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

[saimo]

Problem found. Caused by a lazy cut'n'paste, which included in the initialization code this line...

dc.l $4e7b0006 ;enable copyback (movec.l d0,dttr0/dacr0)

... instead of this one...

dc.l $4e7b0007 ;enable copyback (movec.l d0,dttr1/dacr1)

[Wrangler]

Repeated benchmark results based on the 17:12 version posted today on various Amigas. RB CC CS now works btw

AA3000+, BFG9060 50MHz
RB 37.171
RB CC 36.744
RB CS 31.011
RB CC CS 31.003

AA3000+, BFG9060 100MHz
RB 63.586
RB CC 62.822
RB CS 53.422
RB CC CS 53.411

A4000D, Warp Engine 80MHz
RB 38.426
RB CC 38.363
RB CS 36.861
RB CC CS 36.855

A4000T, CSPPC 60MHz
RB 43.017
RB CC 42.759
RB CS 34.641
RB CC CS 34.632

There's a clear progression on all machines of:
RB marginally faster than RB CC, which in turn are both clearly faster than the CS versions (and those two are almost identical)

It's also clear the WE is out of line with the two other cards, I guess because as I recall the on-card mem is running at 40MHz.

If you ignore the WE results, the RB speeds are close to a linear relation to the CPU clock speed:

FPS ~= 11.23 + 0.524*clock speed

I'll go back into my corner now...

[Karlos]

AA3000+

Say what?!

[pixie]

Originally Posted by Karlos

AA3000+

Say what?!

The holy Amiga 3000. The Amiga 4000 that never was.

AA3000+ is an Amiga 3000 AGA motherboard based on the Commodore AA3000 schematics with some improvements and extra features that were missing from the original AA3000 motherboard, such as the scandoubler / flicker fixer.

The specifications of the AA3000+ motherboard are:
- Amiga 3000 motherboard with AGA chipset
- on-board 68030 PGA CPU and 68881/68882 PLCC FPU running at 25MHz
- DSP processor (AT&T DSP3210)
- CPU slot for A3000/A4000 accelerators
- 16MB Fast RAM using two SIMM modules (no expensive ZIP RAM chips needed)
- 2MB Chip RAM (no expensive ZIP RAM chips needed)
- extended Fast RAM support (up to 64MB?, work in progress)
- extended Kickstart ROM support up to 4MB
- on-board full 24bit AGA scandoubler / flicker fixer
- support for two external video inputs (for example a graphics card and another scandoubler such as Indivision AGA)
- monitor switch between the video inputs, similar to Ratte monitor switch (controllable by software or external switch)

AA3000+ Motherboard

[saimo]

Originally Posted by Wrangler

Repeated benchmark results based on the 17:12 version posted today on various Amigas.

I'm most thankful to you for the patience.

RB CC CS now works btw

Evidently, the race condition I mentioned a few posts earlier and fixed shortly thereafter was the only culprit (so, I had underestimated it).

AA3000+, BFG9060 50MHz
RB 37.171
RB CC 36.744
RB CS 31.011
RB CC CS 31.003

AA3000+, BFG9060 100MHz
RB 63.586
RB CC 62.822
RB CS 53.422
RB CC CS 53.411

A4000D, Warp Engine 80MHz
RB 38.426
RB CC 38.363
RB CS 36.861
RB CC CS 36.855

A4000T, CSPPC 60MHz
RB 43.017
RB CC 42.759
RB CS 34.641
RB CC CS 34.632

There's a clear progression on all machines of:
RB marginally faster than RB CC, which in turn are both clearly faster than the CS versions (and those two are almost identical)

It's also clear the WE is out of line with the two other cards, I guess because as I recall the on-card mem is running at 40MHz.

If you ignore the WE results, the RB speeds are close to a linear relation to the CPU clock speed:

FPS ~= 11.23 + 0.524*clock speed

I'll go back into my corner now...

No need to, that put a big smile on my face!

The main point that your figures stress is that all the data cache fiddling is detrimental: just leaving it on all the time is the best solution...
... other than on TerribleFire cards, where the on/off switching provides a major boost. Copyback, instead, results always in a slight loss.

[saimo]

Updated table, including the results I got privately from klx300r.

------+---------------------+-------+---------+---------+---------+---------+---
AMIGA | EXPANSION BOARD     | CPU   |       D |       C |       S |     C+S | N.
------+---------------------+-------+---------+---------+---------+---------+---
1200  | Blizzard 1230 IV    | 3@50  |  21.137 |       - |  21.237 |       - | 3a
1200  | Blizzard 1260       | 6@50  |         |         |         |         |
1200  | E-Matrix 1230-50    | 3@50  |  20.341 |       - |  20.333 |       - | 3b
1200  | Ex1T                | 2@14  |   6.401 |       - |       - |       - | 2
1200  | M1207               | 2@14  |   6.138 |       - |       - |       - | 2
1200  | PiStorm32-lite      |  -    | 134.171 | 134.242 | 126.232 | 127.680 | R3
1200  | PiStorm32-lite      |  -    | 171.927 | 172.506 | 168.089 | 168.310 | R4
1200  | TerribleFire TF1260 | 6@50  |  23.747 |  23.372 |  27.883 |  27.877 | 6a
1200  | TerribleFire TF1260 | 6@50  |  23.716 |         |  28.002 |         |
1200  | TerribleFire TF1260 | 6@100 |  40.654 |         |  48.256 |         |
3000+ | BFG9060             | 6@50  |  37.171 |  36.744 |  31.011 |  31.003 |
3000+ | BFG9060             | 6@100 |  63.586 |  62.822 |  53.422 |  53.411 |
4000  | A3640               | 4@33  |  14.930 |  14.959 |  17.233 |  17.236 |
4000  | Cyberstorm MK III   | 6@50  |  33.143 |  32.812 |  28.854 |  28.848 | 6b
4000  | Warp Engine         | 6@80  |  38.426 |  38.363 |  36.861 |  36.855 | 6c
4000T | CyberStorm PPC      | 6@60  |  43.017 |  42.759 |  34.641 |  34.632 |
CD³²  | The Beast 030       | 3@70  |  29.240 |       - |  29.260 |       - | 3c

CPU:
 X@F -> 680X0 @ F MHz

DATA CACHE MODE:
 D = Default (always on + writethrough)
 C = Copyback
 S = Switching

NOTES:
 2.  FAST RAM only
 3a. SIMM 60 ns
 3b. SIMM 70 ns
 3c. SRAM
 6a. 68060 rev. 5
 6b. 68060 rev. 1
 6c. SIMM 25 ns (maybe)
 R3. Raspberry Pi 3 A+
 R4. Raspberry Pi 4 B

[hooverphonique]

Originally Posted by saimo

Updated table, including the results I got privately from klx300r.

DATA CACHE MODE:
 D = Default (always on + writethrough)
 C = Copyback
 S = Switching

Is writethrough the default for a 060? On a 040 it defaults to copyback after SetPatch has been run.

[saimo]

@hooverphonique

Originally Posted by hooverphonique

Is writethrough the default for a 060? On a 040 it defaults to copyback after SetPatch has been run.

If the CACHECOPYBACK switch is not specified, PVE sets the data cache to writethrough for both 68040 and 68060 (hance SetPatch doesn't matter).

@all

Updated table above with results from PiStorm32+.

[modrobert]

Don't add this to the table, but tried FS-UAE with JIT enabled.

> pve RUNBENCHMARK
rendered frames number:     256
elapsed seconds:            0.121
frames rendered per second: 2115.702

[saimo]

Originally Posted by modrobert

Don't add this to the table, but tried FS-UAE with JIT enabled.

> pve RUNBENCHMARK
rendered frames number:     256
elapsed seconds:            0.121
frames rendered per second: 2115.702

Yeah, with UAE it's possible to reach ridiculous speeds (even my underpowered PC manages to reach 1000 fps)

[Dunny]

Hopefully this is done right - PiStorm32Lite, Pi4B no overclock.

[pixie]

Originally Posted by saimo

Yeah, with UAE it's possible to reach ridiculous speeds (even my underpowered PC manages to reach 1000 fps)

Even chipset on WinUAE can be way faster then the real thing.

[saimo]

Originally Posted by Dunny

Hopefully this is done right - PiStorm32Lite, Pi4B no overclock.

Definitely right Thanks! Table updated.

By the way, one of the rare cases where copyback gives a little improvement.

[klx300r]

cool seeing all the test results!

would be great if any 040 owners could run the tests too

[hooverphonique]

Originally Posted by saimo

@hooverphonique
If the CACHECOPYBACK switch is not specified, PVE sets the data cache to writethrough for both 68040 and 68060 (hance SetPatch doesn't matter).

Ah,ok.. I thought you meant OS default, when you said default

[hooverphonique]

Originally Posted by klx300r

cool seeing all the test results!

would be great if any 040 owners could run the tests too

Will try to run it on my A4k tomorrow, if I remember.

[saimo]

Originally Posted by hooverphonique

Ah,ok.. I thought you meant OS default, when you said default

No problem, understandable

Will try to run it on my A4k tomorrow, if I remember.

That would be great! A test on 68040 is sorely missing.

[klx300r]

Originally Posted by hooverphonique

Will try to run it on my A4k tomorrow, if I remember.

thanks & don't worry we'll remind you

[hooverphonique]

Please find attached a photo of the results from my A4000 with an A3640 @ 33MHz (i.e. the cpu side of the motherboard also runs at this frequency).
I booted OS 3.1 w/o SS, ran SetPatch, and then PVE - the screen was all black during the benchmark, I hope this is correct/intentional!?

[saimo]

Originally Posted by hooverphonique

Please find attached a photo of the results from my A4000 with an A3640 @ 33MHz (i.e. the cpu side of the motherboard also runs at this frequency).

Thank you!
Table updated.

I booted OS 3.1 w/o SS, ran SetPatch, and then PVE - the screen was all black during the benchmark, I hope this is correct/intentional!?

Yes, it's correct. I have removed the displaying of graphics during the benchmark to avoid the restrictions caused by buffering, which would not allow to measure the actual computational performance.

Did you try it to run it normally? Could you move around without problems?