/*
* Initialize VampirTrace IDs and registers the finalize function.
* This may be done implicitly by vt_cupti_count().
*/
void vt_cupti_events_init()
{
if(!vt_cuptievt_initialized){ /* fast check without lock */
vt_cupti_init();
VT_CUPTI_LOCK();
if(!vt_cuptievt_initialized){
vt_cntl_msg(2, "[CUPTI Events] Initializing ... ");
/* create VampirTrace counter group ID only once */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
vt_cuptievt_rid_init = vt_def_region(VT_MASTER_THREAD, "vtcuptiHostThreadInit",
VT_NO_ID, VT_NO_LNO, VT_NO_LNO, "VT_CUPTI", VT_FUNCTION);
vt_cuptievt_cgid = vt_def_counter_group(VT_MASTER_THREAD, "CUPTI");
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
vt_cupti_events_sampling = (uint8_t)vt_env_cupti_sampling();
vtcuptievtCapList = vt_cuptievt_setupMetricList();
if(NULL == vtcuptievtCapList){
vt_cupti_events_enabled = 0;
}else{
/* register the finalize function of VampirTrace CUPTI to be called before
* the program exits */
atexit(vt_cupti_events_finalize);
}
vt_cuptievt_initialized = 1;
VT_CUPTI_UNLOCK();
}
}
}
/*
* Retrieve the VampirTrace CUPTI context from the CUDA context.
*
* @param cuCtx the CUDA context
* @param ptid the active VampirTrace thread id
*/
static vt_cupti_ctx_t* vt_cupti_getCtx(CUcontext cuCtx, uint32_t ptid)
{
vt_cupti_ctx_t *vtcuptiCtx = NULL;
/* check, if there has been at least one VampirTrace CUPTI context created */
if(vtcuptiCtxlist == NULL) vt_cupti_init();
/* check, if the current VampirTrace thread is enabled for GPU counters */
if((vt_gpu_prop[ptid] & VTGPU_NO_PC) == VTGPU_NO_PC)
return NULL;
/* check if CUDA context is listed (linear search) */
CUPTI_LOCK();
vtcuptiCtx = vtcuptiCtxlist;
while(vtcuptiCtx != NULL){
if(vtcuptiCtx->cuCtx == cuCtx){
CUPTI_UNLOCK();
/*vt_cntl_msg(1, "[CUPTI] host thread %d (MPI rank %d)", ptid, vt_my_trace);*/
return vtcuptiCtx;
}
vtcuptiCtx = vtcuptiCtx->next;
}
CUPTI_UNLOCK();
vt_cntl_msg(2, "[CUPTI] Context for VT tid %d unknown! Creating ... ", ptid);
vtcuptiCtx = vt_cupti_initCtx(ptid, cuCtx);
if(vtcuptiCtx != NULL){
vt_cupti_start(vtcuptiCtx);
}else{
/* no performance counters for this thread available */
vt_gpu_prop[ptid] |= VTGPU_NO_PC;
vt_cntl_msg(2, "[CUPTI] Could not initialize!");
}
return vtcuptiCtx;
}
/*
* Returns the VampirTrace CUPTI context for the CUDA context associated with
* the calling host thread.
*
* @param ptid the VampirTrace thread id of the calling host thread
*/
vt_cupti_ctx_t* vt_cupti_getCurrentContext(uint32_t ptid)
{
CUcontext cuCtx = NULL;
if(!vt_cupti_initialized) vt_cupti_init();
VT_SUSPEND_CUDA_TRACING(ptid);
# if (defined(CUDA_VERSION) && (CUDA_VERSION < 4000))
CHECK_CU_ERROR(cuCtxPopCurrent(&cuCtx), "cuCtxPopCurrent");
CHECK_CU_ERROR(cuCtxPushCurrent(cuCtx), "cuCtxPushCurrent");
# else
CHECK_CU_ERROR(cuCtxGetCurrent(&cuCtx), "cuCtxGetCurrent");
# endif
VT_RESUME_CUDA_TRACING(ptid);
if(cuCtx == NULL) {
vt_cntl_msg(2, "[CUPTI] No context is bound to the calling CPU thread", cuCtx);
return NULL;
}
return vt_cupti_getCtx(cuCtx, ptid);
}
