New BBC Master dual processor demo - Twinhead

[SarahWalker]

Just released at Revision 2021, a new BBC Master demo by myself and some of the guys at Desire - Twinhead. This is I believe the first full dual processor demo, requiring a 6502 copro.

Note that this _probably_ requires a real Acorn copro, at least in the current version - it's been tested on both external 3 MHz and 4 MHz Acorn copros, but I've been told it will hang roughly halfway through on PiTubeDirect. I need to buy the latter and figure out what's going on... It's totally untested on anything else, and in particular absolutely requires the 24-byte FIFO to be present, which I gather isn't the case on all modern Tube ULA replacements.

It will run on both B-em and BeebEm, though the latter makes an absolute hash of the music so is not recommended!

Youtube link for those without the hardware : https://youtu.be/AvUCeJ6SEUI

[marcusjambler]

Wow this is great
Confirmed working on a Master with a Kortink Reco6502mini copro

[danielj]

Wow this is great
Confirmed working on a Master with a Kortink Reco6502mini copro

That has a short FIFO IIRC?

[tricky]

Looks and sounds great, nice one Sarah

[hoglet]

Excellent demo Sarah.

I've run tried the following second processors:

- 3MHz Acorn cheese-wedge - works

- PiTubeDirect Co Pro 0 (279MHz) - hangs
- PiTubeDirect Co Pro 1 (3MHz) - hangs
- PiTubeDirect Co Pro 2 (279MHz) - hangs
- PiTubeDirect Co Pro 3 (4MHz) - hangs

- PiTubeDirect Co Pro 16 (69MHz lib6502) - works

- Matchbox Co Pro 0 - 20171007_0719 Firmware (3MHz - Arlet's 65C02 Core) - hangs
- Matchbox Co Pro 1 - 20171007_0719 Firmware (4MHz - Arlet's 65C02 Core) - hangs
- Matchbox Co Pro 2 - 20171007_0719 Firmware (16MHz - Arlet's 65C02 Core) - hangs
- Matchbox Co Pro 3 - 20171007_0719 Firmware (64MHz - Arlet's 65C02 Core) - hangs

- Matchbox Co Pro 0 - 20160228_1555 Firmware (4MHz - AlanD's 65C02 Core) - works
- Matchbox Co Pro 1 - 20160228_1555 Firmware (8MHz - AlanD's 65C02 Core) - works
- Matchbox Co Pro 2 - 20160228_1555 Firmware (16MHz - AlanD's 65C02 Core) - hangs
- Matchbox Co Pro 3 - 20160228_1555 Firmware (32MHz - AlanD's 65C02 Core) - hangs

The PiTubeDirect hangs all happen at 2 min 1 seconds on this screen:

The Matchbox 20171007_0719 hangs all happen at 2 mins 42 seconds on this screen:

The preceeding sequence is corrupt:

The Matchbox 20160228_1555 hang happens much earlier:

I've also tried BeebFPGA and it runs to completion, but sadly there are lots of 6845 issues, which is rather a shame, since I worked hard to get all the BitShifter's demos running.

So it looks like this is a great test for discovering 65C02 emulation issues!

Are you planning to release the source at some point? That would make debugging these emulation issues much easier.

Dave

[hoglet]

Sarah,

I've started looking into the emulation issues and found something a bit unexepected.

This is from running the PiTubeDirect debugger on Co Pro 16 (the one that works).

On the Co Processor, I'm seeing writes to &FEE3 in this section:

Here's the code with the write:

Code: Select all

1a50 4c df 19 jmp 19DF
1a53 a9 00    lda #00
1a55 8d e3 fe sta FEE3
1a58 20 cd 19 jsr 19CD
1a5b 85 7e    sta 7E
1a5d 85 6b    sta 6B
1a5f 20 cd 19 jsr 19CD

It looks like possibly this section of code is incorrectly using the host address for a tube register, rather than the parasite address?

This was one of the sequences that crashes on the Matchbox. The address range &FEE0-&FEE7 on the Matchbox is used for bank switching. So writing there may cause a section of code to vanish.

Dave

[SarahWalker]

Whoops! Pretty certain I started out with the correct host/parasite addresses, but probably started getting them confused at some point; on Acorn hardware they're all partially decoded so it works by accident. Shouldn't be too difficult to fix...

[hoglet]

on Acorn hardware they're all partially decoded so it works by accident.

That's true on the host side, but on the parasite they are fully decoded:

That's from the external 6502 Co Pro schematic; I believe the internal one is the same.

So I'm not sure what's going on here.

Dave

[Arcadian]

So fluid, fast and smooth - what a great achievement having a dedicated demo for the Master Turbo.

Is there a possibility of all collaborators doing a talk at either a Saturday or Thursday ABug event do you think?

[marcusjambler]

That has a short FIFO IIRC?

I think that could be the case because running it side by side with an actual Master turbo shows up some corruption in a few sections on the kortink copro run through.

[Simon]

This is fantastic, super smooth.

[Multiwizard]

Wow very nice Sarah...

[hoglet]

Sarah,

I've been looking into why the 3D Pyramid sequence (after the two-headed dog) hangs in PiTubeDirect on CoPros 0..3.

Debugging is hard-going, because that particular Co Pro doesn't include a debugger. However, I'm about 95% sure that during that sequence, the stack becomes unbalanced and wraps. There is evidence of this if you inspect the stack after a crash:

Code: Select all

0100  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0110  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0120  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0130  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0140  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0150  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0160  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
0170  c2 71 43 1a 64 7d 64 7d 6a 7d 64 7d 6a 7d 64 7d  .qC.d}d}j}d}j}d}
0180  6a 7d 64 7d 6a 7d 6a 7d 7c 8a ba 7e c2 71 43 1a  j}d}j}j}|..~.qC.
0190  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
01a0  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
01b0  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
01c0  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
01d0  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
01e0  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.
01f0  c2 71 43 1a c2 71 43 1a c2 71 43 1a c2 71 43 1a  .qC..qC..qC..qC.

PiTubeDirect CoPro 0..3 definitely mis-handles the case where JSR or RTS causes the stack to wrap. The stack pointer ends up at the right address, but the data is read/written incorrectly.

The lib6502 Co Processor handles stack wrapping correctly, which is why I think it works.

I've also tried the 65816 Co Processor, and this also fails in an identical way. That has a debugger, and the stack is definitely wrapping during that sequence.

Do you deliberately rely on stack wrapping at this point, or is that accidental?

I'm going to work with Dominic and see if we can fix this in PiTubeDirect....

Dave

BTW, if it's any help, now I know what's going on, I can run a real Acorn 6502 Co Processor and capture an instruction trace with a logic analyzer. But you could probably do the same in b-em.

[SarahWalker]

Totally accidental. Looks to be the result of two effects sharing the same code (triangle sorting + plotting) with different ideas of how the stack should be managed. Shouldn't be too difficult to fix, but PiTubeDirect should really handle this!

[dp11]

I think it is more the stack wraps around ie overflows. PiTubeDirect fast core does handle this except in the case where a JSR/RTS is used and the stack pointer is at an odd address.

[hoglet]

I think it is more the stack wraps around ie overflows. PiTubeDirect fast core does handle this except in the case where a JSR/RTS is used and the stack pointer is at an odd address.

It's slightly worse than this: there are three cases where it behaves incorrectly:
- SP=0x100: JSR writes 0xFF and 0x100
- SP=0x1FE: RTS reads 0x1FF and 0x200
- SP=0x1FF: RTS reads 0x200 and 0x201

I've quicky hacked a correct (but slow) implementation and that fixes the issue with this section of the demo.

So at least that confirms there's nothing else going on.

Shouldn't be too difficult to fix, but PiTubeDirect should really handle this!

I tend to agree with this...

I'll work with Dominic to fix this as efficiently as is possible.

Dave

[BeebMaster]

I thought it would be a good test for "PyClock" if it could transfer a 100K+ disc image file from my NFS share into the Native ARM RAM so I could write it to a floppy disc with my DFS writer without anything going on fire, and it worked fine. It looks superb on the Sony Trinitron, much better than the Dell flat-screen, and runs all the way through no problem with real Turbo co-processor on Station 114. It must do a *TV0 somewhere, as I have Station 114 set to TV255, so the top line was cut off.

Also the screen is a quite a bit left-aligned on the Sony TV, but seems just a tiny bit right-aligned on the Dell monitor (with PiRGBtoHDMI).

It also works from a real 5.25" floppy disc, which probably nobody else is using.

And thank you for putting all the files in one directory, it means I could easily transfer it to Econet and run it from there no problem.

On PiTube co-pro 0, it crashes at the point shown in Hoglet's first picture on Station 201 and Station 114 (although not exactly the same frame, and the music continues.)

[BeebMaster]

I'm not sure if it helps to diagnose what's happening but it doesn't seem to play all the way through when run from Econet on 65C102 Turbo co-processor, or PiTube. Sometimes it breaks down during the "Progen" sequence (Hoglet's third picture) but most often when the two dog-heads are on screen, and there is a file access (07 or 08).

[hoglet]

Just a quick note to say this now runs on BeebFPGA.

The issue turned out to be the default VDU FIFO depth in BeebFPGA was only 1-byte.

This has now been fixed.

[ChrisJC]

Highly impressive. I would be interested to know how it actually works.

Chris.

[Dave Footitt]

Nice work Sarah, very impressive

[cmorley]

I went to set up my Master this afternoon to try it wondering what copro I should use... my home built 65C02 one or the Pi (following the diagnostics from Hoglet) and then I realised... I own a Master Turbo!

So one *co.intube and a shift-BREAK later... wow! Great demo. I am really impressed. It definitely makes the Master stand up to be counted against the 16bit machines with their fancy blitters and bitplanes.

[SarahWalker]

Updated the first post with v2, which should now work on PiTubeDirect and Matchbox. Many thanks to hoglet for tracking down what was going wrong!

[danielj]

Having finally managed to run this on proper hardware, it is excellent - very elegantly done!

d.

[Simon]

Just set my master up to test something else and ran this for the first time on hardware, and damn it's even better.

Amazing.