Re: For loop benchmarks between Octave versions

[Daniel J Sebald]

On 06/24/2014 02:47 AM, Daniel J Sebald wrote:

> In theory, we could do something similar, i.e., just write our own
> little utility program in C that implements the tests done in Octave
> code. Since we are always using a C compiler, we could make the building
> of this utility part of the Octave build process. For example, say the C
> program (call it Cperf.c for now) generates as an output a string of
> characters that lists the test and the CPU consumption:
>
> systemcommline> Cperf
> doubleForLoop/1: 181.8
> doubleForLoop/a=1: 294.7
> doubleForLoop/a=b: 275.6
> doubleForLoop/a=a+b: 335.8
> doubleForLoop/a=sin(b*i): 383.4
> vectorOfSquares: 677.7
> mandel: 224.9
> fib: 687.6
> parse_int: 2392.1
> quicksort: 1244.9
> pi_sum: 18236.1
> rand_mat_stat: 572.8
> rand_mat_mul: 515.9
> printfd: 1233.3
>
> Then, the Octave test script that implement perf.m could call the system
> function Cperf to get the values and do the relative-to-C comparisons.
>
> That seems like a pretty good comparison to me, and the benefit is we
> know that Cperf is built with pretty much the same environment as Matlab
> and is always present.

Actually, it might make better sense to structure Cperf.c so that it 
accepts inputs:

  Cperf func arg1 arg2 ...

and then returns the test result.  That would simplify the parsing of 
the results to a simple string scan, make the program flexible for the 
number of times through a loop and ensure that the number of times 
through a loop matches whatever Octave is using.  For example, here 
might be a string of script commands to access the C performance:

  result(end+1) = timeit ('vectorOfSquares', @vectorOfSquares, 100, 1e3);
  [stat, out] = system("Cperf vectorOfSquares 100 1000");
  cresult(end+1) = str2double(out);
  relative_perf(end+1) = result(end)/cresult(end);

Dan