/*
* De-initialize the VampirTrace CUPTI context without destroying it.
*
* @param vtcuptiCtx pointer to the VampirTrace CUPTI context
*/
static void vt_cupti_finish(vt_cupti_ctx_t *vtcuptiCtx)
{
CUptiResult cuptiErr = CUPTI_SUCCESS;
if(vtcuptiCtx == NULL || vt_gpu_debug) return;
/*uint64_t time = vt_pform_wtime();
vt_cupti_resetCounter(vtcuptiCtx, 0, &time);*/
/* stop CUPTI counter capturing */
vt_cupti_stop(vtcuptiCtx);
/* destroy all CUPTI event groups, which have been created */
{
vt_cupti_grp_t *vtcuptiGrp = vtcuptiCtx->vtGrpList;
while(vtcuptiGrp != NULL){
cuptiErr = cuptiEventGroupRemoveAllEvents(vtcuptiGrp->evtGrp);
CHECK_CUPTI_ERROR(cuptiErr, "cuptiEventGroupRemoveAllEvents");
cuptiErr = cuptiEventGroupDestroy(vtcuptiGrp->evtGrp);
CHECK_CUPTI_ERROR(cuptiErr, "cuptiEventGroupDestroy");
vtcuptiGrp = vtcuptiGrp->next;
}
}
}
/*
* Finalizes CUPTI device.
*
* @param ptid VampirTrace process/thread id
* @param cleanExit 1 to cleanup CUPTI event group, otherwise 0
*/
void vt_cuptievt_finalize_device(uint8_t cleanExit){
CUptiResult cuptiErr = CUPTI_SUCCESS;
vt_cupti_ctx_t *vtcuptiCtx = NULL;
vt_cntl_msg(2, "[CUPTI Events] Finalize device ... ");
{
CUcontext cuCtx;
#if (defined(CUDA_VERSION) && (CUDA_VERSION < 4000))
VT_CUDRV_CALL(cuCtxPopCurrent(&cuCtx), "cuCtxPopCurrent");
VT_CUDRV_CALL(cuCtxPushCurrent(cuCtx), "cuCtxPushCurrent");
#else
VT_CUDRV_CALL(cuCtxGetCurrent(&cuCtx), "cuCtxGetCurrent");
#endif
vtcuptiCtx = vt_cupti_removeCtx(&cuCtx);
if(vtcuptiCtx == NULL)
return;
}
if(vtcuptiCtx->events == NULL)
return;
if(cleanExit && vt_gpu_debug != 0){
/*uint64_t time = vt_pform_wtime();
vt_cupti_resetCounter(vtcuptiCtx, 0, &time);*/
/* stop CUPTI counter capturing */
vt_cuptievt_stop(vtcuptiCtx->events);
/* destroy all CUPTI event groups, which have been created */
{
vt_cupti_evtgrp_t *vtcuptiGrp = vtcuptiCtx->events->vtGrpList;
while(vtcuptiGrp != NULL){
cuptiErr = cuptiEventGroupRemoveAllEvents(vtcuptiGrp->evtGrp);
VT_CUPTI_CALL(cuptiErr, "cuptiEventGroupRemoveAllEvents");
cuptiErr = cuptiEventGroupDestroy(vtcuptiGrp->evtGrp);
VT_CUPTI_CALL(cuptiErr, "cuptiEventGroupDestroy");
vtcuptiGrp = vtcuptiGrp->next;
}
}
}
/* free VampirTrace CUPTI event context */
vt_cuptievt_freeEventCtx(vtcuptiCtx->events);
}
/*
* Finalizes CUPTI device.
*
* @param cleanExit 1 to cleanup CUPTI event group, otherwise 0
*/
void vt_cupti_finalize_device(uint32_t ptid, uint8_t cleanExit){
CUptiResult cuptiErr = CUPTI_SUCCESS;
vt_cupti_ctx_t *vtcuptiCtx = NULL;
vt_cntl_msg(2, "[CUPTI] Finalize device ... ");
{
CUcontext cuCtx = NULL;
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);
vtcuptiCtx = vt_cupti_takeCtxFromList(cuCtx);
if(vtcuptiCtx == NULL) return;
}
if(cleanExit && vt_gpu_debug != 0){
/*uint64_t time = vt_pform_wtime();
vt_cupti_resetCounter(vtcuptiCtx, 0, &time);*/
/* stop CUPTI counter capturing */
vt_cupti_stop(vtcuptiCtx);
/* destroy all CUPTI event groups, which have been created */
{
vt_cupti_grp_t *vtcuptiGrp = vtcuptiCtx->vtGrpList;
while(vtcuptiGrp != NULL){
cuptiErr = cuptiEventGroupRemoveAllEvents(vtcuptiGrp->evtGrp);
CHECK_CUPTI_ERROR(cuptiErr, "cuptiEventGroupRemoveAllEvents");
cuptiErr = cuptiEventGroupDestroy(vtcuptiGrp->evtGrp);
CHECK_CUPTI_ERROR(cuptiErr, "cuptiEventGroupDestroy");
vtcuptiGrp = vtcuptiGrp->next;
}
}
}
/* free VampirTrace CUPTI context */
vt_cupti_freeCtx(vtcuptiCtx);
}
void vt_cupti_events_finalizeContext(vt_cupti_ctx_t *vtCtx)
{
uint64_t time = vt_pform_wtime();
vt_cupti_strm_t *curStrm = NULL;
vt_cupti_evtgrp_t *vtcuptiGrp = NULL;
if(vtCtx == NULL || vtCtx->events == NULL)
return;
/* These CUPTI calls may fail, as CUPTI has implicitly destroyed something */
if(vt_gpu_debug == 0){
curStrm = vtCtx->strmList;
/* for all streams of this context */
while(curStrm != NULL){
/* ensure increasing time stamps */
if(time < curStrm->vtLastTime){
curStrm = curStrm->next;
continue;
}
vt_cuptievt_resetCounter(vtCtx->events, curStrm->vtThrdID, &time);
curStrm = curStrm->next;
}
/* stop CUPTI counter capturing */
vt_cuptievt_stop(vtCtx->events);
/* destroy all CUPTI event groups, which have been created */
vtcuptiGrp = vtCtx->events->vtGrpList;
while(vtcuptiGrp != NULL){
VT_CUPTI_CALL(cuptiEventGroupRemoveAllEvents(vtcuptiGrp->evtGrp),
"cuptiEventGroupRemoveAllEvents");
VT_CUPTI_CALL(cuptiEventGroupDestroy(vtcuptiGrp->evtGrp),
"cuptiEventGroupDestroy");
vtcuptiGrp = vtcuptiGrp->next;
}
}else{
/* set at least the VampirTrace counter to zero */
curStrm = vtCtx->strmList;
/* for all streams of this context */
while(curStrm != NULL){
/* ensure increasing time stamps */
if(time < curStrm->vtLastTime){
curStrm = curStrm->next;
continue;
}
vtcuptiGrp = vtCtx->events->vtGrpList;
while(vtcuptiGrp != NULL){
size_t i;
for(i = 0; i < vtcuptiGrp->evtNum; i++){
vt_count(curStrm->vtThrdID, &time, *(vtcuptiGrp->vtCIDs+i), 0);
}
vtcuptiGrp = vtcuptiGrp->next;
}
curStrm = curStrm->next;
}
}
/* free previously allocated memory */
vt_cuptievt_freeEventCtx(vtCtx->events);
}
