[Bug 34266] 0*inf in D3D9 shaders outputs NaNs to floating-point framebuffers

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=34266

--- Comment #34 from Paul Gofman <gofmanp at gmail.com> ---
A patch with two distinct workarounds for this bug has been added to Staging.
They are controlled by HKCU\\Software\\Wine\\Direct3D\\multiply_special
registry value (DWORD value). None of the workarounds are enabled by default.
Both workarounds have no effect on shaders 4.0 and higher (that is, Directx
10+), the workarounds concern Directx 8 and 9 shaders only.

1. Setting multiply_special to 1 avoids infinity in RCP, RSQ and LOG
instructions (returns [-]FLT_MAX instead of infinity). This is supposed to
avoid rendering problems in most, if not all, games affected by this bug. Yet
there can potentially exist games which suffer from this bug but for which this
won't help. It can happen if, e. g., the game gets infinity through floating
point overflow which is later multiplied by zero, or passes infinity or nan
through shader constants. I observed the latter in some games but ironically
those were not affected by this bug.
     This approach can have some impact on shader performance, but the impact
is supposed to be not very big. It is also potentially possible that some games
which are not affected by this bug will regress when using it, though I am not
aware of real examples of that.


2. Setting multiply_special to 2 wraps each multiplication and dot product in
shader with a check for zero operand, so that 0 * something result is always 0.
This is supposed to workaround this bug more or less universally. But this
approach is disastrous for shader performance. So it is mostly meant for
debugging purposes. Yet maybe if the game is not GPU bound on the given
hardware this mode can be used to play it, if setting multiply_special to 1
does not help.


What is particularly interesting for testing is:

1. Are there any games which are fixed by setting multiply_special to 2 but not
to 1?

2. Are there any games (not affected by this bug) which have rendering problems
with multiply_special set to 1?

Any reports on that are appreciated.

-- 
Do not reply to this email, post in Bugzilla using the
above URL to reply.
You are receiving this mail because:
You are watching all bug changes.