Optimize exp and log

1 files changed, 24 insertions, 0 deletions

diff --git a/mathfuncs_exp.h b/mathfuncs_exp.h
index ceb89c0..99fd063 100644
--- a/mathfuncs_exp.h
+++ b/mathfuncs_exp.h
@@ -35,6 +35,7 @@ namespace vecmathlib {
     realvec_t r;
     switch (sizeof(real_t)) {
     case 4:
+#ifdef VML_HAVE_FP_CONTRACT
       // float, error=4.55549108005200277750378992345e-9
       r = RV(0.000154653240842602623787395880898);
       r = fma(r, x, RV(0.00133952915439234389712105060319));
@@ -43,8 +44,18 @@ namespace vecmathlib {
       r = fma(r, x, RV(0.240226511015459465468737123346));
       r = fma(r, x, RV(0.69314720007380208630542805293));
       r = fma(r, x, RV(0.99999999997182023878745628977));
+#else
+      // float, error=1.62772721960621336664735896836e-7
+      r = RV(0.00133952915439234389712105060319);
+      r = fma(r, x, RV(0.009670773148229417605024318985));
+      r = fma(r, x, RV(0.055503406540531310853149866446));
+      r = fma(r, x, RV(0.240222115700585316818177639177));
+      r = fma(r, x, RV(0.69314720007380208630542805293));
+      r = fma(r, x, RV(1.00000005230745711373079206024));
+#endif
       break;
     case 8:
+#ifdef VML_HAVE_FP_CONTRACT
       // double, error=9.32016781355638010975628074746e-18
       r = RV(4.45623165388261696886670014471e-10);
       r = fma(r, x, RV(7.0733589360775271430968224806e-9));
@@ -58,6 +69,19 @@ namespace vecmathlib {
       r = fma(r, x, RV(0.240226506959101382690753994082));
       r = fma(r, x, RV(0.69314718055994530864272481773));
       r = fma(r, x, RV(0.9999999999999999978508676375));
+#else
+      // double, error=3.74939899823302048807873981077e-14
+      r = RV(1.02072375599725694063203809188e-7);
+      r = fma(r, x, RV(1.32573274434801314145133004073e-6));
+      r = fma(r, x, RV(0.0000152526647170731944840736190013));
+      r = fma(r, x, RV(0.000154034441925859828261898614555));
+      r = fma(r, x, RV(0.00133335582175770747495287552557));
+      r = fma(r, x, RV(0.0096181291794939392517233403183));
+      r = fma(r, x, RV(0.055504108664525029438908798685));
+      r = fma(r, x, RV(0.240226506957026959772247598695));
+      r = fma(r, x, RV(0.6931471805599487321347668143));
+      r = fma(r, x, RV(1.00000000000000942892870993489));
+#endif
       break;
     default:
       __builtin_unreachable();