On 2002.06.19 20:40 Michael Thaler wrote:
> Hello Jose
> I hope to find the time to compile the mach64-0-0-4 branch at the
> weekend or at least early next week. I am still using the driver
> without dma with which I get about 30 fps playing UT or Quake3 in 640
> x 480. I am curious how the driver performs now.
> It would be quite good to make a short list of how one can measure the
> speed of the driver accurately under different settings. I don't know
> anything about graphics driver programming so I don't know how I can
> measure "dma speed" or other crucial things.
Michael, I think that the best measure remains the fps of the OpenGL
applications one currently use (glxgears, UT, tuxracer), as it's a
objective measure and easy to get in these applications.
The CPU load is also important:
- big idle slices means the card can't cope with with the requests
(not much to do here besides buying a new card, except perhaps trying
to improve the texture management)
- big kernel usage means that the DRM is not being effective (improve
the DMA code)
- big user space usage is a good sign as it means that neither the
card or the DRM are bottlenecks, but it can still mean that we could
try to optimize the Mesa driver (which is on the user side) to use less
To measure the effect of certain things (such as DMA, AGP bus, AGP
texturing, onboard memory) is just a matter of chaging _one_ of these
variables, and measuring the difference in fps. For example:
- compare the fps of glxgears one month ago and and now with the new
DMA code. - not load agpgart to see the effect of the AGP texturing
(if there is no effect then the problem is elsewhere)
- use small screen resolutions to see the effect of more memory (if
there is no effect then the problem is elsewhere)
And so on...
Later on I'll try to make benchmarks in apps such as glperf and
viewperf, but for now it's too much trouble for me since there is quite
alot left to do.