Re: benchmarking m-file replacements

[Rik]

On 04/21/2016 07:25 AM, address@hidden wrote:
> Yes, maybe deprecation is no option at all then. But I think we should
> use the power of Octave, to make the code base as easy as possible, if
> the m-file does not result in a total performance penalty.

As long as there is no performance penalty, I am not opposed.  But
benchmarking can definitely be difficult.  I presented a talk at OctConf
2015 on improving m-file performance which talked briefly about some of the
pitfalls.

>
> tic; for i = 1:5, printf_m("%s", "hi everyone"); end, disp(toc)
> 6.3205e-04

For a small sample size like this the results are often skewed by the first
run through the loop where the printf_m function is parsed and placed in
the symbol table.  To get around this, execute the function printf_m once
before calling the benchmark.  Even so, small sample sizes usually don't
work because you may hit a window when other processes on your
computer--which can steal CPU cycles--are either more or less active. 
Switching to cputime rather than the wall time measured by tic/toc can
help, but still does not completely eliminate the effect of background
processes.

> tic; for i = 1:5, printf("%s", "hi everyone"); end, disp(toc)
> 2.1100e-04
> tic; for i = 1:20000, printf_m("%s", "hi everyone"); end, disp(toc)
> 1.5331
> tic; for i = 1:20000, printf("%s", "hi everyone"); end, disp(toc)
> 1.5113
>  

I benchmarked printf_m using the script below.

-- bmark.m script --
N = 1e4;

more off;

tic;
for i = 1:N
#  printf ("Hello World\n");
  printf_m ("Hello World\n");
endfor
bm = toc;

clc;
bm
-- End script --

The results were:
printf_m: .404
printf: .102
% change: +296%

> function numbytes = printf_m (varargin)
>
>   if (nargin < 1)
>     print_usage ();
>   endif
>
>   numbytes = fprintf (varargin{:});
>
> endfunction
>

printf does not return the number of bytes unless requested.  I think you
may need something like this

nbytes = fprintf (varargin{:});
if (nargout > 0)
  numbytes = nbytes;
endif

I didn't benchmark that, but it is likely to slow it down further.

--Rik