Mercurial > octave
changeset 21627:225d88a97104
hist: reduce memory usage by avoiding conversion to double (bug #47707)
* hist.m: do not convert whole input array to double. Instead, convert only
the pieces that are needed (max and min values), and floor the cutoff point
so they are placed in the right place when sorted (for N > 30 only). There
was no issue with overflow as the comments suggested (at least not anymore).
Further reduce memory by avoiding intermediate variables. Add more logic so
that NBINS or specified X can be of class single. Add tests for all this.
| author | Carnë Draug <carandraug@octave.org> |
|---|---|
| date | Sat, 16 Apr 2016 19:15:37 +0100 |
| parents | 1dda942a2514 |
| children | 37a44d8af461 |
| files | scripts/plot/draw/hist.m |
| diffstat | 1 files changed, 76 insertions(+), 8 deletions(-) [+] |
line wrap: on
line diff
--- a/scripts/plot/draw/hist.m Thu Apr 14 12:51:43 2016 -0700 +++ b/scripts/plot/draw/hist.m Sat Apr 16 19:15:37 2016 +0100 @@ -96,6 +96,11 @@ max_val = max (y(:)); min_val = min (y(:)); + ## Do not convert if input is of class single (or if already is double). + if (! isfloat (y)) + max_val = double (max_val); + min_val = double (min_val); + endif iarg = 1; if (nargin == 1 || ischar (varargin{iarg})) @@ -105,6 +110,11 @@ else ## nargin is either 2 or 3 x = varargin{iarg++}; + ## Do not convert if input is of class single (or if already is double). + if (! isfloat (x)) + x = double (x); + endif + if (isscalar (x)) n = x; if (n <= 0) @@ -126,11 +136,11 @@ endif endif - ## Avoid issues with integer types for x and y - x = double (x); - y = double (y); + cutoff = (x(1:end-1,:) + x(2:end,:)) / 2; + if (isinteger (y)) + cutoff = floor (cutoff); + endif - cutoff = (x(1:end-1,:) + x(2:end,:)) / 2; n = rows (x); y_nc = columns (y); if (n < 30 && columns (x) == 1) @@ -144,11 +154,10 @@ ## The following algorithm works fastest for n greater than about 30. ## Put cutoff elements between boundaries, integrate over all ## elements, keep totals at boundaries. - [s, idx] = sort ([y; repmat(cutoff, 1, y_nc)]); - len = rows (y); - chist = cumsum (idx <= len); + m = (nthargout (2, @sort, [y; repmat(cutoff, 1, y_nc)]) <= rows (y)); + chist = cumsum (m); chist = [(zeros (1, y_nc)); - (reshape (chist(idx > len), rows (cutoff), y_nc)); + (reshape (chist(! m), rows (cutoff), y_nc)); (chist(end,:) - sum (isnan (y)))]; endif @@ -212,3 +221,62 @@ ## Test bug #42394 %!assert (isempty (hist (rand (10,2), 0:5, 1)), false) %!assert (isempty (hist (rand (10,2), 0:5, [1 1])), false) + +## Test bug #47707 +%!test +%! y = [1 9 2 2 9 3 8 9 1 7 1 1 3 2 4 4 8 2 1 9 4 1 ... +%! 2 3 1 1 6 5 5 3 9 9 1 1 8 7 7 2 4 1]; +%! [n, x] = hist (y, 10); +%! [nn, xx] = hist (uint8 (y), 10); +%! assert (nn, n) +%! assert (xx, x) +%! +%! ## test again with N > 30 because there's a special case for it +%! [n, x] = hist (y, 35); +%! [nn, xx] = hist (uint8 (y), 35); +%! assert (nn, n) +%! assert (xx, x) + +## Test logical input +%!test +%! y = [0 1 0 0 1 0 1 1 0 1 0 0 0 0 0 0 1 0]; +%! [n, x] = hist (y, 10); +%! [nn, xx] = hist (logical (y), 10); +%! assert (nn, n) +%! assert (xx, x) +%! +%! ## test again with N > 30 because there's a special case for it +%! [n, x] = hist (y, 35); +%! [nn, xx] = hist (logical (y), 35); +%! assert (nn, n) +%! assert (xx, x) + +## Second output argument must be of class single if data is class single. +%!assert (class (nthargout (2, @hist, rand (10, 1, "single"))), "single") + +## Handle second argument correctly, even when it's class integer +%!test +%! y = [2.4 2.5 2.6 5.4 5.5 5.6] +%! n = [2 3 1]; +%! x = [1 4 7]; +%! [nn, xx] = hist (y, uint8 ([1 4 7])); +%! assert (nn, n) +%! assert (xx, x) + +%!test +%! y = [2.4 2.5 2.6 5.4 5.5 5.6]; +%! s = (5.6 - 2.4) / 6; +%! x = [2.4+s 4.0 5.6-s]; +%! n = [3 0 3]; +%! +%! [nn, xx] = hist (y, 3); +%! assert (nn, n) +%! assert (xx, x) +%! +%! [nn, xx] = hist (y, uint8 (3)); +%! assert (nn, n) +%! assert (xx, x) +%! +%! [nn, xx] = hist (y, single (3)); +%! assert (nn, n) +%! assert (xx, single (x))