delays in spin1

evanh

Damn! Spin on the Prop1 is slow executing. I just made a tight loop that was meant to count off microseconds but it failed dismally ... and it took me some time to believe it was due to processing speed.

pub  waitus( us ) | ticks, interval
    interval := div33( clkfreq, 1_000_000 )    'calc 1 us
    ticks := cnt    'start timing
    repeat us
        ticks += interval
        waitcnt( ticks )

On an 80 MHz sysclock, I had to change the interval calculation to 20 microsecond to allow enough time for the repeat loop to cycle. My lack of experience on the Prop1 showing up again.

Anyone have a solution for finer granularity than 20 us?

evanh

I suppose the answer is use the old waitcnt( cnt + x)

ersmith

@evanh said:

Anyone have a solution for finer granularity than 20 us?

Use flexspin . Or waitcnt, as you suggested later.

Wuerfel_21

Just calculate the amount of cycles ahead of time and then wait for that. At 80 MHz it takes a bit less than a minute to overflow the counter, so that should be fine.

Unrelatedly, if you want to get performance from P1 Spin, you kindof have to think inside its box. Your code there is a good example: you add a value to ticks and then use that value on the next line. Spin prefers when you write like waitcnt(ticks+=interval)

Also, try flexspin's bytecode compiler. It will automagically save cycles and bytes everywhere.

evanh

Funnily, I'm using PropIDE on Linux because PropTool can't find a comport in Wine. I note PropIDE defaults to using bstc for compiling. It seems to have its own download facility ...

evanh

Coooool! Dug up proploader and changed to flexspin for compiling ... the div33() calculation time has now dropped from 4300 ticks odd to 816, so a x5 speed up ... and same for the loop time, needs about 4 us now.

I'm happier back using Kate text editor now anyway. PropIDE was annoying.

evanh

Basically identical timing to BST Compiler when using --interp=rom in Flexspin. I'm getting 3552 ticks for the div33() in both cases. Faster than I had thought for bstc.

pub  div33( dividend, divisor ) : quotient    'round to nearest

    quotient := (dividend + (divisor>>1)) / divisor

Wuerfel_21

@evanh said:
Basically identical timing to BST Compiler when using --interp=rom in Flexspin. I'm getting 3552 ticks for the div33() in both cases. Faster than I had thought for bstc.

For reasons I forgot (and I don't think are relevant anymore), --interp==rom doesn't perform any IR optimizations by default. Enable -O1 manually. The gains over BSTC range from none to quite a bit. Note that BSTC is already faster than the Proptool compiler.

evanh

Okay, yep, -O1 helped. Now 3408 ticks for the div33().