Thursday 5 October 2017

Sounds Like a Plan

Sounds like a pain?


I was nervous about the sound effects, having found it in the past to be hard work to get good results. The Dragon and CoCo machines have a 6 bit DAC that is capable of creating some fantastic sounds, but in a rather cruel twist, things are not easy if you want sound at the same time as animation.

The problem is that sound generation requires a fairly regular rate of writes to the DAC, whereas updating graphics often involves a variety of routines of different lengths. The two things don't mesh together very well, so what usually happens is that the sound ends up being produced in short bursts in between screen updates, creating a distinctive thin and choppy kind of sound that can be heard in many Dragon and CoCo games.

Trip down memory lane


In my first game, Balldozer, there wasn't much work to do to keep the graphics updated: Just a ball, a bat and a power-up, along with the occasional brick that needed erasing. That allowed me to have relatively long bursts of sound in between the updates. Low frequencies with decaying volume were generated when the ball bounced off the bat or the sides of the screen, giving a 'boing' sound, and higher frequencies were generated on hitting the bricks, giving a 'ching' sound. This worked out pretty well:





For my second game, ROTABB, I didn't put any sound in until the last minute and it really shows. Because the sound in Balldozer worked out OK, it didn't occur to me that I would have a problem. What I discovered too late was there were no spare cycles left over, and the best sound I could achieve during normal play were some feeble explosions for destroyed enemies. The only way I could find to generate a reasonable sound effect was by having longer bursts of sound and letting the frame rate drop. So to avoid messing up the gameplay, I only did this for destroying the level boss and for the player death, where the lower frame rate gives a nice slow-motion effect. The lesson here is to allow for sound early on in the design:



For the abandoned third game, I had started experimenting with producing sound and graphics at the same time i.e. embedding the DAC update code within the graphics update code. This game spent a lot of time drawing background tiles, so I inserted a DAC write for every second tile drawn. The values to write to the DAC were read from tables, with the volume of the sound being set by a MUL instruction. This gave a much more full and satisfying sound:



Back to the Future


For the new game I would like the sounds to have something of an arcade feel. That generally means lots of noise going on, with a variety of sounds that are rich in low to mid-range frequencies. I felt that the sound in the third game was a step in the right direction, so I decided to build on those ideas:

  • Embed more DAC updates in the game code
  • Make the DAC updates more efficient by not using MUL for volume
  • Instead of a set of fixed wave tables, have a sound buffer that is updated by the sound engine at the start of each game loop
  • Sound engine that generates sounds from parameters
  • Choose a sound effect for playback by pointing to a parameter entry in a table

This list is saying I need a sound engine that is much more sophisticated than anything I've done before. Fortunately the task can be broken down into more manageable pieces. First of all I need a small piece of code that takes one sample from the buffer, writes it to the DAC and updates the buffer pointer using a minimum of registers so that it can be embedded in existing code without causing too much upheaval:

    lda [sndbufptr] ; Get next sample
    sta $ff20       ; Write it to the DAC
    inc sndbufptr+1 ; Update buffer pointer


That takes 20 cycles. It can be reduced slightly by using self-modifying code, but that's only worth doing in a loop with a fair number of iterations because of the overhead of copying the buffer pointer at the start and end. Note that I'm only incrementing the bottom byte of the buffer pointer. This is fine providing the buffer is completely contained within one 256 byte page.

So how many DAC updates do we need to embed in the game loop? Say we wanted an update rate of 1KHz. Assuming normal CPU speed and a frame rate of 25fps, that means an update approximately every 895 cycles, or 40 updates per game loop.

Those updates need to be sprinkled as evenly as possible throughout the code. This is not an easy thing to do, but it doesn't need to be perfect and it might be easier to change the update rate to suit the code. For example I've put a DAC update inside the loop that copies 128 byte blocks of background to the screen. That performs 21 updates spaced at 523 cycle intervals, forcing me to aim for a higher update rate of 1.7KHz

In the tile drawing routines, the loops are slightly too short, so the DAC update is preceded by a flag toggle to perform the update every second loop.

If a loop is really short, then it might be necessary to break the loop up into two or more chunks by having an inner loop doing the work and an outer loop containing the DAC updates. It's a similar story for the loop that waits for the video frame sync. I have an inner loop checking for the sync flag in the PIA, and the outer loop performs the DAC updates at the target rate.

Where's that noise coming from?


Next we need to generate some data to put into the sound buffer. To keep things simple I decided to have a single sound channel. This means only one sound effect plays at a time, and a priority system is required to prevent a less important sound from interrupting a more important sound.

There are two types of sound commonly generated by sound chips, namely square wave and noise, so we may as well start with those. The following code writes a square wave* to the buffer with adjustable frequency and volume:

snd_square
      ldx #soundbuf
      ldb snd_phase
      lda snd_vol
ssqlp sta ,x+
      addb snd_freq
      bcc ssq1
      eora snd_vol
ssq1  cmpx #soundbuf_end
      blo ssqlp
      bra donesound


The frequency and volume are values in the range 0-255. The higher they are, the higher the frequency and volume respectively. snd_phase stores the value of the B register between successive calls to the routine, so that the square wave continues where it left off on the previous game loop. (Though the placement of DAC updates would have to be very accurate for this to make much difference, and if I did get that accurate, I would have to save the volume of the square wave between calls as well)

The noise routine is a little more complex. It reads values from a table of random numbers, sets the volume by ANDing with a mask and then outputs the values to the buffer. The values are repeated in a similar way to the square waves to give some control over the frequency content of the sound. I've used self-modifying code so that the routine is not too slow compared to the square wave routine:

snd_noise
      ldx #soundbuf
      ldu #rndtable
      ldd snd_vol   ; snd_vol & snd_freq
      sta sns2+1    ; volume
      stb sns3+1    ; frequency
      ldb snd_phase
      anda ,u+      ; first sample
snslp sta ,x+
sns3  addb #0       ; frequency
      bcc sns1
      lda ,u+       ; get next random value
sns2  anda #0       ; set volume
sns1  cmpx #soundbuf_end
      blo snslp
      bra donesound


The random number table is a 64 byte table of pseudo-random values, refreshed at a rate of one new byte per game loop such that the table contents are continuously changing.

Now we need a table to define the different sound effects. Each table entry defines the type of sound, how long it lasts and how it changes over time. That suggests these parameters:

  • Sound duration
  • Start volume
  • Change in volume per game loop
  • Start frequency
  • Change in frequency per game loop
  • Address of sound generator (square or noise)

The good news is those parameters fit into seven bytes, meaning we can pack a lot of different sound effects into a small space.

When a sound effect starts, the volume and frequency are copied to variables for the sound generator to use. Then for each game loop, the sound generator is called and then the volume and frequency have the 'change' parameters added to them. The duration is used to initialise a counter that is decremented once per game loop to set the length of the sound.

The sound priority is defined very simply by the address of the parameters. Sounds nearer to the start of the table have lower priorities than sounds nearer to the end of the table. When a sound effect request is made, the address of the new sound is compared to the address of the current sound. If the new address is lower, then the sound is not started, otherwise the new sound will replace the existing sound.

If the current sound has address zero, then it means no sound is playing, which happens when the sound duration counter has expired, for example. Having the lowest possible address also means the silent state has the lowest priority, ensuring the next sound request will be granted.

Put it all together and what do you get?


I've implemented the ideas discussed but haven't to date put a lot of effort into spreading out the DAC updates evenly. There are about 44 DAC writes per loop, some bunched up, some spread out and a big gap when the sprites are being drawn. Even so, the result is surprisingly good: (Please excuse the jumpy video. The game runs much more smoothly than this)




Back to the Future II


At some point I would like to add more parameters to make the sound effects more complex, such as periodic frequency variation (vibrato) and variable duty cycle for the square waves. I also need to do a better job of spreading out the DAC updates plus there is the issue of supporting both PAL and NTSC machines, as the different frame rates mean different sizes of sound buffer. But on the whole I am extremely pleased with the new sound engine. I just wish I had thought of it back in the day :)

As a final thought it occurred to me that the same sound engine could be used to drive a variety of different sound hardware (such as sound carts) just by replacing the lowest levels of the driver. Definitely worth considering for a future 'deluxe' version.

I'd better get on and write some more code before I run out of things to write about...


*Caution: Under no circumstances attempt to generate square waves in real life. The infinite accelerations and energy densities required would rapidly bring on the end of the universe and possibly void the warranty on your speakers.