[Bf-committers] using checksum instead of bundling images for regression tests

Anders Nilsson bf-committers@blender.org
22 May 2004 21:07:06 +0200

I think checksums sound a bit naive considering small differences in
FPU's due to different compilers settings etc. Another alternative might
be to compare the rendered picture to the reference picture, pixel by
pixel, calculating the error. Then within a threshold the picture would
be OK. Once could keep track of maximum error, sum of errors etc. This
is very easy/fast to do, only problem is that ref-pic must be included.
This would also catch problems that's not visible (like pic1 is brighter
than pic2 but otherwise they look ok, people might not notice).

Don't know if this voids the purpose of the checksum but it might work
to certain extents. Also rendering quality might change slightly from
release to release due to different things, rendering code must be
allowed to improve slightly quality-wise. This is not a problem if the
reference picture was generated once per release though.

Anders Nilsson (breakin)