Quake 2 on a 'stock' Atari Falcon

[trixster]

I bet this is old news and has been covered on here before, but how the hell is this possible?

http://www.atari-forum.com/viewtopic.php?t=26775

10fps-ish on a stock 14mhz 030. Ok, the falcon has a smart DSP and a chunky mode, but that demo puts the fps I get from my 80mhz 060 somewhat in the shade.

Just amazing programming?

[Thorham]

Sick

Play list of sick videos: [ Show youtube player ]

[Mrs Beanbag]

nice, so they started converting it from floating point to integer, i don't know why it was ever in floating point to begin with to be honest.

floating point is overrated imho.

[eXeler0]

Originally Posted by trixster

I bet this is old news and has been covered on here before, but how the hell is this possible?

http://www.atari-forum.com/viewtopic.php?t=26775

10fps-ish on a stock 14mhz 030. Ok, the falcon has a smart DSP and a chunky mode, but that demo puts the fps I get from my 80mhz 060 somewhat in the shade.

Just amazing programming?

Ye, seems they are making the most out of the DSP to do 3d calcs.
Reminds me of a Dave Hanyie speech a while ago. As most of us already know, he wanted to put a DSP (Motorola 56001) into the A3000+ primarily for sound, but if I recall correctly he suggested that it was integrated in such a way that it could have been used to do 3D calculations.
So, it seems this theory was proven right, eventually...

[Sandro]

Originally Posted by trixster

I bet this is old news and has been covered on here before, but how the hell is this possible?

http://www.atari-forum.com/viewtopic.php?t=26775

10fps-ish on a stock 14mhz 030. Ok, the falcon has a smart DSP and a chunky mode, but that demo puts the fps I get from my 80mhz 060 somewhat in the shade.

Just amazing programming?

I didn't see any proof that quake 2 could run at 10 fps in a 030, in fact that's impossible, quake uses fpu and is very intensive and I don't think the DSP or the fpu 68882 which is very slow can do those math calculations in a fast way
maybe they are running the game on 2x1 graphics or 2x2

[alexh]

It's not a Quake2 source port, it is Doug's own engine which can run Q2 maps, textures etc.

Try the last alpha on Your stock falcon 030@16MHz 14mb

https://dl.dropboxusercontent.com/u/...estpublic2.zip

Or the Hatari emulator (mercurial build)

[kovacm]

Originally Posted by Mrs Beanbag

nice, so they started converting it from floating point to integer, i don't know why it was ever in floating point to begin with to be honest.

floating point is overrated imho.

maybe because Intel Pentium can execute FP and integer instructions in parallel?
Similar, Doug engine use parallel processing of 68030 and DSP in Falcon.

[Locutus]

Doug's stuff on Atari is amazing, check out his Doom 'port' Badmood, which is more a custom engine that can run Doom .WAD's.

These aint your typical amiga ports :-)

[gimbal]

Originally Posted by Thorham

Sick

Play list of sick videos: [ Show youtube player ]

I am inclined to call this guy a programming Einstein. Sheesh!

Originally Posted by Locutus

Doug's stuff on Atari is amazing, check out his Doom 'port' Badmood, which is more a custom engine that can run Doom .WAD's.

These aint your typical amiga ports :-)

Playlist!

[ Show youtube player ]

[eXeler0]

I think we should donate an A1200 to this guy ;-)

[Amiga1992]

Originally Posted by Sandro

quake uses fpu and is very intensive and I don't think the DSP or the fpu 68882 which is very slow can do those math calculations in a fast way

You didn't read it seems. They are converting the floating point math to integer.

[matthey]

Originally Posted by Akira

You didn't read it seems. They are converting the floating point math to integer.

The response of kovacm to Mrs Beanbag's quandary (not phrased as a question) makes sense to me. Understanding would be better if the EAB forum did not discard the previous quote.

Originally Posted by Mrs Beanbag

nice, so they started converting it from floating point to integer, i don't know why it was ever in floating point to begin with to be honest.

Originally Posted by kovacm

maybe because Intel Pentium can execute FP and integer instructions in parallel?
Similar, Doug engine use parallel processing of 68030 and DSP in Falcon.

Q2 was written for processors like the Pentium which have a fast FPU which can operate in parallel. FP is also easier to use than fixed point integers. These are big reasons why FP was used for Q2.

Even the early 6888x does a lot of FP processing in parallel to the CPU (integer unit) but the interface, instruction pipelining and caching are not efficient with an external FPU. Most early DSPs did not support FP but were good at fixed point integer math (often better than the CPU integer unit). DSPs are notoriously difficult to program efficiently though. Had the Amiga or Atari with DSP beat the PC, I expect a fast FPU and later SIMD (also not easy to program) would have replaced most DSP processing.

[kovacm]

Originally Posted by matthey

Most early DSPs did not support FP but were good at fixed point integer math (often better than the CPU integer unit). DSPs are notoriously difficult to program efficiently though. Had the Amiga or Atari with DSP beat the PC, I expect a fast FPU and later SIMD (also not easy to program) would have replaced most DSP processing.

yes, Motorola 56001 in Falcon (at 32MHz) is declared as 16 MIPS.
Beside that MC DSP has SIMD characteristics and they are self sufficient (btw today many Falcon coders works in game industry as optimisation specialist for PlayStation, Xbox...) Falcon implementation of MC DSP allow only ~800KB/s data transfer "to" and "from" DSP (or up to 3MB/s if you skip confirmations...).

[SKOLMAN_MWS]

Originally Posted by kovacm

yes, Motorola 56001 in Falcon (at 32MHz) is declared as 16 MIPS.

68040 30MHz 21,68 MIPS 7,23 MFLOPS link

DSP in Falcon in reality faster than 68040.

[beezle]

Floating point is also a lot more precise than interger, something that at times is important, especially when it comes to 3d.

[kovacm]

Originally Posted by SKOLMAN_MWS

68040 30MHz 21,68 MIPS 7,23 MFLOPS link

DSP in Falcon in reality faster than 68040.

It is really complicated; I am not sure what instructions Motorola take into account when declare DSP as 16MIPS (e.g. multiplication are really fast but bit-wise operation are really slow on DSP). You also need to remember that DSP has only 96KB of memory (!) and access to main memory is only 8bit wide so you need very carefully how to "craft" 3D engine!

There are really lot of details at atari-forum; there are diagrams with chunk of codes and how they are managed between CPU and DSP...

MC 68040 (and espec. 060) are quite different than 68030 founded in Falcon. In many cases, using solo 040 would perform faster than 030+DSP. This engine is specially crafted for stock Falcon (beside DSP, using MC68030 256bytes cache, MMU...) and would not benefit from e.g. 040 or 060 CPU as it could.

Originally Posted by beezle

Floating point is also a lot more precise than interger, something that at times is important, especially when it comes to 3d.

Ofcourse. DSP in Falcon have 24bit words and as you saw, Doug manage to implement 3D engine that will work well with given precision.

[beezle]

Sure, it's very cool stuff regardless of it only using intergers.

My comment was more a follow up to some people responding to Mrs. Beanbags suggestion that fp can be overrated.

[AnimaInCorpore]

Originally Posted by kovacm

(e.g. multiplication are really fast but bit-wise operation are really slow on DSP).

I guess you're referring mainly to shifting instructions. Shifting can be done by multiplying with constant values so it isn't not really slow. However, these parts don't look quite elegant within the sources.

[gimbal]

Originally Posted by beezle

Floating point is also a lot more precise than interger, something that at times is important, especially when it comes to 3d.

True enough, that makes it all the more impressive that the graphics of this effort are so stable if you look past the expected low resolution scale blurring.

[kovacm]

^

I remember when I first time saw work-in-progress with flat shaded polygons:

[ Show youtube player ]

it was fastest flat polygons engine on Atari so far and I wonder (I believe as many) if it would be sensible to add any kind of textures on it preserving reasonable frame rate?

eventually Doug implement more than just textures; there are lightmaps, transparency... pure joy