Re: Re: how to write snappy benchmark?

On Tue, Jul 08, 2003 at 10:57:20AM +0100, Torgeir Veimo wrote:
| Can you provide some details? The SGI way is to virtualize all gfx
| hardware, while DRI is a hardware acces arbitrator?

All direct rendering systems virtualize the graphics hardware to some
extent.  The mechanisms vary.

The DRI was intended to support fully virtualized graphics as well as
partly virtualized systems on which the most you can do is arbitrate
access.  Here's the original high-level design document that talks about
the goals:

	http://dri.sourceforge.net/doc/design_high_level.html

Here's the best publically-available description of the direct rendering
approach that SGI used in its high-end systems:

	http://www.opengl.org/developers/documentation/Direct/direct.html

(Incidentally, it mentions the HP Starbase implementation that
influenced the DRI.)  Also, the references are useful.  The only seminal
paper that comes to mind that I don't see in the list is the GTX
description:

	Kurt Akeley, "The Silicon Graphics 4D/240GTX Superworkstation,"
	IEEE Computer Graphics and Applications, July 1989, pp. 71-83.

That one's particularly interesting because it describes the clever
three-way data transfer mechanism that SGI used to build systems that
had high throughput, low latency, and good CPU cache behavior.

Linas Vepstas (formerly of IBM's AIX graphics group) wrote an
introduction to direct rendering issues that also has a good reference
list:

	http://linas.org/linux/graphics.html

Because of the business models in the PC world, companies are a lot less
willing to describe their implementation techniques than was the case in
the workstation world.  I haven't seen any public description of the
direct rendering architectures for 3Dlabs, ATI, or NVIDIA.  If anyone
has references, please pass them along.

Allen