Writing a “bare metal” operating system for Raspberry Pi 4 (Part 9)

< Go back to part8-breakout-ble

Playing sound from the audio jack

One thing our game is missing is the excitement of sound! Some beeps and squeaks would be a wonderful addition to make the gameplay more compelling. Let’s work to do just that!

I wrote this code as I referenced Peter Lemon’s work, for which I am very grateful. It did need some significant modification to work on the Raspberry Pi 4 hardware.

Design goal

Perhaps most importantly, we must be able to play sounds in the background. If our audio playback ties up the CPU, then gameplay will stop whilst the sound is playing. I think any player would be immediately put off by the rude intrusion into their adventure!

One solution for this is to implement multi-tasking, thereby making use of the four CPU cores (so far we’ve only used one). This is no small feat, and a big commitment for a few beeps and squeaks.

Fortunately, the Raspberry Pi 4’s hardware allows us to avoid this pain (for now), using something called DMA. This allows specific hardware subsystems to access main system memory completely independently of the CPU.

Notes to self

Here are a few things I learned on this journey, which helped me along significantly:

And always, always have a browser tab open on the BCM2711 ARM Peripherals document. It’s a very handy reference for register addresses and bitmaps etc.

Audio sample format

audio.bin is the audio file we’ll be playing. Technically speaking, it’s 8-bit, unsigned PCM data at 44.1 KHz. This is an unusual format in this modern day and age, but it’s easily converted using a tool like ffmpeg.

To convert from our .bin file to a .wav file that any laptop can play natively, do this:

ffmpeg -f u8 -ar 44.1k -ac 2 -i audio.bin audio.wav

To convert back to our binary format, do this:

ffmpeg -i audio.wav -f u8 -ar 44.1k -ac 2 audio.bin

This should help you try the code with your own audio samples! This one is a short excerpt of me singing (argh!), with my wife playing woodwind. As an aside, I highly recommend checking out the original track.

Testing playback using the CPU and PWM module

I knew DMA transfers might be a tricky beast, so I began by just proving I could play audio to the jack output of the Raspberry Pi 4.

Looking at main(), you’ll see that we first call audio_init(). This function ensures that PWM1 is correctly mapped. PWM is a technique used to control analogue devices using digital signals. Whilst digital signals are either on (1 - full power) or off (0 - no power), analogue signals may be an infinite number of values between 1 and 0. PWM fakes an analogue signal by applying power in quick pulses/bursts of regulated voltage. The resultant average voltage will end up looking roughly like an analogue signal, despite not being one. Clever, eh?

These pulses/bursts do need to be highly accurate for this trick to work, and so we need a reliable clock source. Just like your kitchen clock ticks every second, so the oscillator on the Raspberry Pi 4 has a regular ‘tick’ - in this case, 54,000,000 times per second (54 MHz)! Our audio sample is at 44.1 KHz though, so we need to ‘slow it down’. We do this by first stopping the clock, then setting a clock divisor, setting the PWM range, and enabling the clock again. In my code, I use 2 as the clock divisor (so we’re down to 27 MHz) and set the range to 612 (0x264). Essentially, this means that our PWM module will move to a new sample every 27,000,000/612 times per second - roughly equivalent to 44.1 KHz, which just happens to be the sample rate of the included audio sample audio.bin (I’ve included audio.wav so you can listen normally on your laptop too!).

We then enable the PWM module, telling it to wait for sample data on its FIFO input, and we’re good to go. Until we start filling the buffer, no audio will play.

Hopefully you’ll notice that we set both channel 0 and channel 1 up similarly. This is because we’re working with a stereo sample.

Starting the playback

In playaudio_cpu() we use the CPU to drive our sample data into the FIFO buffer. The sample data is built into the kernel and referenced exactly as we did with the Bluetooth firmware file back in part7-bluetooth (we still haven’t done SD card file access, sorry!).

The code is fairly self-documenting. We essentially check the FIFO buffer isn’t full before we send the left channel byte and the right channel byte (stereo, remember!). If we see errors, we clear them as we go.

And that’s it… The PWM will pick up the digital data in the buffer and send it, PWM-style (faking an analogue signal), at the right speed (thanks to our clock divisor/PWM range mastery), to the audio jack.

Doing it with DMA

In playaudio_dma() our first challenge is that DMA transfers and FIFO registers are 4 bytes wide, so our single byte samples need some zero-padding. Because data is an unsigned char * and safe is (technically) an unsigned int *, we can do this copy with a simple for loop. You’ll notice that safe is a memory location in RAM which we know to be beyond our program code (SAFE_ADDRESS is defined in io.h).

We then set up the DMA Control Block. This is an in-memory structure that will tell the DMA engine everything it needs to know to perform the transfer. This is very well-documented already, but worth pointing out is DMA_DEST_DREQ and DMA_PERMAP_1. These settings ensure that we use our previously-set hardware clock to ‘pace’ the transfer. If we didn’t do this, the DMA engine would just get it done as fast as it could (and our audio wouldn’t sound great!). SRC_INC simply tells the DMA engine to increment the source address throughout the transfer. We’ll be keeping the destination address constant though, since we want it to always point to the PWM module’s FIFO input. Note also the use of PWM_LEGACY_BASE rather than PWM_BASE to address this peripheral memory. This is another quirk of the Raspberry Pi 4 hardware!

Note finally how we set .nextconbk to 0x00. This tells the DMA engine that there is no more work to do after this job is complete. If we wanted to loop the audio sample infinitely (nobody needs that much of me singing!), we could simply set this to address the same control block structure again.

As we enable the DMA engine, playback begins. Notably, however, we’re returned to main() immediately and the CPU can get on with other things, thereby meeting our design goal.

Conclusion

We’re now ready to integrate sound into our Breakout game. But rather than do that, let’s go down a rabbit hole and see if we can’t get CPU playback running on a separate core! I said it was hard, but I love a challenge.

Go to part10-multicore >