PyOpenCL with multiprocessing INVALID_DEVICE - opencl

Question

My use case is a main process that compiles a bunch of Open CL kernels. Later in the program several subprocesses are forked and will execute one or more of the kernels. For some reason the subprocesses have device errors.
I have determined the problem has nothing to do with compilation and have reproduced it with the following simple script:
import multiprocessing
import pyopencl as cl
def printme():
platforms = cl.get_platforms()
for p in platforms:
print 75*'!'
print p
print 75*':'
printme()
p = multiprocessing.Process(target = printme)
p.start()
p.join()
Seems to be tied to the NVIDIA OpenCL implementation and not related to PyOpenCL as I initially thought. The same problem occurs with the example below.
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <CL/cl.h>
#define CHECK(A) if ((status = A) != 0) { \
fprintf(stderr, "failed status %d at line %d\n", status, __LINE__); \
exit(1); \
}
static void
runit() {
int i;
cl_int status;
cl_platform_id *platforms;
cl_uint num_platforms;
cl_uint num_devices;
CHECK(clGetPlatformIDs(0, NULL, &num_platforms));
if (num_platforms == 0) {
fprintf(stderr, "no platforms\n");
exit(1);
}
platforms = malloc(sizeof(cl_platform_id)*num_platforms);
CHECK(clGetPlatformIDs(num_platforms, platforms, NULL));
CHECK(clGetDeviceIDs(platforms[0], CL_DEVICE_TYPE_ALL, 0, NULL,
&num_devices));
free(platforms);
}
int main(void) {
runit();
if (fork() == 0) {
runit();
}
else {
wait(NULL);
}
return 0;
}
Exits with failed status = -33 (Invalid Device). Seems that there is some stored state inside the implementation. Note this is running from an NVidia driver.