Ok this is time to share a nice routine for playing 8bit PCM samples on the PSG without the ugly spikes that are generated by the previous conventional techniques that play sample by sample each PCM sample using a 3 channels table.

Go to http://map.tni.nl/articles/psg_sample.php for details on the “standard” solution.

In a word, the standard solution associates each 8bit sample to a triplet of channel levels suitably chosen in order to minimize the error of the PSG overall output level with respect to the given sample. As the sum of the 3 channels leads to 608 disjoint values, you would expect a the output quality of a 9bit quantizer, but, unfortunately, the resulting levels are oddly spaced, so the theorical quality corresponds in dB to about 48dB of SNR, which is turn is equivalent to a 8bit uniform quantizer.

Problem:

The “standard” solution has a very annoying problem: the z80 cannot change the 3 PSG channels at the same time, this implies that, during the transition between two successive samples, the PSG passes trough undefined states where some channels are from the previous triplet and some channels are from the triplet corresponding to the next sample.

The result is that, at each transition between two successive samples, the PSG can produce 1 or 2 spikes of short duration. This effect, repeated at each sample transition, becomes a very annoying wideband noise that sensibly reduce the overall quality of the output, far below the optimal 48dB of SNR.

Solution:

The sole solution is to control the PSG trough every transition.

As the z80 can change a single channel at time and each channel has 16 values, the PSG can be modelled at any time as a state machine with 256 states, 16 inputs, and 608 outputs.

The 256 states are the 16*16 values of the old channels (those that don’t change at the current time), the 16 inputs are the allowed values of the current channel, the 608 outputs are the sum of the volume values from the “state” and of the volume value from the current input.

Each sample transition corresponds to three steps of the state machine.

Assume ABC the current values playing and A’B’C’ the new values.

The transitions are:

ABC -> ABC’

ABC’-> AB’C’

AB’C’-> A’B’C’

So, between the level ABC corresponding to the previous sample and the level A’B’C’ corresponding to the next sample, we get :

ABC

ABC’

AB’C’

A’B’C’

Assume x be the sequence of input samples, we want to minimize the mean square error between the x and the output sequence of levels in any single transition, this implies we need to develop a search algorithm that explores all the sequences of state transitions of the PSG corresponding to the allowed outputs, and finds the best sequence of channel levels that minimize the mean square error.

The algorithm is already existing, and is the classical Viterbi algorithm, I do not run into details here, but in case I will go back on this classical topic.

The output of such an algorithm is the sequence of levels for each channel that encodes the input sequence x.

Now the new problem: each triplet is 4*3= 12 bit : a huge waste of space!

Solution run length encoding.

The best solution should be a RLE based on nibbles, as some "detail channels" can have very long bursts of non equal symbols. Due to lack of time and of energy I did as you can see in the following.

I know that the implementation could be improved, but I cannot fit any a better solution in 447 t cycles (8KHz of PCM player).

Enjoy