void mp_barrier(cycles_t *measurement)
{
coreid_t tid = get_core_id();
#ifdef QRM_DBG_ENABLED
++_num_barrier;
uint32_t _num_barrier_recv = _num_barrier;
#endif
debug_printfff(DBG__REDUCE, "barrier enter #%d\n", _num_barrier);
// Recution
// --------------------------------------------------
#ifdef QRM_DBG_ENABLED
uint32_t _tmp =
#endif
mp_reduce(_num_barrier);
#ifdef QRM_DBG_ENABLED
// Sanity check
if (tid==get_sequentializer()) {
assert (_tmp == get_num_threads()*_num_barrier);
}
if (measurement)
*measurement = bench_tsc();
#endif
// Broadcast
// --------------------------------------------------
if (tid == get_sequentializer()) {
mp_send_ab(_num_barrier);
} else {
#ifdef QRM_DBG_ENABLED
_num_barrier_recv =
#endif
mp_receive_forward(0);
}
#ifdef QRM_DBG_ENABLED
if (_num_barrier_recv != _num_barrier) {
debug_printf("ASSERTION fail %d != %d\n", _num_barrier_recv, _num_barrier);
}
assert (_num_barrier_recv == _num_barrier);
// Add a shared memory barrier to absolutely make sure that
// everybody finished the barrier before leaving - this simplifies
// debugging, as the programm will get stuck if barriers are
// broken, rather than some threads (wrongly) continuing and
// causing problems somewhere else
#if 0 // Enable separately
debug_printfff(DBG_REDUCE, "finished barrier .. waiting for others\n");
shl_barrier_shm(get_num_threads());
#endif
#endif
debug_printfff(DBG__REDUCE, "barrier complete #%d\n", _num_barrier);
}
amp *
mp_sqrt_to(amp *r, amp *a, amp *rp)
{
amp *c;
amp *b;
int l;
int i;
int j;
mp_reduce(a);
if (a->denom)
return 0;
l = a->len;
l /= 2;
if (r)
c = r;
else
c = new_amp();
MP_NEED(c,l+1);
if (l) {
for(i = l-1, j = a->len-1; i >= 0; i--,j--)
c->data[i] = a->data[j];
c->len = l;
} else {
if (a->len == 1 && a->data[0] <= 1) {
mp_copy_to(c,a);
return c;
}
MP_ASSIGN_SMALL(c,1);
}
MP_TOUCH(c);
b = new_amp();
mp_div_to(b,a,c,(amp*)0);
mp_add_to(b,b,c);
mp_div_x_to(b,b,2,(mp_long*)0);
do {
mp_copy_to(c,b);
mp_div_to(b,a,c,(amp*)0);
mp_add_to(b,b,c);
mp_div_x_to(b,b,2,(mp_long*)0);
} while (mp_cmp(c,b) > 0);
if (rp) {
mp_mul_to(b,c,c);
mp_sub_to(rp,a,b);
}
mp_free(b);
return c;
}
