void
printcounters(struct counter *ctrs, uint64_t duration)
{
struct metrics s = {0};
s.timestamp = _rdtsc();
s.duration = duration;
// We skip the last core
int corethreads =0;
for (int cpu = 1; cpu < gbl.ncpus-3; ++cpu)
{
double delta[NEVENTS];
// volatile because another thread is changing it.
volatile struct counter *p = &ctrs[cpu];
for (int i = 0; i < NEVENTS; ++i)
{
union {
__m512d c;
uint64_t values[8];
} t;
t.c = _mm512_load_pd((void *)&p->counts[i][0]);
delta[i] = perf_scale_delta(t.values, lastctr[cpu].counts[i]);
_mm512_storenrngo_pd((void *)&lastctr[cpu].counts[i][0], t.c);
if (delta[i] < 0)
delta[i] = 0;
sevents[i] += delta[i];
}
if (2*delta[clocks1] > duration)
{
s.nthreads += 1;
corethreads += 1;
}
if ((cpu % 4) == 0) // Last thread on this core
{
if (corethreads)
s.ncores += 1;
corethreads = 0;
}
s.vpu_ea += delta[vpu_ea];
s.instrs += delta[instrs];
s.vinstrs += delta[vpu_ie];
}
uint64_t nreads = 0, nwrites = 0;
for (int i = 0; i < NGBOXES; ++i)
for (int j = 0; j < 2; ++j)
{
nreads += pmu_rdctr(i, j, 0);
nwrites += pmu_rdctr(i, j, 1);
}
s.rbytes = (nreads - prevnreads) * 64;
s.wbytes = (nwrites - prevnwrites)* 64;
prevnreads = nreads;
prevnwrites = nwrites;
sample(&s);
}
inline void transfer_omp_loop_nontemp(uintptr_t rbuf, uintptr_t sbuf, size_t size, HMPI_Request recv_req, HMPI_Request send_req){
int N_DOUBLES_PER_BLOCK = (64/sizeof(char)) ;
size_t total = size / 64 ;
int i = 0;
//#pragma vector nontemporal
#pragma omp parallel for
for (i = 0; i < total; i++) {
__m512d v_b = _mm512_load_pd(sbuf+ N_DOUBLES_PER_BLOCK*i);
_mm512_storenrngo_pd(rbuf+ N_DOUBLES_PER_BLOCK*i, v_b);
}
}
static void
readcounters(int cpu)
{
struct counter *cp = &counters[cpu];
for (int i = 0; i < NEVENTS; ++i)
{
struct evinfo *ep = &events[cpu][i];
union {
__m512d c;
uint64_t values[8];
} u;
readctr((perf_event_mmap_page *)ep->buf, ep->fd, u.values);
if (u.values[0])
_mm512_storenrngo_pd(&cp->counts[i][0], u.c);
}
}
