PED81C - pseudo-native, no C2P chunky screens for AGA

[pixie]

I cannot get it to run in WinUAE, is there any special setting one should be aware?

[saimo]

Originally Posted by paraj

About the same. Stationary: 17.4fps, Moving about: 19.3.

:/

Just to make sure: did the red band look similar to the on in my previous video?

Two more things to try (later):
* No Allocate Mode for disabling the data cache (fetching the map data by 16-byte lines is counterproductive, and this is in fact why on 68030 the cache burst gets disabled);
* squeezing some cycles by means of better pipelining (if possible at all).

[saimo]

Originally Posted by Reynolds

Awesome stuff. looking forward to see where it ends up in several stuffs used.

If it's ever going to be used at all...

Keep up the great work!

Thanks!

[saimo]

Originally Posted by pixie

I cannot get it to run in WinUAE, is there any special setting one should be aware?

Maybe you're simply running the executables I attached earlier? They also need the data files contained in the archive you can download from https://retream.itch.io/ped81c. If you've done it already, what happens, exactly?
Requirements-wise, basically you need 68020+, AGA, PAL SHRES and 4 MB of FAST RAM.

[pixie]

Originally Posted by saimo

Maybe you're simply running the executables I attached earlier? They also need the data files contained in the archive you can download from https://retream.itch.io/ped81c. If you've done it already, what happens, exactly?
Requirements-wise, basically you need 68020+, AGA, PAL SHRES and 4 MB of FAST RAM.

I'm running it from that site, it appears some graphic glitches, it appears I am not using shires from what I read on the demo, but if I go on screenmode prefs the mode is there... My system is running on 030 emulated with 128mb RAM but on top of RTG

[saimo]

Originally Posted by pixie

I'm running it from that site, it appears some graphic glitches, it appears I am not using shires from what I read on the demo, but if I go on screenmode prefs the mode is there... My system is running on 030 emulated with 128mb RAM but on top of RTG

The screenmode prefs are irrelevant, as the program hits the hardware directly. But they could be useful for a test: what happens if you set the Workbench screenmode to PAL SHRES? Maybe your RTG card has trouble dealing with it. Also, it would be great if you could show me a picture.

[pixie]

Originally Posted by saimo

The screenmode prefs are irrelevant, as the program hits the hardware directly. But they could be useful for a test: what happens if you set the Workbench screenmode to PAL SHRES? Maybe your RTG card has trouble dealing with it. Also, it would be great if you could show me a picture.

I had set on WinUAE>display>resolution>Hires
On:
pvm.png
pvm02.png

I then switch to
WinUAE>display>resolution>SuperHires
In:
01 Workbench_004.png
01 Workbench_005.png
01 Workbench_006.png

[saimo]

Originally Posted by pixie

I had set on WinUAE>display>resolution>Hires
On:
pvm.png
pvm02.png

I then switch to
WinUAE>display>resolution>SuperHires
In:
01 Workbench_004.png
01 Workbench_005.png
01 Workbench_006.png

Ah, you meant WinUAE's preferences! Sorry, I thought you referred to the Workbench's
You found the solution: the last two pictures look correct

[saimo]

New test build: on 68040 and 68060, this one disables the data cache altogether when the frame is being rendered; this avoids many useless data fetches as the map data is almost never read by consecutive locations.
I've also checked whether the core loop could be made to exploit the 68060 pipelines better, but it looks like there isn't any room for improvement (I did swap around a couple of instructions, though, but the result should be just the same).
It would be great to get test results from machines mounting those CPUs and - why not - also 68020s.

Note: for easier testing, the attached archive includes also the data files.

[paraj]

New version crashes (just black screen after starting, doesn't allow exit). Don't see any red part of screen with version from post #38.

[saimo]

Originally Posted by paraj

New version crashes (just black screen after starting, doesn't allow exit).

Ouch, I must have messed something up with the caches handling. Could you try this version here, please? I have reworked the ways caches are handled and the code now has different paths according to the CPU (although, apart from caches handling, there are still no 68060-specific optimizations).

Also, to make benchmarking easier, I have added two command line options.

Template: BLINDBENCHMARK=BB/S,FULLBENCHMARK/FB

BLINDBENCHMARK: run benchmark without copying the data to the PED81C raster
FULLBENCHMARK: run benchmark without skipping anything

The functions make PVE render 100 frames, quit to AmigaOS and print the results to the standard output.
During the execution of benchmarks, joystick and keyboard must not be operated to not affect the results.

Finally, I have reworked the controls:
* replaced mouse with keyboard;
* [FIRE] to start;
* [ESCAPE] to quit;
* [F1] to turn the staggered lines on/off;
* [F2] to turn the fps indicator on/off.

Don't see any red part of screen with version from post #38.

I guess I did a stupid thing: I must have built the executable with the test flag off, so the border got blanked, thus hiding the color band. Well, it doesn't matter - it wasn't important at all.

[Lunda]

Very cool!

FULLBENCHMARK result: 30.48 fps on 70MHz 030.

[DanyPPC]

Congratulations Saimo, I have no words.
Excellent voxel engine for the Amiga !!!

[saimo]

Originally Posted by Lunda

Very cool!

FULLBENCHMARK result: 30.48 fps on 70MHz 030.

Cool, thanks!
Could you run also the blind benchmark and give me the result, please? Also, what is your Amiga model, accelerator board and RAM speed? I'd like to add your figures to the table in the manual.

[saimo]

Originally Posted by DanyPPC

Congratulations Saimo, I have no words.
Excellent voxel engine for the Amiga !!!

Thanks! However, I don't know if (and actually I doubt) it is actually "excellent".

[Karlos]

Originally Posted by Lunda

Very cool!

FULLBENCHMARK result: 30.48 fps on 70MHz 030.

70MHz 030?!

[Lunda]

Originally Posted by Karlos

70MHz 030?!

This: https://eab.abime.net/showthread.php?t=115241

[Karlos]

Originally Posted by Lunda

This: https://eab.abime.net/showthread.php?t=115241

Yeah, I probably said the same thing in that thread too, lol

[paraj]

Works again, but speed is unchanged:
bb: 19.15
fb: 17.42

In your blind benchmark, you're not copying to chipmem, is that correctly understood? Seems weird that it's so much slower on (my) 060, and chip copy has so little impact. Are you doing any CacheControl stuff or lots of cache flushes?

[saimo]

Originally Posted by paraj

Works again, but speed is unchanged:
bb: 19.15
fb: 17.42

Thanks once again.

In your blind benchmark, you're not copying to chipmem, is that correctly understood?

Yes.

Seems weird that it's so much slower on (my) 060, and chip copy has so little impact.

Exactly. I'm totally puzzled.

Are you doing any CacheControl stuff or lots of cache flushes?

None at all (except for at startup, when I flush the caches after modifying the code in a few places)! I simply disable the data cache (directly with movec, without the OS) just before rendering because that involves reading the source pixels mostly from non-consecutive locations in memory, so fetching whole lines of data would be a waste. (The previous versions didn't do this. The same happens for the 68040, while for the 68030 I enable the cache burst only when copying the sky background and the rendered frame to CHIP RAM).

Can you try also the attached version, please? It simply ensures that the superscalar dispatch is enabled and has a little dbf -> subq+bpl optimization for the 68060 code. I don't expect much difference, though.