From 65f47d10ccd35c7d1f728e063b9967c7046170f1 Mon Sep 17 00:00:00 2001
From: Martin Reinecke <martin@mpa-garching.mpg.de>
Date: Mon, 10 Dec 2018 14:37:34 +0100
Subject: [PATCH] clearer macro names

---
 libsharp/sharp_core_inc2.c  | 28 ++++++++++++++++++----------
 libsharp/sharp_vecsupport.h |  8 +++++++-
 2 files changed, 25 insertions(+), 11 deletions(-)

diff --git a/libsharp/sharp_core_inc2.c b/libsharp/sharp_core_inc2.c
index 9a2e26b..017df07 100644
--- a/libsharp/sharp_core_inc2.c
+++ b/libsharp/sharp_core_inc2.c
@@ -41,13 +41,16 @@ if (njobs>1)
     Tb lam_3, lam_4;
     Tv r0=vload(rf[l].f[0]),r1=vload(rf[l].f[1]);
     for (int i=0; i<nvec; ++i)
-      lam_3.v[i] = vsub(vmul(vmul(cth.v[i],lam_2.v[i]),r0),vmul(lam_1.v[i],r1));
+//      lam_3.v[i] = vsub(vmul(vmul(cth.v[i],lam_2.v[i]),r0),vmul(lam_1.v[i],r1));
+      lam_3.v[i] = vabmc(vmul(cth.v[i],lam_2.v[i]),r0,vmul(lam_1.v[i],r1));
     r0=vload(rf[l+1].f[0]);r1=vload(rf[l+1].f[1]);
     for (int i=0; i<nvec; ++i)
-      lam_4.v[i] = vsub(vmul(vmul(cth.v[i],lam_3.v[i]),r0),vmul(lam_2.v[i],r1));
+//      lam_4.v[i] = vsub(vmul(vmul(cth.v[i],lam_3.v[i]),r0),vmul(lam_2.v[i],r1));
+      lam_4.v[i] = vabmc(vmul(cth.v[i],lam_3.v[i]),r0,vmul(lam_2.v[i],r1));
     r0=vload(rf[l+2].f[0]);r1=vload(rf[l+2].f[1]);
     for (int i=0; i<nvec; ++i)
-      lam_1.v[i] = vsub(vmul(vmul(cth.v[i],lam_4.v[i]),r0),vmul(lam_3.v[i],r1));
+//      lam_1.v[i] = vsub(vmul(vmul(cth.v[i],lam_4.v[i]),r0),vmul(lam_3.v[i],r1));
+      lam_1.v[i] = vabmc(vmul(cth.v[i],lam_4.v[i]),r0,vmul(lam_3.v[i],r1));
     for (int j=0; j<njobs; ++j)
       {
       Tv ar2=vload(creal(alm[njobs*l+j])),
@@ -71,7 +74,8 @@ if (njobs>1)
       }
     r0=vload(rf[l+3].f[0]);r1=vload(rf[l+3].f[1]);
     for (int i=0; i<nvec; ++i)
-      lam_2.v[i] = vsub(vmul(vmul(cth.v[i],lam_1.v[i]),r0),vmul(lam_4.v[i],r1));
+//      lam_2.v[i] = vsub(vmul(vmul(cth.v[i],lam_1.v[i]),r0),vmul(lam_4.v[i],r1));
+      lam_2.v[i] = vabmc(vmul(cth.v[i],lam_1.v[i]),r0,vmul(lam_4.v[i],r1));
     l+=4;
     }
   }
@@ -127,13 +131,15 @@ NOINLINE static void Z(map2alm_kernel) (const Tb cth,
   while (l<lmax)
     {
     for (int i=0; i<nvec; ++i)
-      lam_1.v[i] = vload(rf[l].f[0])*(cth.v[i]*lam_2.v[i])
-                 - vload(rf[l].f[1])*lam_1.v[i];
+      lam_1.v[i] = vabmc(vload(rf[l].f[0]),vmul(cth.v[i],lam_2.v[i]),
+                   vmul(vload(rf[l].f[1]),lam_1.v[i]));
     for (int j=0; j<njobs; ++j)
       for (int i=0; i<nvec; ++i)
         {
-        atmp[2*(l*njobs+j)]+=lam_2.v[i]*p1[j].r.v[i];
-        atmp[2*(l*njobs+j)+1]+=lam_2.v[i]*p1[j].i.v[i];
+        vfmaeq(atmp[2*(l*njobs+j)],lam_2.v[i],p1[j].r.v[i]);
+        vfmaeq(atmp[2*(l*njobs+j)+1],lam_2.v[i],p1[j].i.v[i]);
+//        atmp[2*(l*njobs+j)]+=lam_2.v[i]*p1[j].r.v[i];
+//        atmp[2*(l*njobs+j)+1]+=lam_2.v[i]*p1[j].i.v[i];
         }
     for (int i=0; i<nvec; ++i)
       lam_2.v[i] = vload(rf[l+1].f[0])*(cth.v[i]*lam_1.v[i])
@@ -141,8 +147,10 @@ NOINLINE static void Z(map2alm_kernel) (const Tb cth,
     for (int j=0; j<njobs; ++j)
       for (int i=0; i<nvec; ++i)
         {
-        atmp[2*((l+1)*njobs+j)]+=lam_1.v[i]*p2[j].r.v[i];
-        atmp[2*((l+1)*njobs+j)+1]+=lam_1.v[i]*p2[j].i.v[i];
+        vfmaeq(atmp[2*((l+1)*njobs+j)],lam_1.v[i],p2[j].r.v[i]);
+        vfmaeq(atmp[2*((l+1)*njobs+j)+1],lam_1.v[i],p2[j].i.v[i]);
+//        atmp[2*((l+1)*njobs+j)]+=lam_1.v[i]*p2[j].r.v[i];
+//        atmp[2*((l+1)*njobs+j)+1]+=lam_1.v[i]*p2[j].i.v[i];
         }
     l+=2;
     }
diff --git a/libsharp/sharp_vecsupport.h b/libsharp/sharp_vecsupport.h
index 5250948..ff3f573 100644
--- a/libsharp/sharp_vecsupport.h
+++ b/libsharp/sharp_vecsupport.h
@@ -58,6 +58,7 @@ typedef int Tm;
 #define vfmaeq(a,b,c) ((a)+=(b)*(c))
 #define vfmaeq_s(a,b,c) ((a)+=(b)*(c))
 #define vfmseq(a,b,c) ((a)-=(b)*(c))
+#define vabmc(a,b,c) ((a)*(b)-(c))
 #define vfmaaeq(a,b,c,d,e) ((a)+=(b)*(c)+(d)*(e))
 #define vfmaseq(a,b,c,d,e) ((a)+=(b)*(c)-(d)*(e))
 #define vneg(a) (-(a))
@@ -125,6 +126,7 @@ static inline Tv vblend__(Tv m, Tv a, Tv b)
 #define vfmaeq(a,b,c) a=_mm_add_pd(a,_mm_mul_pd(b,c))
 #define vfmaeq_s(a,b,c) a=_mm_add_ps(a,_mm_mul_ps(b,c))
 #define vfmseq(a,b,c) a=_mm_sub_pd(a,_mm_mul_pd(b,c))
+#define vabmc(a,b,c) _mm_sub_pd(_mm_mul_pd(a,b),c)
 #define vfmaaeq(a,b,c,d,e) \
   a=_mm_add_pd(a,_mm_add_pd(_mm_mul_pd(b,c),_mm_mul_pd(d,e)))
 #define vfmaseq(a,b,c,d,e) \
@@ -182,6 +184,7 @@ typedef __m256d Tm;
 #define vfmaeq(a,b,c) a=_mm256_macc_pd(b,c,a)
 #define vfmaeq_s(a,b,c) a=_mm256_macc_ps(b,c,a)
 #define vfmseq(a,b,c) a=_mm256_nmacc_pd(b,c,a)
+#define vabmc(a,b,c) _mm256_msub_pd(a,b,c)
 #define vfmaaeq(a,b,c,d,e) a=_mm256_macc_pd(d,e,_mm256_macc_pd(b,c,a))
 #define vfmaseq(a,b,c,d,e) a=_mm256_nmacc_pd(d,e,_mm256_macc_pd(b,c,a))
 #else
@@ -189,12 +192,14 @@ typedef __m256d Tm;
 #define vfmaeq(a,b,c) a=_mm256_fmadd_pd(b,c,a)
 #define vfmaeq_s(a,b,c) a=_mm256_fmadd_ps(b,c,a)
 #define vfmseq(a,b,c) a=_mm256_fnmadd_pd(b,c,a)
+#define vabmc(a,b,c) _mm256_fmsub_pd(a,b,c)
 #define vfmaaeq(a,b,c,d,e) a=_mm256_fmadd_pd(d,e,_mm256_fmadd_pd(b,c,a))
 #define vfmaseq(a,b,c,d,e) a=_mm256_fnmadd_pd(d,e,_mm256_fmadd_pd(b,c,a))
 #else
 #define vfmaeq(a,b,c) a=_mm256_add_pd(a,_mm256_mul_pd(b,c))
 #define vfmaeq_s(a,b,c) a=_mm256_add_ps(a,_mm256_mul_ps(b,c))
 #define vfmseq(a,b,c) a=_mm256_sub_pd(a,_mm256_mul_pd(b,c))
+#define vabmc(a,b,c) _mm256_sub_pd(_mm256_mul_pd(a,b),c)
 #define vfmaaeq(a,b,c,d,e) \
   a=_mm256_add_pd(a,_mm256_add_pd(_mm256_mul_pd(b,c),_mm256_mul_pd(d,e)))
 #define vfmaseq(a,b,c,d,e) \
@@ -241,7 +246,8 @@ typedef __mmask8 Tm;
 #define vmuleq(a,b) a=_mm512_mul_pd(a,b)
 #define vmuleq_mask(mask,a,b) a=_mm512_mask_mul_pd(a,mask,a,b);
 #define vfmaeq(a,b,c) a=_mm512_fmadd_pd(b,c,a)
-#define vfmseq(a,b,c) a=_mm512_fnmadd_pd(b,c,a)
+//#define vabmc(a,b,c) a=_mm512_fnmadd_pd(b,c,a)
+//#define vfms(a,b,c) _mm512_fnmadd_pd(b,c,a)
 #define vfmaaeq(a,b,c,d,e) a=_mm512_fmadd_pd(d,e,_mm512_fmadd_pd(b,c,a))
 #define vfmaseq(a,b,c,d,e) a=_mm512_fnmadd_pd(d,e,_mm512_fmadd_pd(b,c,a))
 #define vneg(a) _mm512_mul_pd(a,_mm512_set1_pd(-1.))