View Single Post
Old 08-22-2012, 09:35 AM   #121
knc1
Going Viral
knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.knc1 ought to be getting tired of karma fortunes by now.
 
knc1's Avatar
 
Posts: 17,212
Karma: 18210809
Join Date: Feb 2012
Location: Central Texas
Device: No K1, PW2, KV, KOA
Holy War challenge accepted ::

K4&K5 == -mfpu=vfpv3-d16 -mfloat-abi=softfp (rather than: -mfpu=vfpv3 or -mfpu=neon)

Keep in mind that the neon instruction set treats denormals as zero.
Should not make a difference in typical uses, like finding the pixel bounds of a graphic item.
Will make a difference in fft and inverse fft code (big time).
Would probably not be recommended for building gcc with either - very likely to screw up the optimizer's tree coloring.

Hmm...
I wouldn't trust the neon instruction set for building Lua either (Lua's only (default) internal numeric type is double)

Note: Only messing with the K3 code at the moment here, not (yet) with K4&K5.

Last edited by knc1; 08-22-2012 at 09:39 AM.
knc1 is offline   Reply With Quote