Import initial version

1 files changed, 86 insertions, 0 deletions

diff --git a/mathfuncs_convert.h b/mathfuncs_convert.h
new file mode 100644
index 0000000..4dfffef
--- /dev/null
+++ b/mathfuncs_convert.h
@@ -0,0 +1,86 @@
+// -*-C++-*-
+
+#ifndef MATHFUNCS_CONVERT_H
+#define MATHFUNCS_CONVERT_H
+
+#include "mathfuncs_base.h"
+
+#include <cassert>
+#include <cmath>
+
+
+
+namespace vecmathlib {
+  
+  template<typename realvec_t>
+  realvec_t mathfuncs<realvec_t>::vml_convert_float(intvec_t x)
+  {
+    // Convert in two passes. Convert as much as possible during the
+    // first pass (lobits), so that the second pass (hibits) may be
+    // omitted if the high bits are known to be zero.
+    int_t lobits = FP::mantissa_bits;
+    int_t hibits = FP::bits - lobits;
+    
+    // Convert lower bits
+    intvec_t xlo = x & IV((U(1) << lobits) - 1);
+    // exponent for the equivalent floating point number
+    int_t exponent_lo = FP::exponent_offset + lobits;
+    xlo |= exponent_lo;
+    // subtract hidden mantissa bit
+    realvec_t flo = as_float(xlo) - RV(FP::as_float(exponent_lo));
+    
+    // Convert upper bits
+    // make unsigned by subtracting largest negative number
+    x ^= FP::sign_mask;
+    intvec_t xhi = lsr(x, lobits);
+    // exponent for the equivalent floating point number
+    int_t exponent_hi = FP::exponent_offset + FP::bits;
+    xhi |= exponent_hi;
+    // add largest negative number, and subtract hidden mantissa bit
+    realvec_t
+      fhi = as_float(xhi) - RV(FP::as_float(exponent_hi) + R(FP::sign_mask));
+    
+    // Combine results
+    return flo + fhi;
+  }
+  
+  
+  
+  template<typename realvec_t>
+  auto mathfuncs<realvec_t>::vml_convert_int(realvec_t x) -> intvec_t
+  {
+    // Handle zero
+    boolvec_t is_zero = x == RV(0.0);
+    // Handle negative numbers
+    boolvec_t is_negative = signbit(x);
+    x = fabs(x);
+    
+    // Round, by adding a large number that removes the excess
+    // precision
+    int_t large = U(1) << FP::mantissa_bits;
+    x += R(large);
+    
+    intvec_t exponent = ilogb(x);
+    for (int i=0; i<intvec_t::size; ++i) {
+      assert(exponent[i] >= FP::mantissa_bits);
+    }
+    intvec_t ix = as_int(x) & IV(FP::mantissa_mask);
+    // add hidden mantissa bit
+    ix |= U(1) << FP::mantissa_bits;
+    // shift according to exponent
+    ix <<= exponent - IV(FP::mantissa_bits);
+    
+    // Undo the adding above
+    ix -= large;
+    
+    // Handle negative numbers
+    ix = ifthen(is_negative, -ix, ix);
+    // Handle zero
+    ix = ifthen(is_zero, IV(0), ix);
+    
+    return ix;
+  }
+  
+}; // namespace vecmathlib
+
+#endif  // #ifndef MATHFUNCS_CONVERT_H