LuxRender Windows CVS build of 12-08-2010

[Lord Crc]

SSE1 and SSE2 are some processor/CPU extensions, used to make the program faster. If you have an old(!) CPU then you won't have SSE2 and you must use the SSE1 build.

These two are only relevant for x86 (32bit).

It may be that the problem was only apparent on x64, I haven't checked that...

[Abdollah 4 Ever]

thanks again

[xsl]

I'm getting major slowdowns and major regression using photometric lights. Here are some numbers and examples. The scene is a sample scene comparing a square patch and an 50W MR16 model with photometric lights. All cooked to 100spp.

Release 11.61min 65496TotS/s
Precise 35.96min 21179TotS/s
Sloppy 35.91min 21210TotS/s

Release, Sensitivity:50, Exposure:1/125, Fstop:2.8, Gamma:2.2

Precise, Sensitivity:50, Exposure:1/125, Fstop:2.8, Gamma:2.2

Sloppy, Sensitivity:50, Exposure:1/125, Fstop:2.8, Gamma:2.2

Sloppy, Sensitivity:20, Exposure:1/1000, Fstop:11, Gamma:2.2

Release, Sensitivity:20, Exposure:1/1000, Fstop:11, Gamma:2.2

Here is the scene and the photometric data files.

PhotometricData.zip

(62.04 KiB) Downloaded 14 times

[Lord Crc]

Thanks for the feedback xsl, this is probably due to the huge internal changes jeanphi did a week ago.

[binarycortex]

Reguarding slowdown. Its not that much for me. I am using educos' scene he posted in the user support (known firefly issues) forum to test. Both were only rendered for 2 minutes just to get an average samples per second.

standard build = ~54000 samples per second

precise build = ~50500 samples per second

Next I am going to render out for fireflies.

[jeanphi]

Hi,

I'm getting major slowdowns and major regression using photometric lights. Here are some numbers and examples. The scene is a sample scene comparing a square patch and an 50W MR16 model with photometric lights. All cooked to 100spp.

If you're not using the bidirectional, the speed drop is most probably due to the use of the new interfaces instead of the old ones. Now that we don't have to support 2 sets of interfaces it might be easier to optimize the new ones.
Regarding the rendering differences, I just found a nasty bug. A fix is in the repository, can you confirm that it's now working as expected?

Jeanphi

[xsl]

If you're not using the bidirectional, the speed drop is most probably due to the use of the new interfaces instead of the old ones.

The scene uses bidir.

[jeanphi]

Hi,

Then I hope the speed drop is caused by the bug, otherwise I can't explain it, bidir should be slightly faster with the new code.

Jeanphi

[patlecoq]

Hi,

I have made several tests, and I have not found any slowdown with the "precise" version, it is even a little faster (2%).

Patrice

[binarycortex]

My previous test was using luxconsole for both the fast math and precise math. This time i tried the luxrender gui. Both the fast math and precise math versions are giving about the same speed. ~14500 samples per second average. 0.7 release gives ~51000 samples per second in luxrender gui. The only real speed comparison I could do was with the console and if it does reduce fireflies (having trouble reproducing even with 0.7) then the speed drop is acceptable. But the drop in the GUI is unacceptable.