int pthread_create(pthread_t *thread, const pthread_attr_t *attr, void *(*start_routine) (void *), void *arg) {
int core;
int ret;
cpu_set_t mask;
CPU_ZERO(&mask);
ret = old_pthread_create(thread, attr, start_routine, arg);
if(!get_shm()->active)
return ret;
core = get_next_core();
if(!get_shm()->per_node) {
CPU_SET(core, &mask);
} else {
int i, node = numa_node_of_cpu(core);
struct bitmask * bmp = numa_allocate_cpumask();
numa_node_to_cpus(node, bmp);
for(i = 0; i < numa_num_configured_cpus(); i++) {
if(numa_bitmask_isbitset(bmp, i))
CPU_SET(i, &mask);
}
numa_free_cpumask(bmp);
}
old_pthread_setaffinity_np(*thread, sizeof(mask), &mask);
VERBOSE("-> Set affinity to %d\n", core);
return ret;
}
pid_t fork(void) {
pid_t ret;
// When a new process is forked, the refcounter must be incremented
pthread_mutex_lock(&get_shm()->pin_lock);
get_shm()->refcount++;
pthread_mutex_unlock(&get_shm()->pin_lock);
ret = old_fork();
if(ret > 0) {
set_affinity(ret, get_next_core());
}
return ret;
}
pid_t fork(void) {
pid_t ret;
// Increment refcount on fork to avoid parent dying before child and destroying the shm
__sync_fetch_and_add(&get_shm()->refcount, 1);
ret = old_fork();
if(ret > 0) {
set_affinity(ret, get_next_core());
} else if (ret < 0) {
// fork failed, decrement
__sync_fetch_and_sub(&get_shm()->refcount, 1);
}
return ret;
}
static int init_cuda(NvEncoder*enc)
{
NVENCSTATUS nvResult;
CUresult cuResult;
int devid;
cuResult=cuInit(0);
CHK_CUDA_ERR(cuResult);
cuResult=cuDeviceGetCount(&enc->devcnt);
CHK_CUDA_ERR(cuResult);
fprintf(stderr,"Device Count = %d\n",enc->devcnt);
devid=enc->encCfg.deviceID;//default is 0
if(devid==-1 || devid>=enc->devcnt){
devid=get_next_core();
fprintf(stderr,"\033[33mAuto Select DevID:= %d\n\033[0m",devid);
enc->encCfg.deviceID=devid;
}else{
fprintf(stderr,"\033[33mGiven DevID:= %d\n\033[0m",devid);
}
cuResult=cuDeviceGet(&enc->cuDevice,devid);
CHK_CUDA_ERR(cuResult);
cuResult=cuCtxCreate(&enc->cuContext,0,enc->cuDevice);
CHK_CUDA_ERR(cuResult);
return 0;
}