STM32 SAI DMA: copy whole runs of frames, larger halves and queue

[JaxCalder]

Summary

In the STM32 SAI DMA, the writer and the interrupt worked one frame at a time, with atomics on every frame:

• Writer: copies as many frames as fit and moves tx_head once.
• Interrupt: refills a half buffer with one or two memcpy() calls from the queue (two when the queue wraps) and moves tx_tail once. Underruns are counted as before.
• Half buffer: 64 → 128 frames, so at 48 kHz there is an interrupt every 2.7 ms instead of every 1.3 ms. The DMA buffer comes from the small dma heap: 4 KB for 4 TDM slots.
• Queue: 256 → 1024 frames (21 ms, from the ordinary heap). A mixer's 10 ms period (480 frames) now fits, so its writer is woken a few times per period instead of on every half.

Part of bringing dmaudio on the STM32F746G-DISCO from ~45 % to ~10 % of the CPU (see choco-technologies/dmaudio, choco-technologies/dmwm8994, choco-technologies/dmdevfs).

Test plan

• Builds for stm32f7
• STM32F746G-DISCO: MP3 playback through dmaudio → dmwm8994 → dmsai

🤖 Generated with Claude Code