progressive sampler gives different results when render region is applied

[p0las]

We have been using adaptive sampling but seeing all the push for the progressive we decided to try it in the pipeline.
Our usual approach is to work out the sampling setting (min/max subdivs, contrast etc) on a little region. After a fair bit of experimenting on my sample region I ended up with settings I liked:
min shading rate: 6
min/max subdivs 1-15
noise threshold: 0.025

the noise threshold is higher than expected and max subdivs lower than usually but the result was smooth enough.

however the same part of the image when rendered without region looks very different and has much more noise.

the right one is a section from full res render

Looking at render stats they are not even close.

RENDER REGION:
Sampling level reached: 15.20 subdivs (622 passes)
Number of raycasts: 518043891 (6245.71 per pixel)
Camera rays: 12991553 (156.63 per pixel)
Shadow rays: 135790702 (1637.14 per pixel)
GI rays: 69601802 (839.14 per pixel)
Reflection rays: 70941 (0.86 per pixel)
Refraction rays: 0 (0.00 per pixel)
Unshaded rays: 318002894 (3833.95 per pixel)
Number of light evaluations: 264101600 (127.36 per pixel)
Number of intersectable primitives: 162836


FULL FRAME:
Sampling level reached: 15.20 subdivs (801 passes)
Number of raycasts: 5600395985 (2700.81 per pixel)
Camera rays: 168959981 (81.48 per pixel)
Shadow rays: 1472087601 (709.92 per pixel)
GI rays: 824378546 (397.56 per pixel)
Reflection rays: 2659636 (1.28 per pixel)
Refraction rays: 0 (0.00 per pixel)
Unshaded rays: 3339764198 (1610.61 per pixel)
Number of light evaluations: 2856628836 (1377.62 per pixel)
Number of intersectable primitives: 162836


in both cases the same sampling level was reached 15.20 but the amount of passes was different. Why isn't it consistent? Other progressive renderers like maxwell do not behave that way.
is this how vray works? or a bug? can someone explain a bit how the algorithm works and why the passes seem to dynamically change?
shouldn't the noise be exactly the same with the same contrast threshold?

thank you,
Pawel

[vlado]

Do you get a similar message in the V-Ray output log for the full-res image?

Estimated progressive buffer memory usage of XXXX MB is over the limit of 2048 MB. Falling back to per-pixel filtering.

Other progressive renderers like maxwell do not behave that way.

The passes in V-Ray are not the same thing as passes in Maxwell. The subdivs value is the important one, not the number of passes. Also, Maxwell does not perform adaptive sampling. If it makes you feel better, we can hide the the "passes" number and only show you the reached subdivs.

can someone explain a bit how the algorithm works and why the passes seem to dynamically change?

The subdivs value is the important one, not the number of passes. The full resolution image is more complicated and it may contain parts that require more passes to reach the specified noise threshold.

Best regards,
Vlado

[p0las]

Please do not hide the passes. The more info and control we get the better. I'm trying to understand what is happening under the hood to be bale to better predict what settings are needed and to tweak render times. This difference in renders was very unexpected and left me baffled for a bit. I thought I made a mistake somewhere and did all the tests again to make sure.

the only message like the one you mentioned that I got was:
[2016/Mar/17|18:27:56] V-Ray: Progressive sampling base buffer memory usage: 428.09 MB

this is for the full resolution farm render.
thanks.