static int
mpiPi_insert_callsite_records (callsite_stats_t * p)
{
callsite_stats_t *csp = NULL;
mpiPi_query_src (p); /* sets the file/line in p */
/* If exists, accumulate, otherwise insert. This is
specifically for optimizations that have multiple PCs for
one src line. We aggregate across rank after this.
The collective_report reporting approach does not aggregate individual
process callsite information at the collector process.
*/
if (mpiPi.collective_report == 0)
{
if (NULL == h_search (mpiPi.global_callsite_stats, p, (void **) &csp))
{
int j;
callsite_stats_t *newp = NULL;
newp = (callsite_stats_t *) malloc (sizeof (callsite_stats_t));
bzero (newp, sizeof (callsite_stats_t));
newp->op = p->op;
newp->rank = p->rank;
for (j = 0; j < MPIP_CALLSITE_STACK_DEPTH; j++)
{
newp->pc[j] = p->pc[j];
newp->filename[j] = p->filename[j];
newp->functname[j] = p->functname[j];
newp->lineno[j] = p->lineno[j];
}
newp->csid = p->csid;
newp->count = p->count;
newp->cumulativeTime = p->cumulativeTime;
newp->cumulativeTimeSquared = p->cumulativeTimeSquared;
newp->maxDur = p->maxDur;
newp->minDur = p->minDur;
newp->maxDataSent = p->maxDataSent;
newp->minDataSent = p->minDataSent;
newp->cumulativeDataSent = p->cumulativeDataSent;
newp->maxIO = p->maxIO;
newp->minIO = p->minIO;
newp->cumulativeIO = p->cumulativeIO;
newp->arbitraryMessageCount = p->arbitraryMessageCount;
newp->cookie = MPIP_CALLSITE_STATS_COOKIE;
/* insert new record into global */
h_insert (mpiPi.global_callsite_stats, newp);
}
else
{
csp->count += p->count;
csp->cumulativeTime += p->cumulativeTime;
assert (csp->cumulativeTime >= 0);
csp->cumulativeTimeSquared += p->cumulativeTimeSquared;
assert (csp->cumulativeTimeSquared >= 0);
csp->maxDur = max (csp->maxDur, p->maxDur);
csp->minDur = min (csp->minDur, p->minDur);
csp->maxDataSent = max (csp->maxDataSent, p->maxDataSent);
csp->minDataSent = min (csp->minDataSent, p->minDataSent);
csp->cumulativeDataSent += p->cumulativeDataSent;
csp->maxIO = max (csp->maxIO, p->maxIO);
csp->minIO = min (csp->minIO, p->minIO);
csp->cumulativeIO += p->cumulativeIO;
csp->arbitraryMessageCount += p->arbitraryMessageCount;
}
}
/* Collect aggregate callsite summary information indpendent of rank. */
if (NULL == h_search (mpiPi.global_callsite_stats_agg, p, (void **) &csp))
{
int j;
callsite_stats_t *newp = NULL;
newp = (callsite_stats_t *) malloc (sizeof (callsite_stats_t));
bzero (newp, sizeof (callsite_stats_t));
newp->op = p->op;
newp->rank = -1;
for (j = 0; j < MPIP_CALLSITE_STACK_DEPTH; j++)
{
newp->pc[j] = p->pc[j];
newp->filename[j] = p->filename[j];
newp->functname[j] = p->functname[j];
newp->lineno[j] = p->lineno[j];
}
newp->csid = p->csid;
newp->count = p->count;
newp->cumulativeTime = p->cumulativeTime;
newp->cumulativeTimeSquared = p->cumulativeTimeSquared;
newp->maxDur = p->maxDur;
newp->minDur = p->minDur;
newp->maxDataSent = p->maxDataSent;
newp->minDataSent = p->minDataSent;
newp->cumulativeDataSent = p->cumulativeDataSent;
newp->cumulativeIO = p->cumulativeIO;
newp->maxIO = p->maxIO;
newp->minIO = p->minIO;
newp->cookie = MPIP_CALLSITE_STATS_COOKIE;
if (mpiPi.calcCOV)
{
newp->siteData = (double *) malloc (mpiPi.size * sizeof (double));
newp->siteData[0] = p->cumulativeTime;
newp->siteDataIdx = 1;
}
/* insert new record into global */
h_insert (mpiPi.global_callsite_stats_agg, newp);
}
else
{
csp->count += p->count;
csp->cumulativeTime += p->cumulativeTime;
assert (csp->cumulativeTime >= 0);
csp->cumulativeTimeSquared += p->cumulativeTimeSquared;
assert (csp->cumulativeTimeSquared >= 0);
csp->maxDur = max (csp->maxDur, p->maxDur);
csp->minDur = min (csp->minDur, p->minDur);
csp->maxDataSent = max (csp->maxDataSent, p->maxDataSent);
csp->minDataSent = min (csp->minDataSent, p->minDataSent);
csp->cumulativeDataSent += p->cumulativeDataSent;
csp->maxIO = max (csp->maxIO, p->maxIO);
csp->minIO = min (csp->minIO, p->minIO);
csp->cumulativeIO += p->cumulativeIO;
if (mpiPi.calcCOV)
{
csp->siteData[csp->siteDataIdx] = p->cumulativeTime;
csp->siteDataIdx += 1;
}
}
/* Do global accumulation while we are iterating through individual callsites */
mpiPi.global_task_info[p->rank].mpi_time += p->cumulativeTime;
mpiPi.global_mpi_time += p->cumulativeTime;
assert (mpiPi.global_mpi_time >= 0);
mpiPi.global_mpi_size += p->cumulativeDataSent;
mpiPi.global_mpi_io += p->cumulativeIO;
if (p->cumulativeTime > 0)
mpiPi.global_time_callsite_count++;
if (p->cumulativeDataSent > 0)
{
mpiPi.global_mpi_msize_threshold_count += p->arbitraryMessageCount;
mpiPi.global_mpi_sent_count += p->count;
}
return 1;
}
static int
mpiPi_insert_callsite_records (callsite_stats_t * p)
{
callsite_stats_t *csp = NULL;
mpiPi_query_src (p); /* sets the file/line in p */
/* If exists, accumulate, otherwise insert. This is
specifically for optimizations that have multiple PCs for
one src line. We aggregate across rank after this.
The collective_report reporting approach does not aggregate individual
process callsite information at the collector process.
*/
if (mpiPi.collective_report == 0)
{
if (NULL == h_search (mpiPi.global_callsite_stats, p, (void **) &csp))
{
callsite_stats_t *newp = NULL;
newp = (callsite_stats_t *) malloc (sizeof (callsite_stats_t));
memcpy (newp, p, sizeof (callsite_stats_t));
/* insert new record into global */
h_insert (mpiPi.global_callsite_stats, newp);
}
else
mpiPi_merge_individual_callsite_records (csp, p);
}
/* Collect aggregate callsite summary information indpendent of rank. */
if (NULL == h_search (mpiPi.global_callsite_stats_agg, p, (void **) &csp))
{
callsite_stats_t *newp = NULL;
newp = (callsite_stats_t *) malloc (sizeof (callsite_stats_t));
memcpy (newp, p, sizeof (callsite_stats_t));
newp->rank = -1;
if (mpiPi.calcCOV)
{
newp->siteData = (double *) malloc (mpiPi.size * sizeof (double));
newp->siteData[0] = p->cumulativeTime;
newp->siteDataIdx = 1;
}
/* insert new record into global */
h_insert (mpiPi.global_callsite_stats_agg, newp);
}
else
{
mpiPi_merge_individual_callsite_records (csp, p);
if (mpiPi.calcCOV)
{
csp->siteData[csp->siteDataIdx] = p->cumulativeTime;
csp->siteDataIdx += 1;
}
}
/* Do global accumulation while we are iterating through individual callsites */
mpiPi.global_task_mpi_time[p->rank] += p->cumulativeTime;
mpiPi.global_mpi_time += p->cumulativeTime;
assert (mpiPi.global_mpi_time >= 0);
mpiPi.global_mpi_size += p->cumulativeDataSent;
mpiPi.global_mpi_io += p->cumulativeIO;
mpiPi.global_mpi_rma += p->cumulativeRMA;
if (p->cumulativeTime > 0)
mpiPi.global_time_callsite_count++;
if (p->cumulativeDataSent > 0)
{
mpiPi.global_mpi_msize_threshold_count += p->arbitraryMessageCount;
mpiPi.global_mpi_sent_count += p->count;
}
return 1;
}
