Paula Audio DMA general questions

[remz]

Hi Amiga coders,

Would someone be able to kindly point me to some explanation or documentation on the exact workings of Amiga audio registers and paula state machine? (Aside from the hardware reference manual)?

For example, a few questions I have in mind:

When using DMA audio, normal unattached independent channels:

Exactly when is AUDxLC being read? I seem to understand it is re-read every time the currently playing sound has reached its AUDxLEN.
If so, I must assume one should always turn off DMA for an audio channel prior to resetting the AUDxLC 32-bit pointer, otherwise one risks having a partially set address (since split into two 16-bit writes, perhaps except on 32-bit chip (?), but even then, one cannot write both pointer and length atomically so there is always a risk of race condition here.

However AUDxPER and AUDxVOL can be freely written at all time and Paula will fetch them at every scanline? (or perhaps at every DMA read, thus depending on the current Period rate?)

Is it enough to write to DMACON to provoke the re-reading of the audio pointer? Or must that channel have been set to off beforehand?

I assume that leaving a fast, low period, silent 2-byte looping audio is wasting bus bandwidth, and would be more efficient to turn off the dma channel instead? (Alternatively, setting an extremely high period would make this channel waste less bus bandwidth? Is 65535 the maximum period in fact?)

I also assume that running a looping sample, whatever its length, if the volume is set to zero, is exactly equivalent of looping on a [0,0] empty sample?

These are all that come to mind now but any additional useful guide or information would be greatly appreciated!

What a wonderful machine, the Amiga..

[mc6809e]

The main thing is to understand is the division of responsibility between Agnus and Paula. Paula's logic will tell Agnus that it needs a new sample from the stored waveform or that it has reached the end of a waveform but for the most part it's just outputting whatever two samples are in the data register at a rate determined by the period. It's up to Agnus to fetch samples to fill the data register after the samples have been output. And it's up to Agnus to reset pointers when the end of the waveform been reached.

Agnus does its work near the beginning of a scanline. This limits the rate at which Paula's data registers can be filled with new data. This means that if Paula is outputting samples too quickly Agnus may not be able to keep up and Paula will go right on outputting whatever old data is in the data register. It doesn't care that the data haven't been changed.

If DMA is off, Paula still doesn't care. It can continue outputting the two samples wasting no bandwidth.

[a/b]

Originally Posted by remz

Exactly when is AUDxLC being read?

1. When the audio DMA is activated and the first word has to be fetched.
2. After playing the last byte (length counter has hit 0) and then after the period counter has hit 0 (or: when the audio DMA is used to fetch the next word and you are at the end).
Also, audio hw will check its DMA state only when it has to fetch, so a too quick disable + re-enable would most likely go unnoticed.

Safe way:
- disable audio DMA (makes your writes atomic, among other things)
- write pointer, length, and other stuff as needed
- wait maximum* amount of time so that audio hw detects it's been disabled
- enable audio DMA
* This depends on the highest period value you are using, the period counter must hit 0 before you re-enable DMA so that the hw catches up. One option could be to set the period to 1 right after disabling the audio DMA, to speed it up. Period change is applied (copied to the internal period counter register) immediately IIRC, someone will surely correct me if I'm wrong about this one.

Might be hard to decipher everything: http://amigadev.elowar.com/read/ADCD.../node02DC.html

[Toni Wilen]

Here are some details, hope this helps:

AUDxPT: when Agnus does first DMA, it gets both DMA request and DMA reset bits via DMAL, does DMA from old AUDxPT -> AUDxDAT and loads new AUDxPT (state 1, this word is not used for anything), second fetch fetches from new PT (state 5).

AUDxLEN: Paula loads it to internal register immediately when DMA gets enabled (and when sample repeats)

AUDxVOL/AUDxPER: Paula loads both to internal registers when new 8-bit sample starts playing (=Volume/Period changes do nothing when Paula plays the sample and period is counting down)

Also when audio DMA gets disabled, it only prevents Paula from generating NEW DMA requests. Possible already DMAL queued DMA will still happen in next scan line. Agnus ignores DMACON audio (and floppy) bits.

[remz]

I have a small bonus question: Is there an official mention that forbids writing 8-bit (byte) to a custom register?
Taking for example AUDxVOL, where only the 7 low bits are used, with the other bits being marked as "not used" in the doc, which means must be set to zero.
I wondered how the hardware reacted to a byte write in that scenario? (from testing at least in WinUAE, it appears to work fine.)

[a/b]

Upper byte (bits 8-15) = lower byte (might be different on some 020+, e.g 060, but in this case it doesn't matter), because audio volume only uses bits 0-6, so it works fine. You can also have any value in bit7 (I'm doing that in my 4kb intro PT player, no problems).