Aquila-Consortium/libsharp2
0
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

better

Browse source
This commit is contained in:
Martin Reinecke 2018-12-18 15:45:55 +01:00
parent b80c2a5543
commit 0ee3a87e2d
1 changed files with 99 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

125
libsharp/sharp_core.c
View file

@ -283,54 +283,113 @@ NOINLINE static void calc_alm2map (sharp_job * restrict job,
NOINLINE static void calc_alm2map_alt (sharp_job * restrict job,
const sharp_Ylmgen_C * restrict gen, s0data_v * restrict d, int nth)
{
int l=gen->m, lmax=gen->lmax;
int lmax=gen->lmax;
int nv2 = (nth+VLEN-1)/VLEN;
job->opcnt += (lmax+1-l) * 6*nth;
const double inv_sqrt4pi = 0.2820947917738781434740397257803862929220;
job->opcnt += (lmax+1-gen->m) * 6*nth;
int full_ieee;
Tv csq[nv0];
Tv mfac = vload(gen->mfac[gen->m]);
if (gen->m&1) mfac=-mfac;
int below_limit = 1;
for (int i=0; i<nv2; ++i)
{
csq[i] = d->cth[i]*d->cth[i];
d->lam1[i]=vzero;
mypow(d->sth[i],l,gen->powlimit,&d->lam2[i],&d->scale[i]);
mypow(d->sth[i],gen->m,gen->powlimit,&d->lam2[i],&d->scale[i]);
d->lam2[i] *= mfac;
Tvnormalize(&d->lam2[i],&d->scale[i],sharp_ftol);
below_limit &= vallTrue(vlt(d->scale[i],vload(sharp_limscale)));
}
int l=gen->m, il=0;
while (below_limit)
{
if (l+2>gen->lmax) {l=gen->lmax+1;return;}
below_limit=1;
for (int i=0; i<nv2; ++i)
{
Tv tmp = (gen->a[il]*csq[i] + gen->b[il])*d->lam2[i] + d->lam1[i];
d->lam1[i] = d->lam2[i];
d->lam2[i] = tmp;
if (rescale(&d->lam1[i], &d->lam2[i], &d->scale[i], vload(sharp_ftol)))
below_limit &= vallTrue(vlt(d->scale[i],vload(sharp_limscale)));
}
l+=2; ++il;
}
const dcmplx * restrict alm=job->almtmp;
dcmplx * restrict alm2=RALLOC(dcmplx, gen->lmax+5);
const dcmplx * restrict alm=job->almtmp;
full_ieee=1;
for (int i=0; i<nv2; ++i)
{
getCorfac(d->scale[i], &d->corfac[i], gen->cf);
full_ieee &= vallTrue(vge(d->scale[i],vload(sharp_minscale)));
}
{
for (int il=0, l=gen->m; l<=gen->lmax; ++il,l+=2)
// int il=0, l=gen->m;
for (; (l<=lmax) && (!full_ieee); ++il, l+=2)
{
dcmplx al = alm[l];
dcmplx al1 = (l+1>gen->lmax) ? 0. : alm[l+1];
dcmplx al2 = (l+2>gen->lmax) ? 0. : alm[l+2];
alm2[l ] = gen->alpha[il]*(gen->eps[l+1]*al + gen->eps[l+2]*al2);
alm2[l+1] = gen->alpha[il]*al1;
}
}
for (int i=0; i<nv2; ++i)
getCorfac(d->scale[i], &d->corfac[i], gen->cf);
for (int il=0, l=gen->m; l<=lmax; ++il, l+=2)
{
Tv ar1=vload(creal(alm2[l ])), ai1=vload(cimag(alm2[l ]));
Tv ar2=vload(creal(alm2[l+1])), ai2=vload(cimag(alm2[l+1]));
Tv ar1=vload(creal(alm[l ])), ai1=vload(cimag(alm[l ]));
Tv ar2=vload(creal(alm[l+1])), ai2=vload(cimag(alm[l+1]));
full_ieee=1;
for (int i=0; i<nv2; ++i)
{
d->p1r[i] += d->lam2[i]*d->corfac[i]*ar1;
d->p1i[i] += d->lam2[i]*d->corfac[i]*ai1;
d->p2r[i] += d->cth[i]*d->lam2[i]*d->corfac[i]*ar2;
d->p2i[i] += d->cth[i]*d->lam2[i]*d->corfac[i]*ai2;
Tv tmp = (gen->a[il]*d->cth[i]*d->cth[i] + gen->b[il])*d->lam2[i] + d->lam1[i];
d->p2r[i] += d->lam2[i]*d->corfac[i]*ar2;
d->p2i[i] += d->lam2[i]*d->corfac[i]*ai2;
Tv tmp = (gen->a[il]*csq[i] + gen->b[il])*d->lam2[i] + d->lam1[i];
d->lam1[i] = d->lam2[i];
d->lam2[i] = tmp;
if (rescale(&d->lam1[i], &d->lam2[i], &d->scale[i], vload(sharp_ftol)))
{
getCorfac(d->scale[i], &d->corfac[i], gen->cf);
full_ieee &= vallTrue(vge(d->scale[i],vload(sharp_minscale)));
}
// full_ieee=0;
}
}
DEALLOC(alm2);
for (int i=0; i<nv2; ++i)
{
d->lam1[i] *= d->corfac[i];
d->lam2[i] *= d->corfac[i];
}
for (; l<=lmax-2; il+=2, l+=4)
{
Tv ar1=vload(creal(alm[l ])), ai1=vload(cimag(alm[l ]));
Tv ar2=vload(creal(alm[l+1])), ai2=vload(cimag(alm[l+1]));
Tv ar3=vload(creal(alm[l+2])), ai3=vload(cimag(alm[l+2]));
Tv ar4=vload(creal(alm[l+3])), ai4=vload(cimag(alm[l+3]));
Tv a1=vload(gen->a[il]), a2=vload(gen->a[il+1]),
b1=vload(gen->b[il]), b2=vload(gen->b[il+1]);
for (int i=0; i<nv2; ++i)
{
d->p1r[i] += d->lam2[i]*ar1;
d->p1i[i] += d->lam2[i]*ai1;
d->p2r[i] += d->lam2[i]*ar2;
d->p2i[i] += d->lam2[i]*ai2;
d->lam1[i] = (a1*csq[i] + b1)*d->lam2[i] + d->lam1[i];
d->p1r[i] += d->lam1[i]*ar3;
d->p1i[i] += d->lam1[i]*ai3;
d->p2r[i] += d->lam1[i]*ar4;
d->p2i[i] += d->lam1[i]*ai4;
d->lam2[i] = (a2*csq[i] + b2)*d->lam1[i] + d->lam2[i];
}
}
for (; l<=lmax; ++il, l+=2)
{
Tv ar1=vload(creal(alm[l ])), ai1=vload(cimag(alm[l ]));
Tv ar2=vload(creal(alm[l+1])), ai2=vload(cimag(alm[l+1]));
for (int i=0; i<nv2; ++i)
{
d->p1r[i] += d->lam2[i]*ar1;
d->p1i[i] += d->lam2[i]*ai1;
d->p2r[i] += d->lam2[i]*ar2;
d->p2i[i] += d->lam2[i]*ai2;
Tv tmp = (gen->a[il]*csq[i] + gen->b[il])*d->lam2[i] + d->lam1[i];
d->lam1[i] = d->lam2[i];
d->lam2[i] = tmp;
}
}
}
}
NOINLINE static void map2alm_kernel(s0data_v * restrict d,
@ -833,6 +892,17 @@ NOINLINE static void inner_loop_a2m(sharp_job *job, const int *ispair,
{
if (job->spin==0)
{
//adjust the a_lm for the new algorithm
dcmplx * restrict alm=job->almtmp;
for (int il=0, l=gen->m; l<=gen->lmax; ++il,l+=2)
{
dcmplx al = alm[l];
dcmplx al1 = (l+1>gen->lmax) ? 0. : alm[l+1];
dcmplx al2 = (l+2>gen->lmax) ? 0. : alm[l+2];
alm[l ] = gen->alpha[il]*(gen->eps[l+1]*al + gen->eps[l+2]*al2);
alm[l+1] = gen->alpha[il]*al1;
}
const int nval=nv0*VLEN;
int ith=0;
int itgt[nval];
@ -868,6 +938,9 @@ NOINLINE static void inner_loop_a2m(sharp_job *job, const int *ispair,
calc_alm2map_alt (job, gen, &d.v, nth);
for (int i=0; i<nth; ++i)
{
//adjust for new algorithm
d.s.p2r[i]*=d.s.cth[i];
d.s.p2i[i]*=d.s.cth[i];
int tgt=itgt[i];
int phas_idx = tgt*job->s_th + mi*job->s_m;
complex double r1 = d.s.p1r[i] + d.s.p1i[i]*_Complex_I,