Virtualization
/
berkeley-softfloat-3
mirror of https://github.com/ucb-bar/berkeley-softfloat-3.git
1
0
Fork 0
Code Issues Projects Releases Wiki Activity

fixing bf16_to_f32 for subnormal numbers

This commit is contained in:
Nicolas Brunie 2023-09-22 18:30:13 -07:00
parent 577fcd787c
commit b1bbccf04d
1 changed files with 5 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
source/bf16_to_f32.c
View File

@ -74,17 +74,11 @@ float32_t bf16_to_f32( bfloat16_t a )
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
if ( ! exp ) {
if ( ! frac ) {
uiZ = packToF32UI( sign, 0, 0 );
goto uiZ;
}
normExpSig = softfloat_normSubnormalBF16Sig( frac );
exp = normExpSig.exp - 1;
frac = normExpSig.sig;
}
/*------------------------------------------------------------------------
*------------------------------------------------------------------------*/
// packToF32UI simply packs bitfields without any numerical change
// which means it can be used directly for any BF16 to f32 conversions which
// does not require bits manipulation
// (that is everything where the 16-bit are just padded right with 16 zeros, including
// subnormal numbers)
uiZ = packToF32UI( sign, exp, ((uint_fast32_t) frac) <<16 );
uiZ:
uZ.ui = uiZ;