Am 11.09.2015 um 05:14 schrieb Dave Airlie: > On 11 September 2015 at 12:37, Roland Scheidegger <srol...@vmware.com> wrote: >> Just on a very quick glance, seems somewhat odd this function is named >> "cypress" whereas the other 64bit ones are named "cayman". >> In any case, saying enabling support for chips which have hative support >> isn'quite correct, since there's a couple more which should be able to >> do it (rv670, rv770, maybe more) as far as I know (no idea though how >> complex it would be to implement, maybe they are missing some things...) > > probably should say evergreen native, > > I doubt anyone will care about rv670/770, their fp64 support is pretty lacking > compared to evergreen. We really only care on evergreen because GL4.0, > and none of the older gpus can do GL4.0 >
True enough. I'm kind of surprised the blob gets away with the uber-cheap cheated implementation of the int to double / double to int conversion, needless to say that probably doesn't get you quite the results you'd expect and doesn't look very conformant to me... At least for int to double, a better implementation wouldn't be all that tricky, just convert upper and lower bits separately and add them back together (rescaled of course), albeit of course quite a few more instructions. The reverse probably could work similarly too, but anyway if the blob is happy with the cheat I guess that's good enough... Roland _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
