improve fifo performance?

[myklemykle]

Hi,

I'm profiling the main loop in a multicore pico program that sends a single 32-bit value from core0 to core1 on every loop.

The entire loop takes about 300 usec to finish, on average. My profiling suggests that a single call to rp2040.fifo.push_nb(), to send 32 bits from core0 to core1, is taking about 30 usec to complete, on a chip that's running at 133mhz. That comes to about 4000 clock cycles to complete the non-blocking function. That's a tenth of the loop, for one line of code. Is that the expected performance of this fifo implementation? Anything I can to to make it faster?

Thanks in advance,
-mykle-

p.s. In general I am really loving the arduino-pico environment! Works great, with great documentation.

[earlephilhower]

The FIFO now is built for correctness, not for speed, sorry. Because we need to use the HW FIFO to cause an IRQ on the other core for freeze-CPU operations (i.e. while doing flash erase/write) it doesn't use the HW FIFO for rp2040.push()/etc.

It uses the Pico-SDK's multi-core, multi-thread safe queues:

arduino-pico/cores/rp2040/RP2040Support.h

Lines 70 to 84 in f33dfd2

	bool push_nb(uint32_t val) {
	// Push to the other FIFO
	return queue_try_add(&_queue[get_core_num() ^ 1], &val);
	}
	uint32_t pop() {
	uint32_t ret;
	while (!pop_nb(&ret)) { /* noop */ }
	return ret;
	}
	bool pop_nb(uint32_t *val) {
	// Pop from my own FIFO
	return queue_try_remove(&_queue[get_core_num()], val);
	}

If you have a simple 1-producer/1-consumer setup it should be possible for you to implement a simple, lockless FIFO (as done in the SerialUART code) in your own code, and not use the built-in heavy duty queueing functions.

[myklemykle]

Fair enough. That's probably the right approach for this situation -- one producer, one consumer.