FIFO driven PIO vs bitstream, what's best for accurate transmission of fixed and high rate protocols?

[kevinjwalters]

The default timings for the neopixel library in MicroPython on the rp2 port are very close to the boundaries for practical timings for some RGB LED pixels. The discussion in #18295 raised the choice of two implementations:

1. MicroPython's bitstream implementation &
2. An RP2030/RP2350 PIO based implementation (using FIFO)

for implementing the WS2812B wire protocol, for latter Get Started with MicroPython on Raspberry Pi Pico's PIO_WS2812.py is one example. Simplistically the PIO option gives perfect timing at up to CPU clock frequency resolution but I think this is more complex as there's no DMA here and the majority of uses would require data being fed into the PIO code via the FIFO. As the data rate increases this becomes increasingly likely to be interrupted with the stall likely to cause problems for many protocols, the WS2812B is a prime example.

For MicroPython I haven't yet dived into what mp_hal_quiet_timing_enter() does and what mp_event_handle_nowait() can do on the rp2 port which appear in the implementations. Does anyone have experience of this to be able to pinpoint the risks for the two approaches from interrupts and background tasks, e.g. packets arriving over WiFi, data arriving on USB serial, etc? For (GRB) WS2812B a 4 word FIFO would have 4 * 24 = 96 bits of data in it equating to 120us at 800 kilobaud protocol meaning anything that runs must complete comfortably within that small time window. For this unidirectional protocol the FIFOs could be "joined" to give 8 words.

MicroPython's bitstream implementation

The bitstream implementation

• reads data from memory in a loop and output giving a close approximation of desired timing
• has some minor additional delays on final bit of each byte causing jitter
• wrapped with mp_hal_quiet_timing_enter() and mp_hal_quiet_timing_exit() to perhaps tame interrupts in some way.

RP2030/RP2350 PIO implementation (using FIFO)

The MicroPython StateMachine implementation

• this reads data from a MicroPython buffer
• filling the FIFO with 8-32 bits using pio_sm_put() til it's full
• calls mp_event_handle_nowait() if FIFO is full

[GitHubsSilverBullet]

Simplistically the PIO option gives perfect timing at up to CPU clock frequency resolution but I think this is more complex as there's no DMA here

What makes you think that there's no DMA available for the PIOs?

[kevinjwalters]

Forgetfulness or ignorance, perhaps aided by some of the MicroPython example code I've been looking at!

In the context of MicroPython is it easy to use? Does MicroPython have a way to hint to (a future) GC system that a piece of memory must not be moved?

I'll have a read through Eric Styger's RP2040 with PIO and DMA to address WS2812B LEDs article later, this uses C++.

[GitHubsSilverBullet]

In the context of MicroPython is it easy to use?

Pretty easy to use, actually.

Does MicroPython have a way to hint to (a future) GC system that a piece of memory must not be moved?

A bytearray (or array.array) once allocated will stay in place until deleted.