Re: [avr-gcc-list] Trouble with division

[Larry Barello]

Here is an old email with relative times for various math operations in avr-gcc.

----- Original Message ----- 
From: "Ulrich Schweitzer" <address@hidden>
...
Yes you are right, the counter is running at full speed. Unfortunatelly I 
need the resolution of full speed and also cannot afford to lose 
interrupts. If I can't avoid the division (I'm not sure, yet), I could 
try to raise the clock frequency (it's only 1Mhz now) and then use the 
prescaler to keep the counter at the same rate. How many clock cycles 
does the divison of a uint16 through a uint8 take?

----- Original Message ----- 
From: "Dean Ferreyra" <address@hidden>
To: "Uwe Bonnes" <address@hidden>; <address@hidden>
Sent: Wednesday, May 22, 2002 10:22 AM
Subject: RE: [avr-gcc-list] 10 bit division in AVR-GCC


> Uwe,
> 
> About 9 months ago I wrote a small AVR program to time various operations.
> Here are the results I got:
> 
> atmega103 at 4 MHz
> 
>              Hz     Cycles
> 
> Overhead:
> ; (u):       190378    21
> 
> uint8_t:
> x+y (u):     142747     7
> x-y (u):     142762     7
> x<<5 (u):    142762     7
> x*y (o):      46497    65
> x/y (o):      37026    87
> x%y (o):      36686    88
> 
> uint16_t:
> x+y (u):     114218    14
> x-y (u):     114225    14
> x<<9 (u):    124927    11
> x*y (o):      19412   185
> x/y (o):      16662   219
> x%y (o):      16802   217
> div (o):      14436   256
> 
> uint32_t:
> x+y (u):      81598    28
> x-y (u):      81601    28
> x<<17 (o):    25633   135
> x*y (o):       9214   413
> x/y (o):       6105   634
> x%y (o):       6105   634
> ldiv (o):      5228   744
> 
> uint64_t:
> x+y (o):      26481   130
> x-y (o):      26482   130
> x*y (o):       2936  1341
> x<<33 (o):    13510   275
> x/y (o):       1040  3825
> x%y (o):        977  4073
> 
> float:
> fabs (u):    105206    17
> floor (o):     7810   491
> ceil (o):      7918   484
> x+y (o):      27578   124
> x-y (o):      22592   156
> x*y (o):       9387   405
> x/y (o):       7310   526
> inverse (o):   7297   527
> square (o):    9566   397
> sqrt (o):      2366  1670
> cos (o):        801  4973
> sin (o):        716  5566
> tan (o):        699  5701
> pow (o):        379 10533
> exp (o):        920  4327
> log (o):        688  5793
> log10 (o):      633  6298
> acos (o):       702  5677
> asin (o):       719  5542
> 
> The Hz column is how many of the given operations the AVR was able to do in
> one second.  The Cycles column is how many cycles it took to do the
> operation once.  The Cycles column takes overhead into account with the
> overhead calculated in the first test (labeled ";", i.e. an empty C line).
> (For example, the test of "tan" above can be performed 699 times in one
> second, so it takes 4 MHz / 699 = 5722 cycles per operation.  Since the
> overhead is 21 cycles, the result you see is 5701 cycles.)  Each operation
> was run 65536 times or until a timer overflowed, whichever came first.  (The
> "u" or "o" in the parentheses indicate whether or not the timer overflowed.)
> 
> I did this to get in the ballpark and kept this list handy when making
> descisions on how to implement my algorithms.
>