[PATCH 3/5] wined3d: Check relative addressing indices when accessing uniforms in GLSL shaders.

Thu Feb 11 15:03:36 CST 2016

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA256

> +                        sprintf(register_name, "(%s + %u >= 0 &&
> %s + %u < %u ? %s_c[%s + %u] : vec4(0.0))",
Did you do any testing on the performance impact of this? This seems
like something where it's convenient to sacrifice performance
everywhere for a corner case, but later on we wonder why games are
unplayably slow. Keep in mind that currently it's pretty difficult to
be GPU limited. With the CSMT patches you'll have a higher chance of
being GPU limited, but there's no guarantee of course.

dx10+ GPUs don't return 0.0 reliably on Windows here. My Geforce GPUs
returned the int constants at c[-32] to c[-48].

I think there are additional options of increasing our odds of
behaving like Windows. E.g. with constant buffers we could create a
buffer that matches the hardware page size (which we can't query, I
know) and fill everything except the 256 d3d9 constants with zeroes.
If the out of bounds read falls into the buffer we get a zero. If not,
the hardware should detect the invalid read and I'd expect it to
return zero due to exactly this d3d requirement.

Also I think d3d9 HW should "just work" in this regard because only
256 registers exist.

Yes, I am aware that there are more games than just The Witcher that
are affected by this problem.