GitLab

This avoids both function call overhead and extra checks for values that
are outside MININT/MAXINT range, for the uses inside GeglBuffer we only
care about coordinates that are representable as integers - thus these
inline-able shortcut versions are good enough for us.

The speedup seems to consistently make all samplers about 10% faster
according to the tests in perf/

All resampling for GeglBuffers used to happen in RaGaBaA float, now
grayscale formats end up using YaA float instead, which will improve
performance of scaling/rotating in GIMP as well as other tools using
GeglBuffers resampling like the sculpt tool when in grayscale mode.

buffer: add members for interpolation bpp/components

buffer: make linear sampler component generic

buffer: generalize cubic sampler

buffer: generalize lohalo innerloop

For now, assuming that input is still a premultiplied format and treat last component
differently, should sniff whether the interpolation format has alpha and only do
different treatment in that case.

buffer: generalize nohalo for components

buffer: nohalo slight simplification

foo

This reverts earlier indentation breaking changes when GeglRectangle was
renamed GeglBufferRectangle this change also retains the existing released API
and ABI more exactly.

The 2x2 matrix is only used by GeglBuffer and its API, gimp-2.10 does not
use the symbols - we are remaining abi compatible but are rearranging
locations of symbols.

This reverts commit 3bae8f7e.
The effect of changing NOHALO_OFFSET_0 needs further investigation
momentarily reverting.

Remnants of removed and buggy GeglBuffer mipmap awareness in resampler.

This does the same modification that preceding commits did to lohalo, for
sigificant downsampling this might result in sharper/aliased results, and maybe
the user would be better of having chosen linear or cubic resamplers which use
a box-filter for scaling down. A similar box-filter might at a later stage be
integrated as a replacement for the mipmap using code that were in nohalo and
lohalo.

This should fix bug #763557

Making the samplers slow again, it seems like the parameter cannot be used
directly as a speed/performance trade-off.

For affine like buffer sampling a 5.23x speed improvement for nohalo and 2.45x
speed improvement for lohalo. This tunes the algorithms for more reasonable
performance; without much loss of quality.

Implement an adaptive caching policy; that meets the needs of affine
transformations. Running averages are kept of x/y deltas as well as the
magnitudes of the deltas of coordinates leading to cache misses.

The initial rectangle used is a small one; subsequent rectangles have extents
up to 64 pixels wide/tall; when tracked deltas are more diagonal than axis
aligned smaller more square regions are tracked.