact this code works on your GTX 480 is accidental.