audio-guide

Low-level Audio Guidelines (Real-Time Software Mixing)

Engine-agnostic architecture for a software audio mixer: a real-time render thread feeds the device, voices are summed into the output buffer, and the game thread talks to the audio thread only through a lock-free queue. The lock-free queue itself is owned by lock-free-guide; preallocated pools by memory-management-guide; struct-of-arrays voice state by data-oriented-design-guide.

Essentials

• Mix in float, ramp every gain - Accumulate voices in deinterleaved float through a gain matrix; never snap a gain or you click, see references/mixing-and-buffers.md

The render path

• Render quantum - Produce fixed-size blocks sized to fit the device buffer; bounds per-block cost, see references/audio-callback-thread.md
• Queue ahead, sleep on the event - Top up ~1.5 quanta, then wait on the device's buffer-low event, see references/audio-callback-thread.md
• Latency vs buffer size - Bigger buffers resist underrun but add latency; tune against the worst real device, see references/audio-callback-thread.md

Mixing and signal

• Gain matrix m[ic][oc] - Volume, pan, and spatialization are all one per-voice input->output matrix, see references/mixing-and-buffers.md
• Fit before output - Scale the summed float mix to range, then convert/interleave to the device's integer format, see references/mixing-and-buffers.md
• Interpolate fractional reads - Linear by default, cubic/Hermite where fidelity is worth it, see references/resampling-and-dsp.md

Voices and control

• Steal the least important - Full pool: cut the lowest-priority/quietest/oldest voice, never refuse silently, see references/voice-management.md
• Handles, not pointers - Hand out slot+generation so stale references to recycled voices are detectable, see references/voice-management.md
• Commands are value-complete - A command must not point at game-thread-mutable state, see references/command-handoff.md