for loops in c++

[wysota]

well I'm nut sure whether or not this is representative; I ran each test a couple of times and each time the result was more or less the same (times in mSec btw).

The fact that you can reproduce it doesn't mean it's representative.

Can't I conclude then, that for these particular pieces of code, the second one is noticably faster than the first one, though they both yield the same result?

Sure, it's faster. Noticably faster - I'm not sure. Biased to prove the point - definitely so

It is true that post-increment is in most cases slower than pre-increment, but the for() loop has nothing to do with it, so let's not exagerrate it - we mostly iterate fors using integers.

how do I measure the process time btw?

Using clock(), times() or "time" shell command. Or a profiler.

[stinos]

The fact that you can reproduce it doesn't mean it's representative.

mmm, good point

Biased to prove the point - definitely so

yeah I know, I'm a bit too keen on details sometimes..

Using clock(), times() or "time" shell command. Or a profiler.

I'm afraid I don't understand this: what's the difference between using QueryPerformanceCounter() or rdtsc() and using eg clock()? Each returns a timer value, possibly from different timers, but how else then measurig a start and stop time and subtracting the two can one mesure execution time? Doesn't a profiler work like that?
Suppose I use time() on the shell, wouldn't time at process stop - time at process start be equal to the two time values I measure inside the test() function + time needed for program init etc?

@TS: sorry for the off-topic discussion, but on the other hand, we'll probably learn from it, seems to me wysota is rather experienced!

[wysota]

Each returns a timer value, possibly from different timers, but how else then measurig a start and stop time and subtracting the two can one mesure execution time?

clock(), time() and "time" ask the system (scheduler) to provide information how much cpu time was spent on the process, not how many miliseconds have elapsed between starting and finishing the execution.

Doesn't a profiler work like that?

Depends what it does. The number of cpu instructions executed is independent of time. If a profiler only measures world time elapsed, then it's biased as well.

Suppose I use time() on the shell, wouldn't time at process stop - time at process start be equal to the two time values I measure inside the test() function + time needed for program init etc?

No, because I can stop the process (CTRL+Z) and restart it two days later ("fg" or "bg"). time will give me a proper result (like 20ms), substracting end and start times will give me 2 days and 20ms. The process can be preemptied practically at any time to yield the CPU to some other process which makes results of such primitive calculations incorrect. The difference between those two measurements is "how much time it took to execute the code" and "how much time it took to execute the code in particular conditions" where "particular conditions" are impossible to recreate, so the measurement is not representative.

[stinos]

clock(), time() and "time" ask the system (scheduler) to provide information how much cpu time was spent on the process, not how many miliseconds have elapsed between starting and finishing the execution.

thanks for clearing that out! I really didn't know that.

[stinos]

The process can be preemptied practically at any time to yield the CPU to some other process which makes results of such primitive calculations incorrect. The difference between those two measurements is "how much time it took to execute the code" and "how much time it took to execute the code in particular conditions" where "particular conditions" are impossible to recreate, so the measurement is not representative.

I've been thinking about this some more: does this mean that using clock() would, for a program like the above, always return the same results, no matter what the system load is?
There's no sleep or waiting conditions, only one thread, so the code executed should always follow the exact same path, so using clock() in the places where I used QueryPerformanceCounter() should always return the exact same value, or is there a pitfall of some kind?

[wysota]

I've been thinking about this some more: does this mean that using clock() would, for a program like the above, always return the same results, no matter what the system load is?

It is approximate so the exact values might vary a bit, but in general - yes, at least I think so.