/*
* Remove a context from the global context list and return it.
*
* @param cuCtx pointer to the CUDA context
* @return the VampirTrace CUPTI context, which has been removed
*/
vt_cupti_ctx_t* vt_cupti_removeCtx(CUcontext *cuCtx)
{
vt_cupti_ctx_t *currCtx = NULL;
vt_cupti_ctx_t *lastCtx = NULL;
VT_CUPTI_LOCK();
currCtx = vt_cupti_ctxList;
lastCtx = vt_cupti_ctxList;
while(currCtx != NULL){
if(currCtx->cuCtx == *cuCtx){
/* if first element in list */
if(currCtx == vt_cupti_ctxList){
vt_cupti_ctxList = vt_cupti_ctxList->next;
}else{
lastCtx->next = currCtx->next;
}
VT_CUPTI_UNLOCK();
return currCtx;
}
lastCtx = currCtx;
currCtx = currCtx->next;
}
VT_CUPTI_UNLOCK();
vt_cntl_msg(2, "[CUPTI] Could not remove context (CUDA Context not found)!");
return NULL;
}
/*
* Finalize the CUPTI common interface.
* - free the VampirTrace CUPTI context list
*/
void vt_cupti_finalize()
{
if(!vt_cupti_finalized && vt_cupti_initialized){
VT_CUPTI_LOCK();
if(!vt_cupti_finalized && vt_cupti_initialized){
vt_cntl_msg(2, "[CUPTI] Finalizing ... ");
/* free VampirTrace CUPTI context structures */
while(vt_cupti_ctxList != NULL){
vt_cupti_ctx_t *tmp = vt_cupti_ctxList;
vt_cupti_ctxList = vt_cupti_ctxList->next;
vt_cupti_finalizeCtx(tmp);
tmp = NULL;
}
vt_cupti_finalized = 1;
VT_CUPTI_UNLOCK();
#if (defined(VT_MT) || defined (VT_HYB))
VTTHRD_LOCK_ENV();
VTThrd_deleteMutex(&VTThrdMutexCupti);
VTTHRD_UNLOCK_ENV();
#endif /* VT_MT || VT_HYB */
}
}
}
/*
* Prepend the given VampirTrace CUPTI context to the global context list.
*
* @param vtCtx pointer to the VampirTrace CUPTI context to be prepended
*/
void vt_cupti_prependCtx(vt_cupti_ctx_t *vtCtx)
{
VT_CUPTI_LOCK();
vtCtx->next = vt_cupti_ctxList;
vt_cupti_ctxList = vtCtx;
VT_CUPTI_UNLOCK();
}
/*
* Get a VampirTrace CUPTI context by CUDA context
*
* @param cuCtx the CUDA context
*
* @return VampirTrace CUPTI context
*/
vt_cupti_ctx_t* vt_cupti_getCtx(CUcontext cuCtx)
{
vt_cupti_ctx_t* vtCtx = NULL;
/* lookup context */
VT_CUPTI_LOCK();
vtCtx = vt_cupti_ctxList;
while(vtCtx != NULL){
if(vtCtx->cuCtx == cuCtx){
/* workaround to set the correct device number
if(vtCtx->devID == VT_CUPTI_NO_DEVICE_ID){
CUdevice cuDev;
if(CUDA_SUCCESS != cuCtxGetDevice(&cuDev)){
vt_warning("[CUPTI] Could not get CUdevice from context");
}
vtCtx->devID = (uint32_t)cuDev;
vtCtx->cuDev = cuDev;
}*/
VT_CUPTI_UNLOCK();
return vtCtx;
}
vtCtx = vtCtx->next;
}
VT_CUPTI_UNLOCK();
return NULL;
}
void vt_cuptiact_markStreamAsDestroyed(CUcontext cuCtx, uint32_t strmID)
{
vt_cupti_ctx_t *vtCtx = NULL;
vt_cuptiact_strm_t *currStrm = NULL;
VT_CUPTI_LOCK();
if(cuCtx == NULL){
vt_warning("[CUPTI Activity] No CUDA context given in "
"vt_cuptiact_markStreamAsDestroyed()!");
VT_CUPTI_UNLOCK();
return;
}
vtCtx = vt_cupti_getCtxNoLock(cuCtx);
if(vtCtx == NULL){
vt_warning("[CUPTI Activity] No context found in "
"vt_cuptiact_markStreamAsDestroyed()!");
VT_CUPTI_UNLOCK();
return;
}
currStrm = vtCtx->activity->strmList;
while(currStrm != NULL){
if(currStrm->strmID == strmID){
currStrm->destroyed = 1;
VT_CUPTI_UNLOCK();
return;
}
currStrm = currStrm->next;
}
VT_CUPTI_UNLOCK();
}
void vt_cupti_activity_finalize()
{
if(!vt_cuptiact_finalized && vt_cuptiact_initialized){
VT_CUPTI_LOCK();
if(!vt_cuptiact_finalized && vt_cuptiact_initialized){
vt_cupti_ctx_t *vtCtx = vt_cupti_ctxList;
vt_cntl_msg(2, "[CUPTI Activity] Finalizing ... ");
while(vtCtx != NULL){
/* write buffered activities, which have not been dumped yet */
vt_cuptiact_flushCtxActivities(vtCtx);
/* free the context */
vt_cuptiact_destroyContext(vtCtx->activity);
vtCtx->activity = NULL;
/* set pointer to next context */
vtCtx = vtCtx->next;
}
vt_cuptiact_finalized = 1;
VT_CUPTI_UNLOCK();
}
}
}
/*
* Initialize the CUPTI events data of the given VampirTrace CUPTI context.
*
* @param vtCtx pointer to the VampirTrace CUPTI context
*/
void vt_cupti_events_initContext(vt_cupti_ctx_t *vtcuptiCtx)
{
vt_cupti_events_t *vtcuptiEvtCtx = NULL;
vt_cntl_msg(2, "[CUPTI Events] Initializing VampirTrace CUPTI events context");
/* get a pointer to eventIDArray */
{
CUresult cuErr = CUDA_SUCCESS;
int dev_major, dev_minor;
vt_cupti_device_t *cuptiDev;
/* TODO: do not trace this driver API function call */
cuErr = cuDeviceComputeCapability(&dev_major, &dev_minor, vtcuptiCtx->cuDev);
VT_CUDRV_CALL(cuErr, "cuDeviceComputeCapability");
/* check if device capability already listed */
VT_CUPTI_LOCK();
cuptiDev = vtcuptievtCapList;
VT_CUPTI_UNLOCK();
cuptiDev = vt_cupti_checkMetricList(cuptiDev, dev_major, dev_minor);
if(cuptiDev){
/* allocate the VampirTrace CUPTI events context */
vtcuptiEvtCtx = (vt_cupti_events_t *)malloc(sizeof(vt_cupti_events_t));
if(vtcuptiEvtCtx == NULL)
vt_error_msg("[CUPTI Events] malloc(sizeof(vt_cupti_events_t)) failed!");
vtcuptiEvtCtx->vtDevCap = cuptiDev;
vtcuptiEvtCtx->vtGrpList = NULL;
vtcuptiEvtCtx->counterData = NULL;
vtcuptiEvtCtx->cuptiEvtIDs = NULL;
vtcuptiCtx->events = vtcuptiEvtCtx;
}else{
return;
}
}
/* create and add the VampirTrace CUPTI groups to the context */
vt_cupti_addEvtGrpsToCtx(vtcuptiCtx);
/* allocate memory for CUPTI counter reads */
{
size_t allocSize = vtcuptiEvtCtx->vtGrpList->evtNum;
vtcuptiEvtCtx->counterData =
(uint64_t *)malloc(allocSize*sizeof(uint64_t));
vtcuptiEvtCtx->cuptiEvtIDs =
(CUpti_EventID *)malloc(allocSize*sizeof(CUpti_EventID));
}
vt_cuptievt_start(vtcuptiEvtCtx);
}
/*
* Finalizes the VampirTrace CUPTI events interface.
*/
void vt_cupti_events_finalize()
{
if(!vt_cuptievt_finalized && vt_cuptievt_initialized){ /* fast check without lock */
VT_CUPTI_LOCK();
if(!vt_cuptievt_finalized && vt_cuptievt_initialized){
vt_cupti_ctx_t *vtcuptiCtxList = vt_cupti_ctxList;
/* needed because of the atexit in vt_cupti_events_init() */
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
vt_cntl_msg(2, "[CUPTI Events] Finalizing ...");
/* free VampirTrace CUPTI events context structures */
while(vtcuptiCtxList != NULL){
if(vtcuptiCtxList->events != NULL){
vt_cupti_events_finalizeContext(vtcuptiCtxList);
free(vtcuptiCtxList->events);
vtcuptiCtxList->events = NULL;
}
vtcuptiCtxList = vtcuptiCtxList->next;
}
/* free capability metric list */
while(vtcuptievtCapList != NULL){
vt_cupti_device_t *tmp = vtcuptievtCapList;
vtcuptievtCapList = vtcuptievtCapList->next;
/* free VampirTrace CUPTI events */
while(tmp->vtcuptiEvtList != NULL){
vt_cupti_evtctr_t *tmpEvt = tmp->vtcuptiEvtList;
tmp->vtcuptiEvtList = tmp->vtcuptiEvtList->next;
free(tmpEvt);
tmpEvt = NULL;
}
free(tmp);
tmp = NULL;
}
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
vt_cuptievt_finalized = 1;
VT_CUPTI_UNLOCK();
}
}
}
void vt_cupti_init()
{
if(!vt_cupti_initialized){
#if (defined(VT_MT) || defined(VT_HYB))
VTThrd_createMutex(&VTThrdMutexCupti);
#endif
VT_CUPTI_LOCK();
if(!vt_cupti_initialized){
vt_cntl_msg(2, "[CUPTI] Initializing ... ");
/* register the finalize function of VampirTrace CUPTI to be called before
* the program exits */
atexit(vt_cupti_finalize);
vt_cupti_initialized = 1;
VT_CUPTI_UNLOCK();
}
}
}
void vt_cuptiact_enableConcurrentKernel(vt_cupti_ctx_t* vtCtx)
{
/*
* Disable collection of kernels for the given CUDA context.
* !!! does not work yet !!!
VT_CUPTI_CALL(cuptiActivityDisableContext(cuCtx, CUPTI_ACTIVITY_KIND_KERNEL),
"cuptiActivityDisableContext");*
* flush the already buffered activities for this CUDA context *
vt_cuptiact_flushCtxActivities(cuCtx);
* Enable collection of kernels for the given CUDA context
VT_CUPTI_CALL(cuptiActivityEnableContext(cuCtx, CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL),
"cuptiActivityEnableContext");*/
if((vt_gpu_config & VT_GPU_TRACE_CONCURRENT_KERNEL)
!= VT_GPU_TRACE_CONCURRENT_KERNEL){
vt_cntl_msg(2, "[CUPTI Activity] Enable concurrent kernel tracing.");
/*
* Disable normal (lower overhead) kernel tracing.
*/
VT_CUPTI_CALL(cuptiActivityDisable(CUPTI_ACTIVITY_KIND_KERNEL),
"cuptiActivityDisable");
/*
* Flush the already buffered activities for this CUDA context.
*/
VT_CUPTI_LOCK();
vt_cuptiact_flushCtxActivities(vtCtx);
VT_CUPTI_UNLOCK();
/*
* Enable concurrent kernel tracing (higher overhead).
*/
VT_CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL),
"cuptiActivityEnable");
vt_gpu_config |= VT_GPU_TRACE_CONCURRENT_KERNEL;
}
}
/*
* 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();
}
}
}
/*
* Increases the "Allocated CUDA memory" counter.
*
* @param ctxUID CUDA context identifier (@see CUPTI callback info)
* @param devPtr pointer to the allocated memory (needed for vtcudaFree())
* @param size the number of bytes allocated
*/
void vt_cuptiact_writeMalloc(uint32_t ctxID, CUcontext cuCtx,
void *devPtr, size_t size)
{
uint64_t vtTime;
vt_cupti_ctx_t* vtCtx = NULL;
vt_cupti_activity_t *vtcuptiActivity = NULL;
vt_cupti_gpumem_t *vtMalloc = NULL;
if(devPtr == NULL) return;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
vtMalloc = (vt_cupti_gpumem_t*)malloc(sizeof(vt_cupti_gpumem_t));
vtMalloc->memPtr = devPtr;
vtMalloc->size = size;
/* check for VampirTrace CUPTI context */
vtCtx = vt_cupti_getCtx(cuCtx);
if(vtCtx == NULL){
vtCtx = vt_cupti_createCtx(cuCtx, VT_CUPTI_NO_CUDA_DEVICE, ctxID, VT_CUPTI_NO_DEVICE_ID);
vt_cupti_prependCtx(vtCtx);
}
/* check for VampirTrace CUPTI activity context */
if(vtCtx->activity == NULL){
vtCtx->activity = vt_cuptiact_createCtxActivity(cuCtx);
}
vtcuptiActivity = vtCtx->activity;
/* lock the work on the context */
VT_CUPTI_LOCK();
/* flush activity buffer */
vt_cuptiact_flushCtxActivities(vtCtx);
/* add malloc entry to list */
vtMalloc->next = vtcuptiActivity->gpuMemList;
vtcuptiActivity->gpuMemList = vtMalloc;
/* increase allocated memory counter */
vtcuptiActivity->gpuMemAllocated += size;
/* check if first CUDA stream is available */
if(vtcuptiActivity->strmList == NULL){
if(vt_gpu_init_time < vt_start_time)
vt_gpu_init_time = vt_start_time;
vtcuptiActivity->strmList = vt_cuptiact_createStream(vtCtx, vtcuptiActivity->defaultStrmID);
vt_count(vtcuptiActivity->strmList->vtThrdID, &vt_gpu_init_time, vt_gpu_cid_memusage, 0);
}
VT_CUPTI_UNLOCK();
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
/* write counter value */
vtTime = vt_pform_wtime();
vt_count(vtcuptiActivity->strmList->vtThrdID, &vtTime, vt_gpu_cid_memusage,
(uint64_t)(vtcuptiActivity->gpuMemAllocated));
}
/*
* Decreases the "Allocated CUDA memory" counter.
*
* @param ctxUID CUDA context identifier (@see CUPTI callback info)
* @param devPtr pointer to the allocated memory
*/
void vt_cuptiact_writeFree(uint32_t ctxID, CUcontext cuCtx, void *devPtr)
{
uint64_t vtTime;
vt_cupti_ctx_t* vtCtx = NULL;
vt_cupti_activity_t *vtcuptiActivity = NULL;
vt_cupti_gpumem_t *curMalloc = NULL;
vt_cupti_gpumem_t *lastMalloc = NULL;
if(devPtr == NULL) return;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
/* check for VampirTrace CUPTI context */
vtCtx = vt_cupti_getCtx(cuCtx);
if(vtCtx == NULL){
vtCtx = vt_cupti_createCtx(cuCtx, VT_CUPTI_NO_CUDA_DEVICE, ctxID, VT_CUPTI_NO_DEVICE_ID);
vt_cupti_prependCtx(vtCtx);
}
/* check for VampirTrace CUPTI activity context */
if(vtCtx->activity == NULL){
vtCtx->activity = vt_cuptiact_createCtxActivity(cuCtx);
}
vtcuptiActivity = vtCtx->activity;
VT_CUPTI_LOCK();
/* flush activity buffer */
vt_cuptiact_flushCtxActivities(vtCtx);
curMalloc = vtcuptiActivity->gpuMemList;
lastMalloc = curMalloc;
/* lookup the CUDA malloc entry by its memory pointer */
while(curMalloc != NULL){
if(devPtr == curMalloc->memPtr){
/* decrease allocated counter value and write it */
vtTime = vt_pform_wtime();
vtcuptiActivity->gpuMemAllocated -= curMalloc->size;
vt_count(vtcuptiActivity->strmList->vtThrdID, &vtTime, vt_gpu_cid_memusage,
(uint64_t)(vtcuptiActivity->gpuMemAllocated));
/* set pointer over current element to next one */
lastMalloc->next = curMalloc->next;
/* if current element is the first list entry, set the list entry */
if(curMalloc == vtcuptiActivity->gpuMemList){
vtcuptiActivity->gpuMemList = curMalloc->next;
}
/* free VT memory of CUDA malloc */
curMalloc->next = NULL;
free(curMalloc);
curMalloc = NULL;
/* set mallocList to NULL, if last element freed */
if(vtcuptiActivity->gpuMemAllocated == 0) {
vtcuptiActivity->gpuMemList = NULL;
}
VT_CUPTI_UNLOCK();
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
return;
}
lastMalloc = curMalloc;
curMalloc = curMalloc->next;
}
VT_CUPTI_UNLOCK();
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
vt_warning("[CUPTI Activity] free CUDA memory, which has not been allocated!");
}
