/*
* Initialize Mutex, VampirTrace ids and registers the finalize function.
* This may be done implicitly by vt_cupti_count().
*/
void vt_cupti_init()
{
if(!vt_cupti_initialized){
#if (defined(VT_MT) || defined(VT_HYB))
VTThrd_createMutex(&VTThrdMutexCupti);
#endif
CUPTI_LOCK();
if(!vt_cupti_initialized){
vt_cntl_msg(2, "[CUPTI] Initializing ... ");
/* create VampirTrace counter group ID only once */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
rid_cupti_init = vt_def_region(VT_MASTER_THREAD, "vtcuptiHostThreadInit",
VT_NO_ID, VT_NO_LNO, VT_NO_LNO, "VT_CUPTI", VT_FUNCTION);
cgid_cupti = vt_def_counter_group(VT_MASTER_THREAD, "CUPTI");
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
vt_cupti_capList = vt_cupti_setupMetricList();
/* register the finalize function of the CUDA wrapper to be called before
* the program exits and CUDA has done its implicit clean-up */
atexit(vt_cupti_finalize);
vt_cupti_initialized = 1;
CUPTI_UNLOCK();
}
}
}
unsigned int VT_User_def__(const char* name, const char* file, int lno)
{
uint32_t rid;
VT_INIT;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
/* -- get region identifier by name, file, and line number -- */
if ( (rid = hash_get_region(name, file, lno)) == VT_NO_ID )
{
/* -- register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (rid = hash_get_region(name, file, lno)) == VT_NO_ID )
rid = register_region(0, name, file, lno);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
rid = register_region(0, name, file, lno);
#endif /* VT_MT || VT_HYB */
}
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
return rid;
}
void VT_User_start__(const char* name, const char *file, int lno) {
uint32_t rid;
uint64_t time;
/* -- if not yet initialized, initialize VampirTrace -- */
if ( vt_init ) {
VT_MEMHOOKS_OFF();
vt_init = 0;
vt_open();
VT_MEMHOOKS_ON();
}
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( (rid = hash_get((unsigned long) name)) == VT_NO_ID ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (rid = hash_get((unsigned long) name)) == VT_NO_ID )
rid = register_region(name, 0, file, lno);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
rid = register_region(name, 0, file, lno);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(&time, rid);
VT_MEMHOOKS_ON();
}
void VT_User_start__(const char* name, const char* file, int lno)
{
uint32_t rid;
uint64_t time;
VT_INIT;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
time = vt_pform_wtime();
/* -- get region identifier by address -- */
if ( (rid = hash_get_addr((unsigned long)name)) == VT_NO_ID )
{
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (rid = hash_get_addr((unsigned long)name)) == VT_NO_ID )
rid = register_region((unsigned long)name, name, file, lno);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
rid = register_region((unsigned long)name, name, file, lno);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(VT_CURRENT_THREAD, &time, rid);
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
}
VT_DECLDEF(void __profile_loop(struct profile_gen_struct* d))
{
uint64_t time;
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( d->data == NULL )
{
/* -- loop entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( d->data == NULL )
register_region(d, VT_LOOP);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
register_region(d, VT_LOOP);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(&time, *((uint32_t*)(d->data)));
VT_MEMHOOKS_ON();
}
void ___rouent2(struct s1 *p) {
uint32_t tid;
uint64_t time;
/* -- if not yet initialized, initialize VampirTrace -- */
if (rou_init)
{
rou_init = 0;
vt_open();
}
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
/* -- get calling thread id -- */
GET_THREAD_ID(tid);
VT_SUSPEND_MALLOC_TRACING(tid);
time = vt_pform_wtime();
if (!p->isseen)
{
char* rname = p->rout;
char* modpos;
/* fix opari output file names */
if ( (modpos = strstr(p->file, ".mod.")) != NULL )
{
strcpy(modpos, modpos+4);
}
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if (!p->isseen)
{
p->fid = vt_def_scl_file(tid, p->file);
p->rid = vt_def_region(tid, rname, p->fid, p->lineno,
VT_NO_LNO, NULL, VT_FUNCTION);
p->isseen = 1;
}
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
p->fid = vt_def_scl_file(tid, p->file);
p->rid = vt_def_region(tid, rname, p->fid, p->lineno,
VT_NO_LNO, NULL, VT_FUNCTION);
p->isseen = 1;
#endif /* VT_MT || VT_HYB */
}
/* write enter trace record */
vt_enter(tid, &time, p->rid);
VT_RESUME_MALLOC_TRACING(tid);
}
void VT_Dyn_start(uint32_t index, const char* name, const char* fname,
uint32_t lno, uint32_t loop)
{
uint64_t time;
uint32_t* rid;
vt_libassert(index < VT_MAX_DYNINST_REGIONS);
/* Ignore events if VT is initializing */
if( !dyn_init && !vt_is_alive ) return;
/* If not yet initialized, initialize VampirTrace */
if ( dyn_init )
{
VT_MEMHOOKS_OFF();
dyn_init = 0;
rtab = (uint32_t*)calloc(VT_MAX_DYNINST_REGIONS, sizeof(uint32_t));
if ( rtab == NULL )
vt_error();
vt_open();
vt_comp_finalize = VT_Dyn_finalize;
VT_MEMHOOKS_ON();
}
/* If VampirTrace already finalized, return */
if ( !vt_is_alive ) return;
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* Get region identifier
*/
rid = &(rtab[index]);
if ( *rid == 0 )
{
/* If region entered the first time, register region
*/
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( *rid == 0 )
*rid = register_region(name, fname, lno, loop);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
*rid = register_region(name, fname, lno, loop);
#endif /* VT_MT || VT_HYB */
}
/* Write enter record */
vt_enter(VT_CURRENT_THREAD, &time, *rid);
VT_MEMHOOKS_ON();
}
void vt_win_create( MPI_Win win, MPI_Comm comm )
{
MPI_Group group;
/* check if window already exists */
if (win_search( win ) == (uint32_t)-1)
{
uint32_t wid;
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif /* VT_MT || VT_HYB */
wid = vt_get_curid();
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif /* VT_MT || VT_HYB */
/* ask for group of win and register them */
PMPI_Win_get_group(win, &group);
vt_libassert(group != MPI_GROUP_NULL);
vt_group_create(group);
/* enter win in wins[] array */
if( free_win != (uint32_t)-1 )
{
wins[free_win].win = win;
wins[free_win].comm = comm;
wins[free_win].gid = vt_group_id(group);
wins[free_win].wid = wid;
}
else
{
/* raise maximum number of windows, if necessary */
if (last_win == max_wins)
RAISE_MAX(wins, max_wins, struct VTWin);
wins[last_win].win = win;
wins[last_win].comm = comm;
wins[last_win].gid = vt_group_id(group);
wins[last_win].wid = wid;
last_win++;
}
}
}
unsigned int VT_User_marker_def__(const char* mname, int mtype)
{
uint32_t mid;
uint32_t mtype_otf = OTF_MARKER_TYPE_UNKNOWN;
VT_INIT;
VT_MEMHOOKS_OFF();
switch(mtype)
{
case VT_MARKER_TYPE_ERROR:
{
mtype_otf = OTF_MARKER_TYPE_ERROR;
break;
}
case VT_MARKER_TYPE_WARNING:
{
mtype_otf = OTF_MARKER_TYPE_WARNING;
break;
}
case VT_MARKER_TYPE_HINT:
{
mtype_otf = OTF_MARKER_TYPE_HINT;
break;
}
default:
{
vt_error_msg("Unknown marker type %i", mtype);
break;
}
}
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
mid = (uint32_t)vt_def_marker(mname, mtype_otf);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
VT_MEMHOOKS_ON();
return mid;
}
unsigned int VT_User_marker_def__(const char* mname, int mtype)
{
uint32_t mid;
uint32_t _mtype = VT_MARKER_UNKNOWN;
VT_INIT;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
switch(mtype)
{
case VT_MARKER_TYPE_ERROR:
{
_mtype = VT_MARKER_ERROR;
break;
}
case VT_MARKER_TYPE_WARNING:
{
_mtype = VT_MARKER_WARNING;
break;
}
case VT_MARKER_TYPE_HINT:
{
_mtype = VT_MARKER_HINT;
break;
}
default:
{
vt_error_msg("Unknown marker type %i", mtype);
break;
}
}
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
mid = vt_def_marker(VT_CURRENT_THREAD, mname, _mtype);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
return mid;
}
VT_DECLDEF(void VT_User_start___f(const char* name, const char* file, int* lno,
int nl, int fl))
{
uint32_t rid;
uint64_t time;
int namlen;
int fillen;
char fnambuf[128];
char ffilbuf[1024];
VT_INIT;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
time = vt_pform_wtime();
/* -- convert Fortran to C strings -- */
namlen = ( nl < 128 ) ? nl : 127;
strncpy(fnambuf, name, namlen);
fnambuf[namlen] = '\0';
fillen = ( fl < 1024 ) ? fl : 1023;
strncpy(ffilbuf, file, fillen);
ffilbuf[fillen] = '\0';
/* -- get region identifier by address -- */
if ( (rid = hash_get_addr((unsigned long)name)) == VT_NO_ID )
{
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (rid = hash_get_addr((unsigned long)name)) == VT_NO_ID )
rid = register_region((unsigned long)name, fnambuf, NULL, ffilbuf, *lno);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
rid = register_region((unsigned long)name, fnambuf, NULL, ffilbuf, *lno);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(VT_CURRENT_THREAD, &time, rid);
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
} VT_GENERATE_F77_BINDINGS(vt_user_start__, VT_USER_START__,
void __cyg_profile_func_enter(void* func, void* callsite) {
void* funcptr;
uint64_t time;
HashNode* hn;
funcptr = DEREF_IA64_FUNC_PTR(func);
/* -- if not yet initialized, initialize VampirTrace -- */
if ( gnu_init ) {
VT_MEMHOOKS_OFF();
gnu_init = 0;
vt_open();
vt_comp_finalize = gnu_finalize;
get_symtab();
VT_MEMHOOKS_ON();
}
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( (hn = hash_get((long)funcptr))) {
if ( hn->vtid == VT_NO_ID ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if( hn->vtid == VT_NO_ID )
register_region(hn);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
register_region(hn);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(VT_CURRENT_THREAD, &time, hn->vtid);
}
VT_MEMHOOKS_ON();
}
void VT_User_start___f(const char* name, const char *file, int *lno, int nl, int fl) {
uint32_t rid;
uint64_t time;
int namlen;
int fillen;
/* -- convert Fortran to C strings -- */
namlen = ( nl < 128 ) ? nl : 127;
fillen = ( fl < 1024 ) ? fl : 1023;
strncpy(fnambuf, name, namlen);
fnambuf[namlen] = '\0';
strncpy(ffilbuf, file, fillen);
ffilbuf[fillen] = '\0';
/* -- if not yet initialized, initialize VampirTrace -- */
if ( vt_init ) {
VT_MEMHOOKS_OFF();
vt_init = 0;
vt_open();
VT_MEMHOOKS_ON();
}
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( (rid = hash_get((unsigned long) name)) == VT_NO_ID ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (rid = hash_get((unsigned long) name)) == VT_NO_ID )
rid = register_region(fnambuf, (unsigned long) name, ffilbuf, *lno);
VTTHRD_UNLOCK_IDS();
#else
rid = register_region(fnambuf, (unsigned long) name, ffilbuf, *lno);
#endif
}
/* -- write enter record -- */
vt_enter(&time, rid);
VT_MEMHOOKS_ON();
} VT_GENERATE_F77_BINDINGS(vt_user_start__, VT_USER_START__,
void _ftrace_enter2_() {
char *func = (char *)vftr_getname();
int len = vftr_getname_len();
uint32_t rid;
uint64_t time;
/* -- if not yet initialized, initialize VampirTrace -- */
if ( necsx_init ) {
VT_MEMHOOKS_OFF();
necsx_init = 0;
vt_open();
vt_comp_finalize = &ftrace_finalize;
VT_MEMHOOKS_ON();
}
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
/* -- ignore NEC OMP runtime functions -- */
if ( strchr(func, '$') != NULL ) return;
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( (rid = hash_get((long) func)) == VT_NO_ID ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (rid = hash_get((long) func)) == VT_NO_ID )
rid = register_region(func, len);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
rid = register_region(func, len);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(VT_CURRENT_THREAD, &time, rid);
VT_MEMHOOKS_ON();
}
unsigned int VT_User_msg_comm_def__(const char* cname)
{
uint32_t cid;
VT_INIT;
VT_SUSPEND_MALLOC_TRACING(VT_CURRENT_THREAD);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
cid = vt_def_user_comm(VT_CURRENT_THREAD, cname);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
VT_RESUME_MALLOC_TRACING(VT_CURRENT_THREAD);
return cid;
}
unsigned int VT_User_count_group_def__(const char* gname)
{
uint32_t gid;
VT_INIT;
VT_MEMHOOKS_OFF();
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
gid = vt_def_counter_group(VT_CURRENT_THREAD, gname);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
VT_MEMHOOKS_ON();
return gid;
}
void __func_trace_enter(char* name, char* fname, int lno) {
HashNode *hn;
uint64_t time;
/* -- if not yet initialized, initialize VampirTrace -- */
if ( xl_init ) {
VT_MEMHOOKS_OFF();
xl_init = 0;
vt_open();
vt_comp_finalize = &xl_finalize;
VT_MEMHOOKS_ON();
}
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
/* -- ignore IBM OMP runtime functions -- */
if ( strchr(name, '@') != NULL ) return;
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( (hn = hash_get((long) name)) == 0 ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (hn = hash_get((long) name)) == 0 )
hn = register_region(name, fname, lno);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
hn = register_region(name, fname, lno);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(VT_CURRENT_THREAD, &time, hn->vtid);
VT_MEMHOOKS_ON();
}
void phat_enter(char *str, int *id) {
uint64_t time;
/* -- if not yet initialized, initialize VampirTrace -- */
if ( phat_init ) {
VT_MEMHOOKS_OFF();
phat_init = 0;
vt_open();
vt_comp_finalize = &phat_finalize;
VT_MEMHOOKS_ON();
}
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
/* -- ignore SUN OMP runtime functions -- */
if ( strchr(str, '$') != NULL ) return;
VT_MEMHOOKS_OFF();
time = vt_pform_wtime();
/* -- get region identifier -- */
if ( *id == -1 ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (*id = hash_get((long) str)) == VT_NO_ID )
*id = register_region(str);
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
*id = register_region(str);
#endif /* VT_MT || VT_HYB */
}
/* -- write enter record -- */
vt_enter(VT_CURRENT_THREAD, &time, *id);
VT_MEMHOOKS_ON();
}
void vt_group_create(MPI_Group group)
{
uint32_t i;
/* check if group already exists */
if ((i = group_search( group )) == (uint32_t)-1)
{
uint32_t gid;
/* raise maximum number of groups, if necessary */
if (last_group == max_groups)
RAISE_MAX(groups, max_groups, struct VTGroup);
/* create group entry in grpv */
group_to_bitvector( group );
/* register mpi group definition (as communicator) */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif /* VT_MT || VT_HYB */
gid = vt_def_mpi_comm(VT_CURRENT_THREAD, VT_MPI_GROUP,
world.size_grpv, grpv);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif /* VT_MT || VT_HYB */
/* enter group in groups[] array */
groups[last_group].group = group;
groups[last_group].gid = gid;
groups[last_group].refcnt = 1;
last_group++;
}
else
{
/* count additional reference on group */
groups[i].refcnt++;
}
}
/*
* 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();
}
}
}
/* no need to lock, because it is only called by vt_cupti_callback_init() */
void vt_cupti_activity_init()
{
/*if(!vt_cuptiact_initialized){
vt_cupti_init();
VT_CUPTI_LOCK();*/
if(!vt_cuptiact_initialized){
vt_cntl_msg(2, "[CUPTI Activity] Initializing ... ");
{
vt_cuptiact_bufSize = vt_env_cudatrace_bsize();
/* no buffer size < 1024 bytes allowed (see CUPTI documentation) */
if(vt_cuptiact_bufSize < 1024){
if(vt_cuptiact_bufSize > 0){
vt_warning("[CUPTI Activity] Buffer size has to be at least 1024 "
"bytes! It has been set to %d.", vt_cuptiact_bufSize);
}
vt_cuptiact_bufSize = VT_CUPTI_ACT_DEFAULT_BSIZE;
}
/* queue a global buffer to initialize CUPTI before CUDA init
vt_cuptiact_buffer = (uint8_t *)malloc(vt_cuptiact_bufSize);
VT_CUPTI_CALL(cuptiActivityEnqueueBuffer(NULL, 0,
vt_cuptiact_buffer, vt_cuptiact_bufSize),
"cuptiActivityEnqueueBuffer");*/
}
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
if(vt_gpu_trace_kernels > 1){
/* define kernel counters */
vt_cuptiact_cid_knStaticSharedMem = vt_def_counter(VT_MASTER_THREAD,
"staticSharedMemory", "Bytes",
VT_CNTR_ABS | VT_CNTR_NEXT | VT_CNTR_UNSIGNED,
vt_cupti_cgid_cuda_kernel, 0);
vt_cuptiact_cid_knDynamicSharedMem = vt_def_counter(VT_MASTER_THREAD,
"dynamicSharedMemory", "Bytes",
VT_CNTR_ABS | VT_CNTR_NEXT | VT_CNTR_UNSIGNED,
vt_cupti_cgid_cuda_kernel, 0);
vt_cuptiact_cid_knLocalMemTotal = vt_def_counter(VT_MASTER_THREAD,
"localMemoryPerKernel", "Bytes",
VT_CNTR_ABS | VT_CNTR_NEXT | VT_CNTR_UNSIGNED,
vt_cupti_cgid_cuda_kernel, 0);
vt_cuptiact_cid_knRegistersPerThread = vt_def_counter(VT_MASTER_THREAD,
"registersPerThread", "#",
VT_CNTR_ABS | VT_CNTR_NEXT | VT_CNTR_UNSIGNED,
vt_cupti_cgid_cuda_kernel, 0);
}
/* define region for GPU activity flush */
vt_cuptiact_rid_flush = vt_def_region(VT_MASTER_THREAD, "flushActivities",
VT_NO_ID, VT_NO_LNO, VT_NO_LNO, "VT_CUDA", VT_FUNCTION);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
/*** enable the activities ***/
/* enable kernel tracing */
if(vt_gpu_trace_kernels > 0){
#if (defined(CUPTI_API_VERSION) && (CUPTI_API_VERSION >= 3))
if((vt_gpu_config & VT_GPU_TRACE_CONCURRENT_KERNEL)
== VT_GPU_TRACE_CONCURRENT_KERNEL){
/*VT_CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL),
"cuptiActivityEnable");*/
VT_CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_CONCURRENT_KERNEL),
"cuptiActivityEnable");
}else
#endif
VT_CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_KERNEL),
"cuptiActivityEnable");
}
/* enable memory copy tracing */
if(vt_gpu_trace_mcpy){
VT_CUPTI_CALL(cuptiActivityEnable(CUPTI_ACTIVITY_KIND_MEMCPY),
"cuptiActivityEnable");
}
/* register the finalize function of VampirTrace CUPTI to be called before
* the program exits
atexit(vt_cupti_activity_finalize);*/
vt_cuptiact_initialized = 1;
/*VT_CUPTI_UNLOCK();
}*/
}
}
/*
* Parse the environment variable for CUPTI metrics (including CUDA device
* capabilities) and fill the capability metric list.
*
* @param capList points to the first element of the capability metric list
*/
static void vt_cupti_fillMetricList(vt_cupti_device_t *capList)
{
char *metricString = vt_env_cupti_events();
char *metric_sep = vt_env_metrics_sep();
char *metric, *metric_cap;
metric = strtok(metricString, metric_sep);
while (metric != NULL){
CUptiResult cuptiErr = CUPTI_SUCCESS;
vt_cupti_device_t *cuptiDev = NULL;
vt_cupti_evtctr_t *vtcuptiEvt = NULL;
int metr_major = 0;
int metr_minor = 0;
/* try to get CUDA device capability parsed from metric */
metr_major = atoi(metric);
metric_cap = strchr(metric+1, '.');
if(metric_cap){
metr_minor = atoi(metric_cap+1);
metric_cap = strchr(metric_cap+1, '_');
}
/* check whether device capability is given or not */
if(metric_cap){
metric = metric_cap + 1;
vt_cntl_msg(2, "Metric '%s', %d.%d", metric, metr_major, metr_minor);
cuptiDev = vt_cupti_checkMetricList(capList, metr_major, metr_minor);
if(cuptiDev == NULL){
metric = strtok(NULL, metric_sep);
continue;
}
vtcuptiEvt = (vt_cupti_evtctr_t*)malloc(sizeof(vt_cupti_evtctr_t));
cuptiErr = cuptiEventGetIdFromName(cuptiDev->cuDev, metric,
&vtcuptiEvt->cuptiEvtID);
if(cuptiErr != CUPTI_SUCCESS){
if(!strncmp(metric, "help", 4)) vt_cupti_showAllCounters(capList);
vt_warning("[CUPTI Events] Skipping invalid event '%s' for device %d",
metric, cuptiDev->cuDev);
metric = strtok(NULL, metric_sep);
continue;
}
/* create VampirTrace counter ID */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
vtcuptiEvt->vtCID = vt_def_counter(VT_MASTER_THREAD, metric, "#",
VT_CNTR_ABS | VT_CNTR_LAST | VT_CNTR_UNSIGNED, vt_cuptievt_cgid, 0);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
cuptiDev->evtNum++;
vtcuptiEvt->next = cuptiDev->vtcuptiEvtList;
cuptiDev->vtcuptiEvtList = vtcuptiEvt;
}else{
/* device capability is not given. Try to add metric to all devices */
uint32_t cid_metric = VT_NO_ID;
cuptiDev = capList;
while(cuptiDev != NULL){
vtcuptiEvt = (vt_cupti_evtctr_t*)malloc(sizeof(vt_cupti_evtctr_t));
cuptiErr = cuptiEventGetIdFromName(cuptiDev->cuDev, metric,
&vtcuptiEvt->cuptiEvtID);
if(cuptiErr != CUPTI_SUCCESS){
if(!strncmp(metric, "help", 4)) vt_cupti_showAllCounters(capList);
vt_warning("[CUPTI Events] Skipping invalid event '%s' for device %d",
metric, cuptiDev->cuDev);
}else{
/* create VampirTrace counter ID, if not yet done for other device */
if(cid_metric == VT_NO_ID){
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
cid_metric = vt_def_counter(VT_MASTER_THREAD, metric, "#",
VT_CNTR_ABS | VT_CNTR_LAST | VT_CNTR_UNSIGNED, vt_cuptievt_cgid, 0);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
}
cuptiDev->evtNum++;
vtcuptiEvt->vtCID = cid_metric;
vtcuptiEvt->next = cuptiDev->vtcuptiEvtList;
cuptiDev->vtcuptiEvtList = vtcuptiEvt;
}
cuptiDev = cuptiDev->next;
}
}
metric = strtok(NULL, metric_sep);
}
}
void vt_comm_create(MPI_Comm comm)
{
MPI_Group group, lgroup, rgroup;
VT_MPI_INT inter;
VT_MPI_INT size_grpv = 0;
uint32_t cid;
/* raise maximum number of communicators, if necessary */
if (last_comm == max_comms)
RAISE_MAX(comms, max_comms, struct VTComm);
/* ask for group of comm */
PMPI_Comm_test_inter(comm, &inter);
if (inter)
{
PMPI_Comm_remote_group(comm, &rgroup);
PMPI_Comm_group(comm, &lgroup);
PMPI_Group_union(lgroup, rgroup, &group);
PMPI_Group_free(&lgroup);
PMPI_Group_free(&rgroup);
}
else
{
PMPI_Comm_group(comm, &group);
}
/* create group entry in grpv except for
MPI_COMM_SELF and
MPI_COMM_WORLD (if the current rank isn't the first available one) */
if ((comm != MPI_COMM_SELF && comm != MPI_COMM_WORLD) ||
(comm == MPI_COMM_WORLD && vt_my_trace_is_first_avail))
{
group_to_bitvector( group );
size_grpv = world.size_grpv;
}
/* register mpi communicator definition */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif /* VT_MT || VT_HYB */
cid = vt_def_mpi_comm(VT_CURRENT_THREAD,
comm == MPI_COMM_WORLD ? VT_MPI_COMM_WORLD :
comm == MPI_COMM_SELF ? VT_MPI_COMM_SELF : VT_MPI_COMM_OTHER,
size_grpv, grpv);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif /* VT_MT || VT_HYB */
/* save communicator id for fast access in VT_COMM_ID */
if (comm == MPI_COMM_WORLD) vt_mpi_comm_world_cid = cid;
else if (comm == MPI_COMM_SELF) vt_mpi_comm_self_cid = cid;
/* enter comm in comms[] array */
comms[last_comm].comm = comm;
comms[last_comm].cid = cid;
last_comm++;
/* clean up */
PMPI_Group_free(&group);
}
/*
* This function is called at the entry of each function
* The call is generated by the IBM xl compilers
*
* XL C/C++ 11.x / XLF 13.x or later pass the address of a static pointer
* variable, initialized with NULL, as additional parameter.
*/
#if __IBMC__ > 1100
void __func_trace_enter(char* name, char* fname, int lno, HashNode** ihn)
#else /* __IBMC__ */
void __func_trace_enter(char* name, char* fname, int lno)
#endif /* __IBMC__ */
{
HashNode *hn;
uint32_t tid;
uint64_t time;
/* -- ignore IBM OMP runtime functions -- */
# if __IBMC__ > 1100
if ( *ihn == &ignored_func ) return;
# endif /* __IBMC__ */
if ( ( strchr(name, '@') != NULL ) || (strchr(name, '$') != NULL) ) {
# if __IBMC__ > 1100
*ihn = &ignored_func;
# endif /* __IBMC__ */
return;
}
/* -- if not yet initialized, initialize VampirTrace -- */
if ( xl_init ) {
xl_init = 0;
vt_open();
vt_comp_finalize = &xl_finalize;
}
/* -- if VampirTrace already finalized, return -- */
if ( !vt_is_alive ) return;
/* -- get calling thread id -- */
GET_THREAD_ID(tid);
VT_SUSPEND_MALLOC_TRACING(tid);
time = vt_pform_wtime();
# if __IBMC__ > 1100
/* -- region not yet registered -- */
if ( *ihn == NULL ) {
# endif /* __IBMC__ */
/* -- get region identifier -- */
if ( (hn = hash_get((long) name)) == 0 ) {
/* -- region entered the first time, register region -- */
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
if ( (hn = hash_get((long) name)) == 0 ) {
hn = register_region(tid, name, fname, lno);
# if __IBMC__ > 1100
*ihn = hn;
# endif /* __IBMC__ */
}
VTTHRD_UNLOCK_IDS();
#else /* VT_MT || VT_HYB */
hn = register_region(tid, name, fname, lno);
# if __IBMC__ > 1100
*ihn = hn;
# endif /* __IBMC__ */
#endif /* VT_MT || VT_HYB */
}
#if __IBMC__ > 1100
} else {
/* -- region already registered -- */
hn = *ihn;
}
#endif /* __IBMC__ */
/* -- write enter record -- */
vt_enter(tid, &time, hn->vtid);
VT_RESUME_MALLOC_TRACING(tid);
}
unsigned int VT_User_count_def__(const char* cname, const char* cunit, int ctype,
unsigned int gid)
{
uint32_t cid;
uint32_t cprop = VT_CNTR_ABS | VT_CNTR_NEXT;
VT_INIT;
VT_MEMHOOKS_OFF();
if (gid == (uint32_t)VT_COUNT_DEFGROUP)
{
if (def_gid == 0)
def_gid = VT_User_count_group_def__("User");
gid = def_gid;
}
switch(ctype)
{
case VT_COUNT_TYPE_SIGNED:
case VT_COUNT_TYPE_INTEGER:
case VT_COUNT_TYPE_INTEGER8:
{
cprop |= VT_CNTR_SIGNED;
break;
}
case VT_COUNT_TYPE_UNSIGNED:
{
cprop |= VT_CNTR_UNSIGNED;
break;
}
case VT_COUNT_TYPE_FLOAT:
case VT_COUNT_TYPE_REAL:
{
cprop |= VT_CNTR_FLOAT;
break;
}
case VT_COUNT_TYPE_DOUBLE:
{
cprop |= VT_CNTR_DOUBLE;
break;
}
default:
{
vt_error_msg("Unknown counter type %i", ctype);
break;
}
}
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_LOCK_IDS();
#endif
cid = vt_def_counter(VT_CURRENT_THREAD, cname, cprop, gid, cunit);
#if (defined(VT_MT) || defined(VT_HYB))
VTTHRD_UNLOCK_IDS();
#endif
VT_MEMHOOKS_ON();
return cid;
}
