[Dri-devel] Mach64 offscreen/texture memory

[Leif D. <lde...@re...>]

I've commited some changes to AGP and offscreen memory management for
mach64, which are available now in CVS and the latest binary snapshots for
testing.  I'm particularly interested in finding out if the AGP changes
(the second item below) are unstable for anyone, relative to the previous
code.  Here's a summary of the changes:

- Allocate/free back, depth, and texture memory when transitioning from 2d 
to 3d and back.  (Taken from tdfx and the reinit branch).  Note that this 
doesn't change the static allocation requirement for useable modes, since 
the buffers are still shared by all contexts and the buffers are not 
resized for mode switches.  However, when no GL contexts are active, the 
offscreen mem. allocated (for XAA) is the same as with DRI disabled.

- Use only AGP for textures when available.  Use of a local texture heap
in addition to AGP can be forced with Option "LocalTextures".  PCI cards
(and AGP cards with Option "ForcePCIMode") still use only a local texture
heap.  This means that a few modes will be useable only on AGP cards:

8MB card: 1024x768 @ 24-bit depth, 1280x1024 @ 16-bit depth
6MB card: 1152x864 @ 16-bit depth
(if there's such a thing as a 2MB AGP card: 640x480@16)

This change increases the XAA cache for AGP cards when GL contexts are
active, relative to the previous code, and should reduce the amount of 
texture swapping for multitexturing.

- To try to reduce the performance impact of using AGP only, AGP is
initialized to use the default mode set by agpgart (Option "AGPMode" can
still be used to force 1x or 2x), rather than always defaulting to 1x.  
Also, at 1x the card is configured to generate all AGP reads as high
priority.  This setting was always used in the Utah driver, but it's
unstable for me at 2x, so I've only enabled it at 1x.  On my system the
end result of the changes is a slight increase in performance at 2x, and a
slight decrease in performance at 1x, relative to the previous code (the
difference is <1 FPS in quake3).

- New texture/2D cache allocation algorithm:  Instead of a fixed 128
scanlines for XAA cache, try for a full viewport or enough for
single-buffered DVD (will be needed for Xvideo in the future), whichever
is larger.  If that fails or gives less than half the available offscreen
mem. for textures (after front/back/depth buffers), split the available
mem.  between textures and XAA cache.  If there's not enough space for 2
256x256 textures in the texture heap at once, DRI is disabled.  These
changes decrease the local texture memory used, but shouldn't change the
usable modes and should help 2D acceleration.

So, happy testing, and please post to dri-devel with results if you try 
it out.

-- 
Leif Delgass 
http://www.retinalburn.net