Circular DMA bursts

It’s a bit odd. The NDTR register remembers the value it had when you start DMA. It’s not available elsewhere. I dont believe there is a way to start midway through a buffer.

You could change your buffer to always start at the beginning, though it may have performance limitations. But you have to start/stop dma anyway to “reverse�? direction so maybe this is not so hard to do as well.

Thanks! Indeed, the aim is to avoid rotating the buffers because of computation time. Another approach might be to chain 'normal' mode DMA to the end of the buffer, a number of cycles in circular mode and (if needed) another normal DMA from the start to the final index, using interrupts. This will result in a bit of timing jitter at the start and finish. But an (undocumented?) access to the NDTR preload would result in much cleaner code without jitter =)

Are you talking about timer-triggered memory-to-GPIO transfers?

There's no way to start a circular DMA with an offset. When you enable the DMA channel, the exposed CNDTR is copied to a hidden one, from where it is "refreshed" after each round.

At the cost of memory, you can set up two copies of the required pattern after each other, and then start circular DMA with original size in CNDTR but the memory pointer register pointing with the required offset into that dual-buffer.

JW

Hi JW,

Yes, this is about high speed mem to periph DMA on a GPIO port. Doubling the pattern is a good idea, thanks!

If I go for the circular DMA only approach there's a requirement to stop circular DMA after an exact number of timer events. What is the best way to set that up? The number of DMA transfers can be large, a 16 bit counter is not enough.

Maybe you can chain timers, using the DMA-triggering one as master, and a slave to it in external-clock mode as event counter, to determine the end of transfers. Even 16-bit timers can count up to 32 bits thanks to their prescaler...

JW