Film - the bottleneck

[SATtva]

Want me to keep it in the test branch or?

I'd be glad to see it in the trunk. Frankly, this is the only thing keeping me from making the new weeklies.

[Lord Crc]

I'd be glad to see it in the trunk. Frankly, this is the only thing keeping me from making the new weeklies.

Well... I felt frisky, I've pushed the merge. Here goes nothing...

It compiles here, and for all I can see works but may be some issues still.

edit: Some initial banding is to be expected, since some buffers may not be filled before the initial tonemap. I think we need to add some more flushing before outputting images, I'll try to get this sorted ASAP.

[binarycortex]

HOLY COMMIT MESSAGE BATMAN! What was that, a novel.

Edit: That's even funnier because of your avatar.

[Lord Crc]

Hehe, well guibou wanted more text

I'll add some more comments to the film code once we're satisfied with most of it.

[SATtva]

Awesome! Thanks a lot.

HOLY COMMIT MESSAGE BATMAN!

[Pilchard123]

Some days, you just can't get rid of a bug!

[Lord Crc]

Just wanted to add this encouraging info thanks to JtheNinja, who managed to run some tests on a machine with dual Xeon X5670's for a total of 24 threads.

Setting tiles=1 effectively reverts the code to the old behavior, while tiles=0 lets the code decide the number of tiles to use. The "k=5" is the outlier/firefly rejection parameter, 0 disables it.

<JtheNinja> LordCrc: so i just tested your tiled buffer on a 24 thread machine
<JtheNinja> tiles=1, k=5, ~250ks/s
<JtheNinja> tiles=1, k=0, ~600ks/s
<JtheNinja> tiles=0, k=5, ~750ks/s
<JtheNinja> tiles=0, k=0, ~950ks/s

So, even with outlier rejection disabled we're seeing a very nice speed increase of almost 60%, and with outlier rejection it's a massive 200% increase.

[jeanphi]

Hi,

Some would call it an achievement. Nice job!
It also sets things straight: with one good algorithmic decision you win more than all the micro optimizations we can do.

Jeanphi

[xsl]

Hi,

I was doing some tests with Amazon EC2 instances and it looks like I'm hitting this bottleneck even with the new code (1.0 RC3). I was getting strange (very low) performances when running a cc2.8large instance (2 x Intel Xeon E5-2670, eight-core "Sandy Bridge", 32 threads). See this post and the next for the details. The short of it is that I had to run 6 slaves on this instance to max it out. My image was small (960x540) so I'm sure it's contributing factor.

Could there be something else going on?

[Lord Crc]

The short of it is that I had to run 6 slaves on this instance to max it out. My image was small (960x540) so I'm sure it's contributing factor.

Yes, that is likely the cause. I'll see about making the "tiles", or slabs rather, vertical instead of horizontal, given that most people render wide images. Increase the resolution and you should see it scaling better.