Pixel Processing Problems: On the Road to Pixel Perfection

[MayImilae]

New article on the Dolphin Blog: https://dolphin-emu.org/blog/2014/03/15/pixel-processing-problems/

Feel free to discuss the article on this thread.

[delroth]

Let the performance decrease whining begin!

[Dolp]

I'm curious, how does this affect Nvidia Maxwell, the newest GPU architecture from Nvidia.

http://devblogs.nvidia.com/parallelforall/5-things-you-should-know-about-new-maxwell-gpu-architecture/

Also Nvidia is phasing out driver support for the DX10 era GPUs by April 2016, so it's time to retire those GPUs if anyone still uses them.

http://nvidia.custhelp.com/app/answers/detail/a_id/3473

Quote:After Release 340, any subsequent Windows driver release starting with Release 343 will cease to support the products listed in this section.

The Release 340 drivers will continue to support these products until April 1, 2016, and the NVIDIA support team will continue to address driver issues for these products in driver branches up to and including Release 340. However, future driver enhancements and optimizations in driver releases after Release 340 will not support these products.

[Shonumi]

Great article yet again. You don't get insight into what it takes to emulate a GC/Wii anywhere else (except the IRC channel).

Very interesting to see that this float vs. int issue affects vertex processing. Hopefully that's what's at issue with Mario Tennis and its extruding vertices in some character models. Great work by neobrain et al!

[neobrain]

By the way, huge thanks to JMC47 and MaJoR. IMO the article ended up being really awesome, and without them it wouldn't have nearly been half as good. So... I really appreciate the work you guys put into this and in particular the patience you have shown when having to deal with my pedantry all the time

(03-16-2014, 02:33 AM)Shonumi Wrote: Very interesting to see that this float vs. int issue affects vertex processing. Hopefully that's what's at issue with Mario Tennis and its extruding vertices in some character models. Great work by neobrain et al!

I think the Mario Tennis issue was solved by a combination of tev_fixes_new and zfreeze (JMC probably knows for sure). Either way, I've recently started writing a hardware test suite for figuring lots of pesky details about the hardware, and hopefully this will help resolve issues like this in the future.

[NaturalViolence]

Good use of alliteration, +1

[Oehr]

i love to read these articles. great to see another milestone approaching.

Still, its funny to read "performance decrease of ~5% or more" when simplifying calculations like

Quote:value = value - 2.0 * round(0.5 * value * (255.0/256.0)) * (256.0/255.0)

to basically

Quote:value = value

Still: Accuracy is more important than speed!

another lesson i learned is, that nvidia can suck it. once my gtx480 dies, im getting a amd card.
at least then, i get usable opencl and fast opengl, as long as I stick to windows...

[NaturalViolence]

What's wrong with nvidias openCL support?

So far dolphin is the only openGL application I know of that performs better on comparable AMD cards than nvidia. I'm willing to bet this is just a driver issue that will be fixed soon if nvidia is made aware of it.

[AnyOldName3]

I think Neobrain said his nVidia contact said there appeared to be such a small performance drop on D3D compared to what was expected because the driver was deciding to emulate some of the integer ops in FP (but seemed to do it better than dolphin did as it didn't come with the side effects). If this is the case, it could well be that the OGL driver eventually starts doing this too.

[JMC47]

Prior to this merge, OGL was far, far better on NVIDIAs. Now D3D is faster on both NVIDIA and AMD cards.