For my little retro-sounding software audio system, the sound chip I'm looking for inspiration is the Texas Instruments SN76489. It, its variants and various clones were used in a lot of hardware from the eighties, like the Sega Master System and the Game Gear, the ColecoVision, the BBC Micro and the IBM PCJr/Tandy 1000. I'm not trying to directly emulate the SN76489 as closely as possible, but it will help to start by looking at how this chip works so I can make something that sound appropriately "retro"-ish.

The SN76489 features three mono square wave tone generators and one white noise/periodic noise generator with 4-bit (16) volume settings. I'm picking it as a starting point because it's one of the simpler sound chips in terms of operation yet can still make some nice tunes. Plus I grew up playing adventure games on a Tandy 1000 HX so I've got fond memories of those triple square wave tunes.

Useful SN76489 links:

• Wikipedia page for Texas Instruments SN76489 (links to chip datasheet in footnotes)
  
• SMS Power page on the SN76489 with info for Sega emulator developers
  
• Sierra AGI specification for the sound system, including IBM PCjr
  
  The links go into the details, but I'll sum up the basics:
  
  The chip has eight internal setting registers which are set with byte sized commands to the input pins. The setting registers are:
  
  
  • 3 x 10-bit frequency settings for the square wave channels
    
  • 3-bit control mode for the noise channel ("periodic" or "white" noise with one of four speed settings)
    
  • 4 x 4-bit attenuation (volume) registers for the four channels
    
    
    The chip gets a clock signal which it then turns into a internal clock. Different varieties of the chip allow different input clock frequencies - some allow up to 500kHz and divide by 2 for the internal clock, but the ones used in the consoles and computers appear to be the type that allow up to 4MHz and divide by 16. Either way ends up with an internal clock of up to 250kHz. But "up to" is the critical point - both the Sega and the Sierra links above say the standard operation was to tie the clock to the NTSC TV colour subcarrier frequency of 3.5795MHz, which gives an internal clock of 223.7kHz. The clock frequency directly affects the output frequencies so it will make a difference.
    
    Each square wave tone channel has a 10-bit counter and an output bit. They work by resetting their counter to the value in their 10-bit frequency setting register until they reach zero, where they flip their output bit and reset their counter again.
    
    The white noise generator uses a 15 or 16-bit linear feedback shift register, where the output is the bit that is shifted out the end and the input bit at the other end is determined by XORing two bits together (usually on the shift register). These are described in detail in the SMS Power page above if you want the details. It seems this is the part that varies the most between different chip designs, which might be important for anyone wants a characteristic noise generator for a specific system.
    
    Each of these four channels outputs 0 or 1, which is set at their volume attenuation levels and combined in a mixer to get the analog output.
    
    There are a lot of further subtleties that are better explained in the linked pages (especially the SMS Power one for Sega emulation), but the basics are fairly straightfoward: counters and shift registers.
    
    Thoughts:
    
    
    • Something I'll have to design for is the difference in clock speeds between actual hardware and software (semi-)emulation. My software is (probably) going to output audio at 40-44kHz, rather than the internal clock speed of 220-250kHz.
      
    • I'm going to have to think of a good way to get info from my program to the sound generation code. PortAudio's documentation says to avoid using mutexes in the callback function. I haven't done much tinkering with threads but I think as long as I keep the data transfer small and simple it will be fine, as the worst that could happen if the callback happens mid-data set is a brief sound spike.
      
    • Going to have to think of a way of encoding this type of audio in a more human understandable way than raw register settings, but that can wait until after I get the basics done.
      
    • I think the SN76489 was first used in 1979 or 1980 but it was used all throughout the 80s. I have to pick a year from somewhere in the decade. IBM PCJr and the Tandy 1000 came out in 1984, along with the garish 16 colour wonders of EGA - somewhere else I might be going.