aaron/ispc
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Small improvement to float->half function in stdlib.

Browse Source 
Rewrite things to be able to do a float MINPS, for slightly
better code on SSE2 (which has that but not an signed int
min).  SSE2 code now 23 instructions (vs 21 intrinsics).
This commit is contained in:
Matt Pharr
2012-03-23 16:09:32 -07:00
parent 3bb2dee275
commit 7e18f0e247
1 changed files with 3 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
stdlib.ispc
View File

@@ -3551,8 +3551,9 @@ static inline int16 float_to_half(float f) {
// like recursive filters in DSP - not a typical half-float application. Whether // like recursive filters in DSP - not a typical half-float application. Whether
// FP16 denormals are rare in practice, I don't know. Whatever slow path your HW // FP16 denormals are rare in practice, I don't know. Whatever slow path your HW
// may or may not have for denormals, this may well hit it. // may or may not have for denormals, this may well hit it.
int32 fint2 = intbits(floatbits(fint & round_mask) * floatbits(magic)) - round_mask; float fscale = floatbits(fint & round_mask) * floatbits(magic);
fint2 = (fint2 > f16infty) ? f16infty : fint2; // Clamp to signed infinity if overflowed fscale = min(fscale, floatbits((31 << 23) - 0x1000));
int32 fint2 = intbits(fscale) - round_mask;
if (fint < f32infty) if (fint < f32infty)
o = fint2 >> 13; // Take the bits! o = fint2 >> 13; // Take the bits!