[Bug 38277] Too slow test Advanced Pixel Shaders in 3DMark2001SE

[wine-bugs at winehq.org]

https://bugs.winehq.org/show_bug.cgi?id=38277

Sergey Isakov <isakov-sl at bk.ru> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 OS|Linux                       |Mac OS X

--- Comment #1 from Sergey Isakov <isakov-sl at bk.ru> ---
Hi sirs,
Several day I found the reason and finally got it.
The explanation is in Crossover source

    /* MacOS dx9 GPU drivers more GLSL vertex shader uniforms than supported by
the hardware, and if
     * more are used it falls back to software. While the compiler can detect
if the shader uses all
     * declared uniforms, the optimization fails if the shader uses relative
addressing. So any GLSL
     * shader using relative addressing falls back to software.
     *
     * ARB vp gives the correct amount of uniforms, so use it instead of GLSL.
     *
     * In addition, AMD Radeon HD GPUs advertise > 256 constants and can
support this in HW, but the
     * driver nevertheless falls back to software if more than 256 are used.
This is fixed in MacOS
     * 10.8.3 */

-------
I can confirm, this is very essential note. It is why I have so poor
performance.
But copying the function quirk_arb_constants() into wine sources gives a crash.
The reason is the quirk_table[] is applied too late.
gl_info->limits.glsl_vs_float_constants will change after they used to fill
d3d_info.

So I call the function quirk_arb_constants() in procedure
wined3d_adapter_init_limits(gl_info).
The results:
Game2 Low details 41.4
High details  28.2
Vertex shaders 29.9
Advanced Pixel Shaders 39.2 !!!
Now my engines fastest in all tests.
But my method is not good because it unconditional. I should check as CX did in
procedure match_apple_broken_uniforms()

-- 
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.