[arm64] bug fix for fma and cast in checkrounding test

edf-hpc · Feb 13, 2024 · b58a4c5 · b58a4c5
1 parent aeab80b
commit b58a4c5
Show file tree

Hide file tree

Showing 2 changed files with 32 additions and 3 deletions.
diff --git a/unitTest/checkRounding/checkRounding.cxx b/unitTest/checkRounding/checkRounding.cxx
@@ -397,12 +397,23 @@ class testInvariantProdDivm:public test<REALTYPE>{
   inline double myFma(const double& a, const double& b, const double& c){
     double d;
 #ifdef TEST_FMA
+#if defined(__X86_64__)
     __m128d ai, bi,ci,di;
     ai = _mm_load_sd(&a);
     bi = _mm_load_sd(&b);
     ci = _mm_load_sd(&c);
     di=_mm_fmadd_sd(ai,bi,ci);
     d=_mm_cvtsd_f64(di);
+#elif  defined(__aarch64__)
+  const float64x1_t ai=vld1_f64(&a);
+  const float64x1_t bi=vld1_f64(&b);
+  const float64x1_t ci=vld1_f64(&c);
+  const float64x1_t di=vfma_f64(ci,ai,bi);// warning strange argument order
+  // cf doc : https://developer.arm.com/architectures/instruction-set/intrinsics/#q=vfma
+  vst1_f64(&d, di);
+#else
+#error "notyet implemented for this architecture"
+#endif
 #else
     d=a*b+c;
 #endif
@@ -413,12 +424,30 @@ class testInvariantProdDivm:public test<REALTYPE>{
   inline float myFma(const float& a, const float& b, const float& c){
     float d;
 #ifdef TEST_FMA
+#if defined(__X86_64__)
     __m128 ai, bi,ci,di;
     ai = _mm_load_ss(&a);
     bi = _mm_load_ss(&b);
     ci = _mm_load_ss(&c);
     di=_mm_fmadd_ss(ai,bi,ci);
     d=_mm_cvtss_f32(di);
+#elif  defined(__aarch64__)
+  float av[2]={a,0};
+  float bv[2]={b,0};
+  float cv[2]={c,0};
+
+  float32x2_t ap=vld1_f32(av);
+  float32x2_t bp=vld1_f32(bv);
+  float32x2_t cp=vld1_f32(cv);
+
+  float32x2_t resp= vfma_f32(cp,ap,bp); // warning strange argument order
+  // cf doc : https://developer.arm.com/architectures/instruction-set/intrinsics/#q=vfma
+  float res[2];
+  vst1_f32(res, resp);
+  d=res[0];
+#else
+#error "notyet implemented for this architecture"
+#endif
 #else
     d=a*b+c;
 #endif
@@ -654,7 +683,7 @@ class testCast:public test<REALTYPE,REALTYPEREF>{
   }
 
   REALTYPE compute(){
-    REALTYPEREF ref=0.1;
+    volatile REALTYPEREF ref=0.1;
     return ((REALTYPE)ref);
   }
 };
@@ -670,7 +699,7 @@ class testCastm:public test<REALTYPE,REALTYPEREF>{
   }
 
   REALTYPE compute(){
-    REALTYPEREF ref=-0.1;
+    volatile REALTYPEREF ref=-0.1;
     return ((REALTYPE)ref);
   }
 };

diff --git a/unitTest/nocheckExemples/cast.cxx b/unitTest/nocheckExemples/cast.cxx
@@ -3,7 +3,7 @@
 #include <cstdlib>
 
 
-#include  <immintrin.h>
+//#include  <immintrin.h>
 //#include <avxintrin_emu.h>
 int main(int argc, char** argv){
   double a;