SPPM renderer (CPU-only)

[binarycortex]

film.AddSampleCount(float(film.GetXPixelCount()*film.GetYPixelCount()));

in line 554 of hitpoints.cpp (just after Film &film(*scene.camera->film);)

I had something similar to this part already...

and I stop the rendering by adding

if (scene.camera->film->enoughSamplesPerPixel)
{
renderer->Terminate();
}

in line 608 of sppmrenderer.cpp (just after hitPoints->IncPhotonPass();)

But this is the part i was missing.

[binarycortex]

Idenctical scene, 3 passes each. Lowering startradius to 0.5 or so prevents the grid from being smudged out, so perhaps this is how it's supposed to work?

This is a consequence of using the BSDF of the photon hit point and not the eye path hit point. This was introduced and required by the latest patch, however as you have observed, it is noticeable only with very large start radius.

Don't forget that the initial radius is defined by an heuristic and it can indeed pick up a quite wrong value.

Dade, is that the reason for the outlining on the luxball ring and the base in this image?

startradius 2.0, Alpha 0.7, 1mil P/pass

I noticed that when I start with an insanely small startradius this does not happen.

startradius 0.1, Alpha 0.9, 1mil P/pass

[Dade]

Dade, is that the reason for the outlining on the luxball ring and the base in this image?

Yup, this was avoided by a normal check between hit point normal and photon hit point normal in previous version but I can restore that check. It will help to avoid case where the BSDF of the photon hit point has a normal too different from the one of eye path hit point.

[J the Ninja]

Dade, last night in IRC BinaryCortex and I were messing around a bit with the eye pass sampler, and found the 20%+ speedup he reported with hilbert does indeed work with linear and tile as well. Wouldn't linear avoid the problem of "one tile is harder than the others" you mentioned with hilbert? Also, when used with low-discrepancy sampler for the classic renderer, the threads work on different rays in the same block, rather than on different blocks. I assume this is different for SPPM?

[Dade]

Dade, last night in IRC BinaryCortex and I were messing around a bit with the eye pass sampler, and found the 20%+ speedup he reported with hilbert does indeed work with linear and tile as well. Wouldn't linear avoid the problem of "one tile is harder than the others" you mentioned with hilbert? Also, when used with low-discrepancy sampler for the classic renderer, the threads work on different rays in the same block, rather than on different blocks. I assume this is different for SPPM?

"Tile" would suffer of the same problem because a continuous group of adjacent pixels (i.e. a group of adjacent tiles) are assigned to the same thread. But the easy fix to this problem is just to give to the user the freedom to choose the pixelsampler to use.

[binarycortex]

Dade, last night in IRC BinaryCortex and I were messing around a bit with the eye pass sampler, and found the 20%+ speedup he reported with hilbert does indeed work with linear and tile as well. Wouldn't linear avoid the problem of "one tile is harder than the others" you mentioned with hilbert? Also, when used with low-discrepancy sampler for the classic renderer, the threads work on different rays in the same block, rather than on different blocks. I assume this is different for SPPM?

"Tile" would suffer of the same problem because a continuous group of adjacent pixels (i.e. a group of adjacent tiles) are assigned to the same thread. But the easy fix to this problem is just to give to the user the freedom to choose the pixelsampler to use.

I was working on this, but it wasn't going all that well since I'm new to c++. Perhaps it's time to try again.

I did however get halttime and haltspp to work. I am using pass count for haltspp. Are you ok with me committing this Dade?

[Dade]

I did however get halttime and haltspp to work. I am using pass count for haltspp. Are you ok with me committing this Dade?

As long as we stay in the experimental branch, we can test/try/commit whatever we want.

[J the Ninja]

I did however get halttime and haltspp to work. I am using pass count for haltspp. Are you ok with me committing this Dade?

As long as we stay in the experimental branch, we can test/try/commit whatever we want.

Great, because he implemented user-selectable pixelsamplers while you were asleep.

[SATtva]

Dade never sleeps! When you (and I and everyone else) sleep he's planning for world domination!

[Dade]

Dade never sleeps! When you (and I and everyone else) sleep he's planning for world domination!

MOAR CATZ !