Bugfix: micros() returns smaller number when timer wraps

[Kees-van-der-Oord]

See issue 49

1. Fix for anomaly in micros() when the counter wrapped just before it was read (same solution as in R3 code)
2. Specify timer configuration as constants to make micros() considerably faster

[CLAassistant]

All committers have signed the CLA.

[aentinger]

This fixes #49.

[aentinger]

Hi @Kees-van-der-Oord ☕ 👋

Thank you for your PR. I'm inclined to merge it but there are a couple of suggestions I'd like you to tackle first.

[facchinm]

This might not work on all FSP supported boards (like Portenta C33). It's very important to fix this before merging since otherwise the number of ifdefs in the core would soon become unbearable

[aentinger]

Yes, the defines have to go @Kees-van-der-Oord , as already mentioned here.

I don't see how there can be any runtime speed-up by having them in the code, as its simple string search and replace done by the preprocessor.

[Kees-van-der-Oord]

Please see the comment here. I will try to find the most efficient implementation that also works for the Portenta C33.

[aentinger]

Yes, the defines have to go @Kees-van-der-Oord , as already mentioned here.

I don't see how there can be any runtime speed-up by having them in the code, as its simple string search and replace done by the preprocessor.

[Kees-van-der-Oord]

I just pushed a new version that is compatible with both the Uno R4 and the Portenta C33 (and any FSP chip with a reasonable AGT clock frequency). Optimizing the code a little more, I was able to almost match the execution time of micros() on the R4 to that on the R3:

               average     max
Uno R3:        3.46 us     24 us
Uno R4:        3.61 us      7 us
Portenta C33:  1.26 us      3 us

the code to test the micros() function:

void test_micros3() {
  Serial.println("testing micros3 - send input to stop");
  Serial.flush();
  uint32_t count = 0;
  uint32_t total = 0;
  uint32_t max = 0;
  uint32_t min = 0xFFFFFFFF;
  while(!Serial.available() && (count < 1000000)) {
    uint32_t t1 = micros();
    uint32_t t2 = micros();
    uint32_t period = t2 - t1;
    total += period;
    if(period > max) max = period;
    if(period < min) min = period;
    ++count;
  }
  double ave = -1;
  if(count) {
    ave = double(total) / count;
  }
  Serial.print("testing micros3 average: ");Serial.print(ave);Serial.print(" min: ");Serial.print(min);p(" max: ");p(max);Serial.println(" us");
  while(Serial.available()) Serial.read();
}

[Kees-van-der-Oord]

I found one of reasons why the performance is so slow: the function FspTimer::get_counter(). I just pushed a version in whch the call to this function is replaced with a direct counter registry read,
The performance is now much better:

               average     max
Uno R3:        3.46 us     24 us
Uno R4:        2.12 us      5 us
Portenta C33:  0.60 us      2 us

[aentinger]

Thank you for your work @Kees-van-der-Oord.

We are getting there, but there has been a misunderstanding between you and me. I did not ask you to replace macros with variables and inline functions, I asked you to get rid of the macros alltogether and just inline those statements directly into the code.

[aentinger]

Successfully tested for

• Portenta C33 ✔️
• Uno R4 WiFi ✔️
• Uno R4 Minima ✔️