Add constructors from double for fp8 and fp16 types

2 files changed, 23 insertions, 0 deletions

diff --git a/floattypes.h b/floattypes.h
index 1830d1d..2c22a10 100644
--- a/floattypes.h
+++ b/floattypes.h
@@ -215,11 +215,15 @@ namespace vecmathlib {
   struct fp8 {
     // 1 bit sign, 4 bits exponent, 3 bits mantissa
     vml_std::uint8_t val;
+    fp8() {}
+    fp8(double x) { __builtin_unreachable(); }
   };
   
   struct fp16 {
     // 1 bit sign, 5 bits exponent, 10 bits mantissa
     vml_std::uint16_t val;
+    fp16() {}
+    fp16(double x) { __builtin_unreachable(); }
   };
   
 } // namespace vecmathlib
diff --git a/mathfuncs_fabs.h b/mathfuncs_fabs.h
index 4baec9b..4f31dec 100644
--- a/mathfuncs_fabs.h
+++ b/mathfuncs_fabs.h
@@ -166,9 +166,28 @@ namespace vecmathlib {
   realvec_t mathfuncs<realvec_t>::vml_ldexp(realvec_t x, intvec_t n)
   {
     // TODO: Check SLEEF 2.80 algorithm
+#if 0
     realvec_t r = as_float(as_int(x) + (n << I(FP::mantissa_bits)));
     r = ifthen((as_int(x) & IV(FP::exponent_mask)) == IV(I(0)), x, r);
     return r;
+#endif
+    realvec_t r = as_float(as_int(x) + (n << U(FP::mantissa_bits)));
+    int max_n = FP::max_exponent - FP::min_exponent;
+    boolvec_t underflow = n < IV(I(-max_n));
+    boolvec_t overflow = n > IV(I(max_n));
+    intvec_t old_exp =
+      lsr(as_int(x) & IV(FP::exponent_mask), FP::mantissa_bits);
+    intvec_t new_exp = old_exp + n;
+    // TODO: check bit patterns instead
+    underflow =
+      underflow || new_exp < IV(I(FP::min_exponent + FP::exponent_offset));
+    overflow =
+      overflow || new_exp > IV(I(FP::max_exponent + FP::exponent_offset));
+    r = ifthen(underflow, copysign(RV(R(0.0)), x), r);
+    r = ifthen(overflow, copysign(RV(FP::infinity()), x), r);
+    boolvec_t dont_change = x == RV(R(0.0)) || isinf(x) || isnan(x);
+    r = ifthen(dont_change, x, r);
+    return r;
   }
   
   template<typename realvec_t>