On Tue, Jul 25, 2017 at 8:12 AM, Roland Scheidegger <srol...@vmware.com> wrote: > Am 24.07.2017 um 11:44 schrieb Michel Dänzer: >> On 14/07/17 10:01 PM, Marek Olšák wrote: >>> Reviewed-by: Marek Olšák <marek.ol...@amd.com <mailto:marek.ol...@amd.com>> >> >> This change broke piglit spec@ext_texture_lod_bias@lodbias for me with >> radeonsi (but not with llvmpipe). >> >> > > The S_FIXED function needs some fixing (for r600 too). The compiler > takes full advantage of the fact that float to unsigned conversion is > undefined for negative floats, apparently. > The previous CLAMP would have used the state->lod_bias value if that was > equal to the MIN value (-16) (so no way for the compiler to screw it up) > whereas now it will use the -16 value. > (This also means it was probably broken before, for values which were > smaller than the min, albeit possibly noone cared as that would have > been outside the advertized lod bias limits.) > FWIW gcc generates a really terrible code for this (but that's not > really dependent on the new or old macro...) > (There's the same macro in mesa/main macros, but with signed numbers.)
https://patchwork.freedesktop.org/patch/168931/ Marek _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
