Re: reduction funs optimizations + min/max question

[Jaroslav Hajek]

On Fri, Feb 13, 2009 at 9:33 PM, John W. Eaton <address@hidden> wrote:
> On 13-Feb-2009, Jaroslav Hajek wrote:
>
> | this changeset: http://hg.savannah.gnu.org/hgweb/octave/rev/53b4fdeacc2e
> | reimplements the reduction and cumulative reduction "cores" sum, prod,
> | sumsq, cumsum and cumprod
> | for better performance.
> |
> | A short benchmark (set n to suitable value):
> | n = 5e3;
> | c = {1,single(1),1+1i,single(1+1i)};
> | for i = 1:4
> |   a = c{i}*rand (n);
> |   disp (typeinfo (a));
> |   tic; b = sum (a); toc
> |   tic; b = sum (a, 2); toc
> |   tic; b = prod (a); toc
> |   tic; b = prod (a, 2); toc
> |   tic; b = sumsq (a); toc
> |   tic; b = sumsq (a, 2); toc
> |   tic; b = cumsum (a); toc
> |   tic; b = cumsum (a, 2); toc
> |   tic; b = cumprod (a); toc
> |   tic; b = cumprod (a, 2); toc
> | endfor
> |
> | on Core 2 @ 2.83 GHz, g++ -O3 -funroll-loops -march=native,
> | I get with a recent tip:
> |
> | the relative speed-ups, measured as (old_time / new_time - 1) * 100%
> |
> |   431%
> |  1009%
> | [...]
> |
> | n-d arrays seem to yield even slightly better results.
>
> Thanks for looking at this.  These are impressive speedups.
>
> | I would like to also address min & max. I'm not, however, entirely
> | certain about the expected behaviour of min/max
> | w.r.t NaN and NA. Currently, the following holds:
> | max(NA, NaN) = max (NaN, NA) = NA
> | max([NA, NaN]) = NaN
> | max([NaN, NA]) = NA
> |
> | so, it seems that the behaviour is pretty much arbitrary, i.e. NA is
> | not treated as "weaker" or "stringer" kind of NaN.
> | Matlab has no NA, so that's no source of authority here. So, is there
> | anything else to follow?
>
> Since NaN and NA are both NaNs, does it matter?
>
> What does sort do?
>
> Octave's NA was patterned after NA in R, so what does R do?
>

OK, I've just checked with R 2.4.0 and it seems that it does not
distinguish NA and NaN in min/max, so I'll follow that.

-- 
RNDr. Jaroslav Hajek
computing expert
Aeronautical Research and Test Institute (VZLU)
Prague, Czech Republic
url: www.highegg.matfyz.cz