WIP

Author: HenryHRich

jsrc/ar.c

@@ -228,7 +228,7 @@ REDUCCPFX(tymesinsO, D, I, TYMESO)
   acc1=prim(acc1,acc5); acc2=prim(acc2,acc6); acc3=prim(acc3,acc7); acc0=prim(acc0,acc4); \
   acc2=prim(acc2,acc3); acc0=prim(acc0,acc1); acc0=prim(acc0,acc2); /* combine accumulators vertically */ \
   acc0=prim(acc0,_mm256_permute4x64_pd(acc0,0b11111110)); acc0=prim(acc0,_mm256_permute_pd(acc0,0xf));   /* combine accumulators horizontally  01+=23, 0+=1 */ \
-  *(I*)z=_mm256_extract_epi64(_mm256_castpd_si256(acc0),0x0); ++z;  /* store the single result from 0 */ \
+  *z=_mm256_cvtsd_f64(acc0); ++z;  /* store the single result from 0 */ \
  )
 
 // f/ on rank>1, going down columns to save bandwidth
@@ -429,8 +429,8 @@ DF1(jtcompsum){
    c0=_mm256_add_pd(c0,_mm256_permute_pd(c0,0xf)); acc1=_mm256_permute_pd(acc0,0xf);   // combine c0+c1, acc1<-1
    TWOSUM(acc0,acc1,acc0,c1); c0=_mm256_add_pd(c0,c1);    // combine 0123, combine all low parts
    acc0=_mm256_add_pd(acc0,c0);  // add low parts back into high in case there is overlap
-   *(I*)zv=_mm256_extract_epi64(_mm256_castpd_si256(acc0),0x0); ++zv;  // store the single result
-//    _mm_storel_pd(zv++,_mm256_castpd256_pd128(acc0));
+   *zv=_mm256_cvtsd_f64(acc0); ++zv;  // store the single result
+// obsolete    _mm_storel_pd(zv++,_mm256_castpd256_pd128(acc0));
   }
  }else{
   // rank>1, going down columns to save bandwidth and add accuracy

jsrc/va2.c

@@ -794,10 +794,10 @@ static A jtva2(J jt,AD * RESTRICT a,AD * RESTRICT w,AD * RESTRICT self,UI allran
         ti * RESTRICT wv1=wv+dplen; wv1=j==1?wv:wv1; \
         oneprod2  \
         if(j>1){--j; _mm_storeu_pd(zv,_mm256_castpd256_pd128 (acc000)); _mm_storeu_pd(zv+ndpi,_mm256_castpd256_pd128 (acc100)); wv+=dplen; zv +=2;} \
-        else{*(I*)zv=_mm256_extract_epi64(_mm256_castpd_si256(acc000),0x0); *(I*)(zv+ndpi)=_mm256_extract_epi64(_mm256_castpd_si256(acc100),0x0);  zv+=1;} \
+        else{*zv=_mm256_cvtsd_f64(acc000); *(zv+ndpi)=_mm256_cvtsd_f64(acc100);  zv+=1;} \
        }else{ \
         oneprod1  \
-        *(I*)zv=_mm256_extract_epi64(_mm256_castpd_si256(acc000),0x0); \
+        *zv=_mm256_cvtsd_f64(acc000); \
         zv+=1; \
        } \
        if(!--j)break; \
@@ -932,7 +932,7 @@ static A jtva2(J jt,AD * RESTRICT a,AD * RESTRICT w,AD * RESTRICT self,UI allran
  acc3=MUL_ACC(acc3,_mm256_maskload_pd(av,endmask),_mm256_maskload_pd(wv,endmask)); av+=((dplen-1)&(NPAR-1))+1;  wv+=((dplen-1)&(NPAR-1))+1; \
  acc0=_mm256_add_pd(acc0,acc1); acc2=_mm256_add_pd(acc2,acc3); acc0=_mm256_add_pd(acc0,acc2); /* combine accumulators vertically */ \
  acc0=_mm256_add_pd(acc0,_mm256_permute4x64_pd(acc0,0b11111110)); acc0=_mm256_add_pd(acc0,_mm256_permute_pd(acc0,0xf));   /* combine accumulators horizontally  01+=23, 0+=1 */ \
- *(I*)zv=_mm256_extract_epi64(_mm256_castpd_si256(acc0),0x0); ++zv;
+ *zv=_mm256_cvtsd_f64(acc0); ++zv;
 #else
 #define ONEPRODD D total0=0.0; D total1=0.0; if(dplen&1)total1=(D)*av++*(D)*wv++; DQ(dplen>>1, total0+=(D)*av++*(D)*wv++; total1+=(D)*av++*(D)*wv++;); *zv++=total0+total1;
 #endif
@@ -1100,7 +1100,7 @@ DF2(jtsumattymes1){
       acc0=_mm256_add_pd(acc0,_mm256_permute_pd(acc0,0xf));
       acc0=_mm256_add_pd(acc0,c0);  // add low parts back into high in case there is overlap
 #endif
-      *(I*)zv=_mm256_extract_epi64(_mm256_castpd_si256(acc0),0x0); ++zv;
+      *zv=_mm256_cvtsd_f64(acc0); ++zv;
       if(!--j)break; av=av0;  // repeat a if needed
      }
     }