How about some Quake? 😈

Perhaps a BIOS hack is needed to fully utilise write allocate on the K5, or perhaps some minor hardware mode? Or maybe it just doesn't do much for the K5...

Nice! K5 PR200 is super rare isn't it? Like never officially released?

feipoa wrote:

Perhaps a BIOS hack is needed to fully utilise write allocate on the K5, or perhaps some minor hardware mode? Or maybe it just doesn't do much for the K5...

The results seemed entirely positive in Windows. Even the OpenGL and D3D performance improved a little in Wintune.

I need to look at some other DOS games. The Wing Commander III slowdown boggles... I tried some other parameters on K5.sys too but nothing improved it. Without the write allocate enabled, WC3 actually rates the K5 PR200 + Verite setup as maximum CPU and video performance (a rating of 0 for both). I think the 3D engine is based on integer arithmetic so it should perform very well on K5.

Artex wrote:

Nice! K5 PR200 is super rare isn't it? Like never officially released?

I got in on a CPU-World group buy of some stock of the PR200 around 8 years ago. So there were some made but yeah it's not exactly commonplace.

Try some other 3D DOS games, MDK perhaps?

OK I've tested Quake. It also gets slightly slower. In both DOS and Windows. I also tried using the Powerleap utility instead of K5.sys.

What is the percent slowdown in Quake (software mode)?

Is there a thing for setting memory regions, or is that only a Cyrix tweak - I'm thinking maybe the graphics memory might be getting tangled up with the write allocate

Matth79 wrote:

Is there a thing for setting memory regions, or is that only a Cyrix tweak - I'm thinking maybe the graphics memory might be getting tangled up with the write allocate

Is that the write gathering feature(s)?

Matth79 wrote:

Is there a thing for setting memory regions, or is that only a Cyrix tweak - I'm thinking maybe the graphics memory might be getting tangled up with the write allocate

There is with the K5.sys driver.

It's all Japanese but take a look at the text file attached. I'm not sure what to do with the fixed region options....

There is also an official AMD document about the registers.
http://bitsavers.trailing-edge.com/pdf/amd/K8 … _Note_Mar97.pdf

feipoa wrote:

What is the percent slowdown in Quake (software mode)?

It was only around 1fps loss, with scores around 28fps.

Just to add another tool: ftp://78.46.141.148/driver/CPU/amd/enwa.exe

It would be good to know which memory ranges are configured non-cacheable without and with WA enabled. Maybe the config for VGA memory access (either windowed or LFB) isn't changed at all.
Also the primary gain of WA seems to be strongly code dependent.
However there should be some difference in the write performance graph in e.g. speedsys. With WA on write it would be Read Burst, then writes to L1 until the cache line is written back. Without it would be write burst for each. (Both for Cache Miss).

I tried enwa.exe earlier but couldn't get it working for some reason. But I've tried it again and now it's working. I don't know what I was doing before...

But I tested Quake again and it is slower with enwa.exe as well. I also tried k5.sys with the fixed range exemption enabled which prevents WA in 000A_0000h-000F_FFFFh. Still slower than without WA.

And I also tried switching the Verite card for a Riva 128 ZX. WA is still slower. 😀

The speed drop is <1 fps. From ~32.2fps down to ~31.5fps at 320x200. But it isn't margin of error. It is always slower.

How does the slow down compare in software mode vs. 3D accelerated mode?

feipoa wrote:

How does the slow down compare in software mode vs. 3D accelerated mode?

Ok I installed VQuake and did more runs of "timedemo demo1" at 320x200 with the Verite V1000. Write allocate is faster in this case. I see 45.5fps without and 45.9-46.0fps with WA.

I used enwa.exe.

Not exactly awe-inspiring but at least the tides have turned. Do higher resolutions fare any better?

feipoa wrote:

Not exactly awe-inspiring but at least the tides have turned. Do higher resolutions fare any better?

Yeah it looks like write allocate is not much of a benefit.

The V1000 is most likely going to be a bottleneck much above 320x200 so I didn't do any comparisons at higher resolutions.

I did do one run of demo3 at 640x480 and scored 22.3fps. It looks like the K5 PR200 is performing like a Pentium 133 with vQuake. http://gona.mactar.hu/v1000/ (see bottom chart)

Try a more powerful graphics card?