Bug 1265 - EXA for tdfx
Summary: EXA for tdfx
Status: RESOLVED WONTFIX
Alias: None
Product: xorg
Classification: Unclassified
Component: Driver/tdfx (show other bugs)
Version: git
Hardware: x86 (IA32) Linux (All)
: medium enhancement
Assignee: Corbin Simpson
QA Contact: Xorg Project Team
URL:
Whiteboard: 2011BRB_Reviewed
Keywords: patch
Depends on:
Blocks:
 
Reported: 2004-09-01 00:27 UTC by Adam Jackson
Modified: 2014-11-17 21:34 UTC (History)
3 users (show)

See Also:
i915 platform:
i915 features:


Attachments
tdfx-render-accel-1.patch (8.81 KB, patch)
2004-09-01 00:27 UTC, Adam Jackson
no flags Details | Splinter Review
tdfx-render-accel-2.patch (9.37 KB, patch)
2004-09-01 21:47 UTC, Adam Jackson
no flags Details | Splinter Review
tdfx-render-4.patch (14.12 KB, patch)
2004-09-27 10:50 UTC, Adam Jackson
no flags Details | Splinter Review

Description Adam Jackson 2004-09-01 00:27:19 UTC
straight port of the Render accel in voodoo(4) to tdfx(4).  passes rendercheck
at both 16 and 24 bpp.  should be free of endianness issues but only tested on x86.

this is probably broken when used with DRI, we touch some 3d registers but never
take a hardware lock or dirty the 3d state.  quake3 still works, but we might
render incorrectly when doing both Render and GL simultaneously.

thanks to Alan Cox for the voodoo driver.
Comment 1 Adam Jackson 2004-09-01 00:27:54 UTC
Created attachment 800 [details] [review]
tdfx-render-accel-1.patch
Comment 2 Adam Jackson 2004-09-01 21:47:52 UTC
Created attachment 812 [details] [review]
tdfx-render-accel-2.patch

first attempt was a no-op, i was using the old setup hooks.  the voodoo driver
needs to be converted to the new hooks too.

this one actually has an effect.  passes rendercheck in 16bpp but not 24bpp,
and unscaled over blends get a 17x speedup according to renderbench.  xcompmgr
-a is entirely usable; -c is broken, but i suspect that's an unrelated bug in
the driver since i'm not seeing the same problem on other hardware.
Comment 3 Eric Anholt 2004-09-09 12:29:54 UTC
A few notes:
- Need to multiply source by alpha in your PictOpSrc case, since you haven't
premultiplied.
- Syncing needs to be done before writing anything with cpu to framebuffer, so
you appear to be correct.
- The DRI lock is taken by the server on wakeup and dropped on block, so your
driver doesn't need to worry about it.  Since you don't appear to rely on any
new set-once state, you don't need to worry about that either.  However, I don't
remember if tdfx had a set of flags in the shared area for sets of context
dirtied.  If so, just dumping 0xffffffff to it is probably the best idea.
- I'm concerned about the x8r8g8b8 format support, but I haven't read the code
thoroughly.  Remember that you have to treat the x bits as 1.  That means that
the XXX should be srcFormat comment appears correct.  But what about x8r8g8b8
src to a8r8g8b8 dest -- do the bits get filled with 1s?

Comment 4 Adam Jackson 2004-09-27 10:50:09 UTC
Created attachment 971 [details] [review]
tdfx-render-4.patch

this version eliminates the magic numbers and accelerates just about all the
simple blend ops.

problem is, none of them work, except for Dst (because i short circuit it) and
Src and In (which with da=1 are equivalent to normal framebuffer writes).  not
really sure what the deal is there.
Comment 5 René Rhéaume 2005-01-26 11:45:26 UTC
I own an antiquated Voodoo3 2000 PCI and still use it as my primary video card
(My next one will sit on a PCI Express slot and will be in my next computer). I
also built my own copy of Xorg-X11 with 19 patches selected from Gentoo,
Mandrake and RedHat Linux (these distinutions contains a lot more), running
smoothly for one month and an half.

What is the status of the patch? I would like to test but I read there is a
conflict between DRI and Render. Also, are there other issues?
Comment 6 Adam Jackson 2005-01-26 11:59:28 UTC
the patch is still somewhat in limbo due to the correctness issues i already
described.  it gives a nice speedup but doesn't render much of anything
correctly.  i'll come back to it sometime before 6.9 probably.  rest assured
it'll be merged once it's correct.

if someone wanted to take a shot at finishing it, the voodoo3 datasheets are
publicly available and should be sufficient.
Comment 7 Daniel Stone 2007-02-27 01:24:02 UTC
Sorry about the phenomenal bug spam, guys.  Adding xorg-team@ to the QA contact so bugs don't get lost in future.
Comment 8 Corbin Simpson 2010-03-26 06:29:42 UTC
Adding myself to the CC. I have a Voodoo3 now, and should be able to finish off this ancient monstrosity if nobody's already done so.
Comment 9 Jeremy Huddleston Sequoia 2011-10-03 09:56:26 UTC
Corbin, do you still plan to do this?
Comment 10 Adam Jackson 2014-11-17 21:34:57 UTC
Let's not pretend this is actually getting worked on.


Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.