New Z3660 card for A3000/4000

[cloverskull]

Curious as I haven't used a Falcon - can someone share some links/videos to show what the DSP was capable of? Maybe games/demos/etc?

[abu_the_monkey]

Originally Posted by cloverskull

Curious as I haven't used a Falcon - can someone share some links/videos to show what the DSP was capable of? Maybe games/demos/etc?

[ Show youtube player ]

IdTech #2 (Quake 2 engine) redesigned for a 16MHz Atari Falcon from 1992

obviously the falcon is from 1992. NOT the engine being demonstrated in the video.

[VladR]

Originally Posted by Cyprian

which Falcon, Falcon 030, Microbox 030 or Falcon 040 Painter with Jaguar chipsets (DSP/GPU)

Touche If we're talking about prototype devices, we need to [as much as it pains me to admit] extend the same courtesy to Atari too
Also, I should have thought of the Painter myself, given that I paid an egregious amount for a Falcon on eBay few yrs ago

[cloverskull]

Originally Posted by abu_the_monkey

[ Show youtube player ]

IdTech #2 (Quake 2 engine) redesigned for a 16MHz Atari Falcon from 1992

obviously the falcon is from 1992. NOT the engine being demonstrated in the video.

Wow, impressive. Thanks.

[abu_the_monkey]

Originally Posted by cloverskull

Wow, impressive. Thanks.

this is rendering the levels and not running the whole game code, but still, a 16mhz 030 + DSP, not too shabby.

[Cyprian]

Originally Posted by Cyprian

which Falcon, Falcon 030, Microbox 030 or Falcon 040 Painter with Jaguar chipsets (DSP/GPU)









well, DSP3210 has a nice feature - direct access to the fast ram with some constrains - only longwords can be fetch from/to fastram (no byte/word access). The weak point is the size of it's RAM 2x 1K x 32bit.


Anyway it looks cool and I would love to see more 3000+ / DSP videos, tests.

interesting note for DSP3210 on the page 3:
"16.6 milion instruction per second (66MHz clock)"

https://dsp-book.narod.ru/c_dsp/c_dsp3.pdf


DSP56001 has the same figure at 32MHz clock.

[VladR]

Originally Posted by Cyprian

interesting note for DSP3210 on the page 3:
"16.6 milion instruction per second (66MHz clock)"

https://dsp-book.narod.ru/c_dsp/c_dsp3.pdf


DSP56001 has the same figure at 32MHz clock.

Not sure if we can really compare two different RISC chips like that given the HW pipelining differences.

In theory, Jaguar's GPU is 26.6 MIPS, but the numerous HW pipeline stalls bring it down to about half.

Some stages of my 3D flatshader went up to almost 20 MIPS, but on average I was getting around ~13-15 MIPS / rendered frame.

That's when I spent 3x as much time rearranging code. Without that effort, 8-9 MIPS were the norm...

A really nice side benefit of RISC pipelining is that I was able to bring down the Division cost to just 1 cycle (from 9, IIRC) by rearranging code so while division is being computed, chip isn't stalled.

Wonder how that compares to Falcon's DSP and A3000+ DSP...

[eXeler0]

Originally Posted by VladR

Not sure if we can really compare two different RISC chips like that given the HW pipelining differences.

In theory, Jaguar's GPU is 26.6 MIPS, but the numerous HW pipeline stalls bring it down to about half.

Some stages of my 3D flatshader went up to almost 20 MIPS, but on average I was getting around ~13-15 MIPS / rendered frame.

That's when I spent 3x as much time rearranging code. Without that effort, 8-9 MIPS were the norm...

A really nice side benefit of RISC pipelining is that I was able to bring down the Division cost to just 1 cycle (from 9, IIRC) by rearranging code so while division is being computed, chip isn't stalled.

Wonder how that compares to Falcon's DSP and A3000+ DSP...

In one of his videos, Wrangler compares a mandelbrot program which has exes for both 060 and the 3210 DSP and they were pretty equal if I remember it right. Having that power in parallell to the CPU certainly would have been nice on a 030 system.
(And one can speculate that if the A3000+ would have been produced, maaaybe the A1200 would also have had that DSP to increase the user base and give the A1200 more compute power.. and then of course also the CD32 would have been a more interesting machine.. (But ok, thats a lot of ifs and buts, lets leave it there)

[Wrangler]

If you want to know what the DSP on the A3000+/AA3000 is capable of, see my What Might Have Been demo video. Full 3D rendering in 256 colours with Z-buffer on an '030+DSP with no cheating. The thing that limits performance there is c2p on an '030

[abu_the_monkey]

@Wrangler

I think I saw that you did some work on winuae to implement the DSP stuff.
If so, how usable is it under emulation?

[Wrangler]

I'll be brief as we're off topic with this. Basically, the answer is that there is a fully functional DSP emulator in the beta for WinUAE but it isn't playing nicely with the rest of WinUAE, I think because it's running in a separate thread which creates synchronisation challenges that don't exist in real hardware (eg the "CPU" thread running instructions in zero time as far as the "DSP" thread is concerned).

[Cyprian]

Originally Posted by Wrangler

If you want to know what the DSP on the A3000+/AA3000 is capable of, see my What Might Have Been demo video. Full 3D rendering in 256 colours with Z-buffer on an '030+DSP with no cheating.

that one [ Show youtube player ] ? Nice video.

Flat shaded 3D an example on DSP56001:
[ Show youtube player ]
[ Show youtube player ]

Originally Posted by Wrangler

The thing that limits performance there is c2p on an '030

what about doing C2P by the DSP?

[grelbfarlk]

I submitted a DMCA takedown request based on the following offenses:
1) No DSP emulation benchmarks on this primitive accelerator
2) No hardware DSP benchmarks for sound decoding from A3000 SCSI
3) No DSP assisted decoding while this card is accessing the PPC on the Mediator, or Prometheus or whatever
4) Somebody was supposed to be rebuilding AHI with DSP, or ARM, or WarpOS support

[Wrangler]

Originally Posted by Cyprian

Flat shaded 3D an example on DSP56001:
[ Show youtube player ]
[ Show youtube player ]

Very impressive! Flat shading wouldn't be difficult to add to my engine if you pick the colour palette carefully.

Originally Posted by Cyprian

what about doing C2P by the DSP?

I did consider it but a few things stopped me doing that. First, the DSP didn't seem particularly suited to the C2P calcs - its strengths are in floating point not manipulating integer bitmaps. Also the 030 was idle for a chunk of time while the DSP was rendering so it seemed more efficient. The killer reason, though, was that I'm lazy and I could just bolt on Kalms' C2P routine with no effort!

[Cyprian]

Originally Posted by Wrangler

Very impressive! Flat shading wouldn't be difficult to add to my engine if you pick the colour palette carefully.

actually flat shaded vectors are cool.
I wonder if DSP3210 could drive the blitter for line drawing and filling vectors. It could be interesting to see how many vectors it can show in one frame.

Originally Posted by Wrangler

I did consider it but a few things stopped me doing that. First, the DSP didn't seem particularly suited to the C2P calcs - its strengths are in floating point not manipulating integer bitmaps. Also the 030 was idle for a chunk of time while the DSP was rendering so it seemed more efficient. The killer reason, though, was that I'm lazy and I could just bolt on Kalms' C2P routine with no effort!

that's a good idea to run both processors in parallel, and Kalms' C2P rox

[Wrangler]

Originally Posted by Cyprian

actually flat shaded vectors are cool.
I wonder if DSP3210 could drive the blitter for line drawing and filling vectors.

I've not tested this but in principle it should be possible because the DSP can read/write to the entire 32 bit mem space

[Wrangler]

To get back on topic, though, part 2 of my review of the Z3660 is out:

[ Show youtube player ]

[fgh]

What an awesome all in one card! Where do I sign up for one?

[emiespo]

Originally Posted by alexh

It isn't. It has a Xilinx Zynq with Dual-core ARM Cortex-A9 @ 866MHz whereas RPi4 has Quad-core A72 @ 1.8GHz. The difference will be night & day.

Pistorm32 + Emu68 ~2000 MIPS, Z3660 120 MIPS

Isn't this huge gap mainly because of Musashi vs Emu68? I presume if Emu68 was made compatible with this board, the difference would be much less.

I also wonder if a similar approach to the ZZ9000 could be taken, to have emulation run on one ARM core and leave the other one free for parallel tasks - like RTG, audio, datatypes, etc... and even better if this actually worked

[qz3fwd]

My Z3660 is on the way. Scheduled delivery in 2 days (when I will be at work).
Exciting times. Mouser has not shipped the MYiR board yet.