"clipka" <nomail@nomail> wrote:
> - I've been able to come up with a solution that allows for 100% reproducable
> shots with multiprocessing, at a performance cost of only 10%-20%.

Exhaustive test results are as follows:

- Most scenes I tested show a perfect match between file output of consecutive
renders, or sparse stray pixels typically differing by 1/255. Occasionally,
individual pixels differ by as much as 2/255. There is no clear structure to
the occurence of these differences, making analysis difficult. The absence of
banded structures indicate that sample placement is a perfect match, and
current assumption is that the differences are caused by rounding effects in
conjunction with different ordering of samples in the sample cache. Given the
small order of magnitude, these differences are probably not enough of an issue
to really bother about them.

- Some Low-quality scenes show areas of uniform difference of 1/255, with a
structure indicating that sample placement may not be matching perfectly, for
yet unexplained reason. However, this is only a guess, and the structure may
have other causes; given that these differences are of minor order of magnitude
as well, and only occur in low-quality shots, they don't seem worth bothering
about either.

However, three scenes deverse special mention:

- In the "balcony.pov" and "cornell.pov" scenes, output file pixel values differ
by up to 80/255 and 16/255, respectively, in the partially shadowed areas; there
is reason to assume that these differences are due to jittered area lights in
both cases. In fully shadowed or unshadowed areas, only individual stray pixels
of 1/255 difference could be observed.

- In my almost-closed-door scene, pixel values differ by up to 21/255 (the
maximum differences being in the bright corner of the room, which is an
artifact). The differences show as grainy splotches, indicating both slight
differences in the brightness of individual samples, but at the same time a
perfect match in their placement. Again, these differences could be caused by
rounding effect differences, due to variations in the ordering of samples in
the sample cache. The comparatively high order of magnitude may be due to the
exceptionally bright illumination used for the shot.

- The "balcony.pov" scene could not be compared yet, because in one of the two
runs it once again crashed with a segfault (at address 0x00000000 this time).

Rendering time increases by close to 40% in a single case, but mostly stays
within an added 10%-20%. Measured over my whole batch of test scenes, "wall
clock" time increases by 10.8%. Only a small portion of this increase is due to
idle cores; in fact, CPU time increases by 10.5% as well.

As I can well imagine that some people might prefer speed over reproducability,
I suggest adding a special command line parameter to switch on or off any
mechanisms that improve reproducability at the cost of speed.

Post a reply to this message