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

legendre transforms: Polishing and bugfixes

Browse source 
vbroadcast didn't work properly for some reason, vload does..
This commit is contained in:
Dag Sverre Seljebotn 2015-04-23 09:57:45 +02:00
parent ea8671c2ec
commit f2fe4f9ca2
6 changed files with 204 additions and 161 deletions
Download patch file Download diff file Expand all files Collapse all files

38
libsharp/sharp_legendre.c.in
View file

@ -46,6 +46,21 @@
* \author Dag Sverre Seljebotn
*/
#ifndef NO_LEGENDRE
#if (VLEN==8)
#error This code is not tested with MIC; please compile with -DNO_LEGENDRE
/* ...or test it (it probably works) and remove this check */
#endif
#ifndef SHARP_LEGENDRE_CS
#define SHARP_LEGENDRE_CS 4
#endif
#define MAX_CS 6
#if (SHARP_LEGENDRE_CS > MAX_CS)
#error (SHARP_LEGENDRE_CS > MAX_CS)
#endif
#include "sharp_legendre.h"
#include "sharp_vecsupport.h"
@ -66,18 +81,18 @@ static void legendre_transform_vec{{cs}}{{T}}({{scalar}} *recfacs, {{scalar}} *b
x{{i}} = vloadu{{T}}(xarr + {{i}} * VLEN{{T}});
Pm1_{{i}} = vload{{T}}(1.0);
P_{{i}} = x{{i}};
b = vbroadcast{{T}}(bl);
b = vload{{T}}(*bl);
y{{i}} = vmul{{T}}(Pm1_{{i}}, b);
/*{ endfor }*/
b = vbroadcast{{T}}(bl + 1);
b = vload{{T}}(*(bl + 1));
/*{ for i in range(cs) }*/
vfmaeq{{T}}(y{{i}}, P_{{i}}, b);
/*{ endfor }*/
for (l = 2; l <= lmax; ++l) {
b = vbroadcast{{T}}(bl + l);
R = vbroadcast{{T}}(recfacs + l);
b = vload{{T}}(*(bl + l));
R = vload{{T}}(*(recfacs + l));
/*
P = x * Pm1 + recfacs[l] * (x * Pm1 - Pm2)
@ -87,8 +102,8 @@ static void legendre_transform_vec{{cs}}{{T}}({{scalar}} *recfacs, {{scalar}} *b
W1 = vmul{{T}}(x{{i}}, Pm1_{{i}});
W2 = W1;
W2 = vsub{{T}}(W2, Pm2_{{i}});
vfmaeq{{T}}(W1, W2, R);
P_{{i}} = W1;
vfmaeq{{T}}(P_{{i}}, W2, R);
vfmaeq{{T}}(y{{i}}, P_{{i}}, b);
/*{ endfor }*/
@ -117,24 +132,21 @@ void sharp_legendre_transform_recfac{{T}}({{scalar}} *r, ptrdiff_t lmax) {
Compute sum_l b_l P_l(x_i) for all i.
*/
/*{set cs=4}*/
#define CS {{cs}}
#define LEN (CS * VLEN)
#define LEN_s (CS * VLEN_s)
#define LEN (SHARP_LEGENDRE_CS * VLEN)
#define LEN_s (SHARP_LEGENDRE_CS * VLEN_s)
/*{ for scalar, T in [("double", ""), ("float", "_s")] }*/
void sharp_legendre_transform{{T}}({{scalar}} *bl,
{{scalar}} *recfac,
ptrdiff_t lmax,
{{scalar}} *x, {{scalar}} *out, ptrdiff_t nx) {
{{scalar}} xchunk[LEN{{T}}], outchunk[LEN{{T}}];
{{scalar}} xchunk[MAX_CS * VLEN{{T}}], outchunk[MAX_CS * LEN{{T}}];
int compute_recfac;
ptrdiff_t i, j, len;
compute_recfac = (recfac == NULL);
if (compute_recfac) {
recfac = memalign(16, sizeof({{scalar}}) * (lmax + 1));
recfac = malloc(sizeof({{scalar}}) * (lmax + 1));
sharp_legendre_transform_recfac{{T}}(recfac, lmax);
}
@ -160,3 +172,5 @@ void sharp_legendre_transform{{T}}({{scalar}} *bl,
}
}
/*{ endfor }*/
#endif