Commen=
t # 14
on bug 95528<=
/a>
from Daniel Scharrer
Created attachment 124721 [details]
Trimmed R600_DEBUG=3Dvs,ps output & shader GL calls with LLVM r273782
I could not record an apitrace (game hangs before getting to the main menu,
possibly performance related), but I was able to dump the shaders and narrow
down (some of) the affected ones.
The game uses two passes to render the flags and banners: The first pass
(possibly a z prepass?) uses pipeline=3D20 in the attached logs (vertex
shader=3D73, program=3D74; fragment shader=3D89, program=3D90).
The second pass is depth-tested against the results of the first, but uses
different shaders - pipeline=3D105 (vertex shader=3D383 program=3D384; frag=
ment
shader=3D381, program=3D382).
With LLVM r266378 reverted, the results of the two passes line up, with cur=
rent
LLVM they do not causing the z-fighting visible in the video linked in the =
bug
description. Even when disabling the depth test in the second pass, there a=
re
still visible parts of "white" background so this is a) most like=
ly an issue
with one of the vertex shaders and b) not just a depth precision issue. You=
can
also see this by looking closely at the flags in the video - the while (hol=
es
punched by the z-prepass) and textured parts sway differently.
Since the game compiles an enormous amount of shaders on startup, the full=
=20
R600_DEBUG=3Dcs,vs,tcs,tes,gs,ps output is quite large - I have attached a
trimmed output that only contains the two vertex/fragment shader pairs used=
for
the banners. The full logs are available at:
http://const=
expr.org/tmp/bioshockinfinite-bad.log (20 MiB)
http://cons=
texpr.org/tmp/bioshockinfinite-good.log (21 MiB)