Merge pull request #392 from lovell/hist-find-perf

vips_hist_find: ~30% performance improvement for 1-4 band uchar
This commit is contained in:
John Cupitt 2016-02-28 11:29:14 +00:00
commit e46e2b7b28

View File

@ -238,12 +238,46 @@ vips_hist_find_uchar_scan( VipsStatistic *statistic,
int i, j, z; int i, j, z;
/* Tried swapping these loops, no meaningful speedup. /* The inner loop cannot be auto-vectorized by the compiler.
* Images with 1-4 bands are manually unrolled to improve
* performance for the most common cases by a factor of two.
*/ */
switch (nb) {
for( i = 0, j = 0; j < n; j++ ) case 1:
for( z = 0; z < nb; z++, i++ ) for( i = 0, j = 0; j < n; j++, i++ )
hist->bins[z][p[i]] += 1; hist->bins[0][p[i]] += 1;
break;
case 2:
for( i = 0, j = 0; j < n; j++ ) {
hist->bins[0][p[i]] += 1;
hist->bins[1][p[i + 1]] += 1;
i += 2;
}
break;
case 3:
for( i = 0, j = 0; j < n; j++ ) {
hist->bins[0][p[i]] += 1;
hist->bins[1][p[i + 1]] += 1;
hist->bins[2][p[i + 2]] += 1;
i += 3;
}
break;
case 4:
for( i = 0, j = 0; j < n; j++ ) {
hist->bins[0][p[i]] += 1;
hist->bins[1][p[i + 1]] += 1;
hist->bins[2][p[i + 2]] += 1;
hist->bins[3][p[i + 3]] += 1;
i += 4;
}
break;
default:
/* Loop when >4 bands
*/
for( i = 0, j = 0; j < n; j++ )
for( z = 0; z < nb; z++, i++ )
hist->bins[z][p[i]] += 1;
}
/* Note the maximum. /* Note the maximum.
*/ */