Posted on March 05, 2014 at 23:21
Probably some clues:
- DWT contains a cycle counter on 32 bits (possibility to trigger an event on overflow), this can be used to measure ... well ... cycles very accurately.
- One of the problems is I don't want to include time spent in interrupts on the sleep time, but using WFI, the interrupt starts immediatly. The SEVONPEND feature seems interesting, because it triggers an event (that can wake the core if WFE is used instead of WFI) when an interrupt in pended but not enabled in the NVIC. This would allow software to stop counting sleep cycles, and start counting run cycles before it enables the NVIC and the interrupt triggers. BUT if all interrupts are disabled through, for examplean CPSID instruction, is the mecanism still working?
Then the idle task would be like (pseudo code):
while(true)
{
stop couting run cycles (through DWT CYCCNT)
disable all interrupts (CPSID)
start counting sleep cycles (CYCCNT)
WFE
stop counting sleep cycles
start counting run cycles
enable interrupts (CPSIE)
}
And gather that data (run cycles, sleep cycles) to calculate CPU usage on average say over one second.
I'll do some testing tomorrow, but do someone already tried this?
Thomas
