I've seen this improve the std. deviation often by a factor of 2
and occasionally up to a factor of 10. It is sometimes not much
better, but never seems to be appreciably worse compared to using
gettimeofday.
Thanks to David A. Schleef <ds@schleef.org> and his liboil for
the implementation.
- add a yield () function that's called before every test. It reduced the std
dev slightly for me
- fix double comparisons to not just compare the integer part
Previously we were setting the target tolerance based on the surface
type. But that doesn't work as multiple backends will provide a surface
of type meta. So instead we put the tolerance as a value in the
cairo_test_target data structure.
With this change, some single-pixel errors of 1 in the SVG backend
should now be ignored.
This patch generates Type 1 fonts for the PS/PDF backends when TrueType
or Type 1 subsetting fails. This has the advantage over the current
Type 3 fallback of reduced font size and better quality rendering
in some PDF viewers. xpdf shows a large improvement in text display
quality with this patch.
PDF Files with Type 1 fonts fail to open in any version of
ghostscript prior to 8.54. The problem is the hex encoding of the
encrypted portion of the font. The PDF reference says this should
only be in binary.
1. Remove all the alarm/signal code, which just isn't doing what we want for some reason.
Instead, for now we'll simply run for a fixed number of iterations, (perhaps we
can tune that per test later).
2. Before computing mean and stdandard deviation of runs, sort them all and discard the
top and bottom 20% of the values.
Now the standard deviation for the paint test is generally 2% or less.
The interface of the various buffer/image_diff functions is improved to
provide the maximum pixel difference in addition to the number of pixels
that differ. This value can then be used to compare against a per-backend
tolerance.
Currently I've set the SVG backend's tolerance to 1 to handle some issues
we're currently seeing of single-bit differences on different systems, (but
we're not exactly sure why yet).
Also I improved the image_diff routines to properly report a status value
on failure rather than the bogus value of -1 for pixels_changed.
