Queues: queues with lengths as powers of 2

Trying out a silly thought re: our queues' max length.

At present, the FIFO is internally just a 1-d memory that I treat as a circular buffer. Two registers, read and write, mark which cell of the memory should next be read from or written to. The registers are incremented as needed, and, when they equal the maximum length of the queue, they are zeroed out to simulate a wraparound.

This PR requires the queue's maximum length to be a power of 2. We are now able to simply keep incrementing these registers forever, trusting actual overflow to do the zeroing-out for us. We save a couple of equality-checking cells and the wiring needed to run those checks and potentially zero the registers.

The PR also goes ahead and pipes this change to PIFOs, and therefore to PIFO trees. The user used to set a constant indicating a queue's maximum length. They now set a "queue-len-factor", and the actual maximum length is determined as $2^{\text{queue-len-factor}}$.

Please ignore the massive diffs in the .data and .expect files. The max queue length was originally 10, and I needed to re-run the oracles with the new queue length of 16 so that my change would be implementable.

Having set the length to 16, I compared both versions:

The old version, with the actual logic to zero out registers in case they hit 16, takes 820002 cycles.
The new version takes 780002 cycles, a 4% decrease for very little work indeed.

Sadly I didn't note register counts, or numbers for the other queues that build atop of FIFOs. One imagines magnifying gains.

The catch, of course, is that the max queue length is determined in a slightly roundabout way... someone who wants a max queue length of 64 can get exactly that, but someone who needs a max queue length of 70 will be forced to go get 128.

calyxir / calyx

Queues: queues with lengths as powers of 2 #2028