/*
* Setup a list of devices with different device capabilities and add the
* metrics, which are specified by the user.
*
* @return a list of CUDA devices with different device capabilities
*/
static vt_cupti_device_t* vt_cuptievt_setupMetricList(void)
{
CUresult err;
int deviceCount, id;
vt_cupti_device_t *capList = NULL;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
/* CUDA initialization */
VT_CUDRV_CALL(cuInit(0), "cuInit");
/* How many GPGPU devices do we have? */
err = cuDeviceGetCount( &deviceCount );
VT_CUDRV_CALL(err, "cuDeviceGetCount");
if(deviceCount == 0){
vt_error_msg("[CUPTI Events] There is no device supporting CUDA available.");
}
/* create list with available compute capabilities */
for(id = 0; id < deviceCount; id++){
CUdevice cuDev;
vt_cupti_device_t *cuptiDev;
int dev_major, dev_minor;
err = cuDeviceGet(&cuDev, id);
VT_CUDRV_CALL(err, "cuDeviceGet");
err = cuDeviceComputeCapability(&dev_major, &dev_minor, cuDev);
VT_CUDRV_CALL(err, "cuDeviceComputeCapability");
/* check if device capability already listed */
cuptiDev = vt_cupti_checkMetricList(capList, dev_major, dev_minor);
if(cuptiDev == NULL){
/* allocate memory for device list entry */
cuptiDev = (vt_cupti_device_t *)malloc(sizeof(vt_cupti_device_t));
cuptiDev->dev_major = dev_major;
cuptiDev->dev_minor = dev_minor;
cuptiDev->cuDev = cuDev;
cuptiDev->vtcuptiEvtList = NULL;
cuptiDev->evtNum = 0;
cuptiDev->next = NULL;
/* prepend to list */
cuptiDev->next = capList;
capList = cuptiDev;
}
}
vt_cupti_fillMetricList(capList);
/* cleanup list: remove entries, which don't have metrics */
{
vt_cupti_device_t *curr = capList;
vt_cupti_device_t *last = capList;
while(curr != NULL){
vt_cupti_device_t *freeDev = curr;
curr = curr->next;
if(freeDev->evtNum == 0){
/* first element */
if(freeDev == capList){
capList = capList->next;
}else{
last->next = freeDev->next;
}
free(freeDev);
}else last = freeDev;
}
}
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
return capList;
}
static vt_cupti_dev_t* vt_cupti_setupMetricList(void)
{
CUresult err;
int deviceCount, id;
vt_cupti_dev_t *capList = NULL;
/* CUDA initialization */
err = cuInit( 0 );
if ( err != CUDA_SUCCESS ) {
printf( "Initialization of CUDA library failed.\n" );
exit( EXIT_FAILURE );
}
/* How many gpgpu devices do we have? */
err = cuDeviceGetCount( &deviceCount );
CHECK_CU_ERROR(err, "cuDeviceGetCount");
if(deviceCount == 0){
printf("[CUPTI]There is no device supporting CUDA.\n");
exit(EXIT_FAILURE);
}
/* create list with available compute capabilities */
for(id = 0; id < deviceCount; id++){
CUdevice cuDev;
vt_cupti_dev_t *cuptiDev;
int dev_major, dev_minor;
err = cuDeviceGet(&cuDev, id);
CHECK_CU_ERROR(err, "cuDeviceGet");
err = cuDeviceComputeCapability(&dev_major, &dev_minor, cuDev);
CHECK_CU_ERROR(err, "cuDeviceComputeCapability");
/* check if device capability already listed */
cuptiDev = vt_cupti_checkMetricList(capList, dev_major, dev_minor);
if(cuptiDev == NULL){
/* allocate memory for device list entry */
cuptiDev = (vt_cupti_dev_t *)malloc(sizeof(vt_cupti_dev_t));
cuptiDev->dev_major = dev_major;
cuptiDev->dev_minor = dev_minor;
cuptiDev->cuDev = cuDev;
cuptiDev->vtcuptiEvtList = NULL;
cuptiDev->evtNum = 0;
cuptiDev->next = NULL;
/* prepend to list */
cuptiDev->next = capList;
capList = cuptiDev;
}
}
vt_cupti_fillMetricList(capList);
/* cleanup list: remove entries, which don't have metrics */
{
vt_cupti_dev_t *curr = capList;
vt_cupti_dev_t *last = capList;
while(curr != NULL){
vt_cupti_dev_t *freeDev = curr;
curr = curr->next;
if(freeDev->evtNum == 0){
/* first element */
if(freeDev == capList){
capList = capList->next;
}else{
last->next = freeDev->next;
}
free(freeDev);
}else last = freeDev;
}
}
return capList;
}