Atmospheric simulation and GPU

[yonosoy]

Hi.
Just to inform about the circumstance. GPU is now able to sample a little bit the atmospheric simulation. Some hope for a future aperture.

[Oscar J]

Isn't the GPU rendering as accurately as the CPU?

[yonosoy]

Hi Oscar, maybe now the question is understanded ...
Is about a bad expression (from our side); all the samples are correct, but only do a litle numbers of they in the time ( the performance is really low in this moments, jumping from stop (most of the time) to a very good count; and this modality of process is not possible until the 3.8 serie).

Actually a test with the last heavy scene at SS 1 is running with incredible results. If a stable count of samples is achieved the speed for this setups could be highly improved (is again a feeling, is the first serious test).
In the try to almost give to this test 1000 spp to observe the noise (with 440 is surprisingly good !) .

Update: Remember SS 1. Lefting the run in the most stable machinne to 2-3 K if possible.

[yonosoy]

Well, for the right moment. Seems promising.

[Oscar J]

16 h for that small image? That doesn't seem very promising to me.

[yonosoy]

Two GPU in this test. Only performs samples in a very low ratio and better for SS 1.
Just believing that was a good new.
Thank you and sorry for the dissapoint.

[Oscar J]

Didn't mean to sound harsh or anything. Just, what's the point of rendering with GPU if it's so slow? Or could there be any optimisations for GPU so that it renders faster?

There's no disappointment mate. You're a valued member and an Indigo genius, and your (and CTZn's) threads have been the most fascinating by far on this forum lately.

[yonosoy]

Hi Oscar, very glad for your words.
The point is that in this moments the algo process an estimated 5 % of effective time. It seems to be the same speed with/without the volume of clouds of 160x160x2 km. Is not so much information but with the plus of a very good convergence in a very noise scene... The spp are 1/5 approx than with PT simple and this mean some hope about speed. Also the number of polygons required for the subdivision in this scenes is very low and fits in the capabilities of RAM of a GPU.
Could be interesant, but also maybe very hard to developed. So the intention at the moment is to info of the experience and the possibilities...
Thanks a lot .

[CTZn]

thanks for the compliment Oscar, although it's all relative since many people are prefering cars to clouds irl
yonosoy, dijiste que el media atmosferico llegava a que altura ? A los 80km ahora por el ozono ? Es cierto esto ?

I've always been discarding GPU for atmospheric tests because my scenes use sphere primitives !

[yonosoy]

...because my scenes use sphere primitives !

And the bells rings from far away... with a atmosphere_mesh:

Something happens with the perfect spheres beyond a size:

from globe_test in atmosphere at 10 m of height. (CPU PT)

In this case GPU isn´t able to process well with a perfect sphere like atmosphere shell. With a mesh, performs satisfactory but please, sure that could process more rounded !!! Still doing accused jumps... (little GPU).
The test is with henyey-greenstein phase function, 12 minutes test.
CPU : 23.36 s/px
GPU: 54.17 s/px

CTZn, thanks fir the tip! The terrestrial atmospheric medium starts at 6378135 m and ends at 6458135 (yes 80 Km.). The inclusion of the ozone maybe changes the distribution and the specific real tabulated medium mixture of gases but don´t modify the range of expression, though.
Thanks.

[Zom-B]

In this case GPU isn´t able to process well with a perfect sphere like atmosphere shell. With a mesh, performs satisfactory but please, sure that could process more rounded !!! Still doing accused jumps... (little GPU).

GPU renderings doesn't support sphere primitives! Maybe in this case its different, but beware...

[yonosoy]

Firstimer with GPU (*), with patient the atmospheric shell works for this case but maybe at risk of the hardware ...
Don´t know more ...
With an atmosphere mesh seems to be mostly correct.
The runs are usually long so, i take this tests carefully; if done.
Thanks Zom-B.
(*) maybe because of this never worked before this atmospheric test, (ever perfect spheres); the surprise is that for 3.8 begins to work (??).

[yonosoy]

Sorry, finally the bug was not specify (because the ability of process the atmospheric simulation with GPU probably was there time ago):
When there is defined a perfect sphere for the earth, materials like specular and phong works well, but materials like diffuse and oren-nayar gives a black zone when the camera is below (don´t remember but near 60 meters of altitude).
This is and old bug that was forgoten !! Don´t know if has any use, the try years ago was to study objects with atmosphere simulation (has no meaning , but i like). Actually is understand that all this means to overload indigo innecesarily. But once the atmosphere is defined, a downspeed for an earth was never noticed for a basic material (?).

So nothing specific to request only inform that the GPU atmospheric simulation with clouds is a lot faster.