GPU - IPR vs Picture Viewer render time

[peterDTE7K]

Hi, the GPU works great in IPR, it's a significant improvement of render speed vs the CPU. However when I render into Picture Viewer the render time often increases over 15x in comparison with IPR when using GPU and the same sample settings. Do you know what might cause this difference? When using CPU the render times in both IPR and Picture Viewer seem to be very consistent. Thanks!

[Stephen.Blair]

Did you have Arnold logs for the different renders?

[peterDTE7K]

Here are the logs. IPR took 32 seconds and Picture Viewer 4:26 with same settings.

[thiago.ize · Accepted solution]

Slow render had all these AOVs which fast render did not have:

00:00:04 7307MB | [aov] registered driver: "/Drivers/_display_driver_" (driver_c4d_display)
00:00:04 7307MB | [aov] * "RGBA" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "diffuse_direct" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "diffuse_indirect" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "specular_direct" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "specular_indirect" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "transmission" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "sss" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] registered driver: "/Drivers/Arnold_Driver" (driver_exr)
00:00:04 7307MB | [aov] * "RGBA_teal" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_red" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_purple" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_dome" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_orange" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_green" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_distant" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_blue" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "RGBA_default" of type RGBA filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "denoise_albedo" of type RGB filtered by "default_filter" (blackman_harris_filter)
00:00:04 7307MB | [aov] * "N" of type VECTOR filtered by "default_filter" (blackman_harris_filter)

And the log says the majority of the time was spent writing out the data (output driver):

00:04:23 10911MB | rendering 4:18.48
00:04:23 10911MB | output driver 3:56.66
00:04:23 10911MB | pixel rendering 0:21.82 (130 sample iterations @ min: 0:00.317, avg: 0:01.955, max: 0:03.757)

So you'll want to take a look at your AOVs and how you're writing them out.

[peterDTE7K]

Cool, many thanks. Indeed removing the AOVs, especially those from driver_c4d_display makes a huge difference.

Question - why does Arnold need to render all driver_c4d_display AOVs when rendering to Picture Viewer? It doesn't make much sense, I thought these AOVs are only used for preview in IPR. Also the driver_c4d_display AOVs seem to significantly impact Picture Viewer rendering but only when using GPU. CPU rendering is mostly unaffected. If I knew this before I could have saved weeks of rendering time.

[peter_horvath]

why does Arnold need to render all driver_c4d_display AOVs when rendering to Picture Viewer? I thought these AOVs are only used for preview in IPR

---

The display driver is a bridge between the Arnold render output and the Cinema 4D frame buffer, it basically manages all render processes, such as Picture Viewer, IPR, Team Render, Command line render. AOVs assigned to the display driver are active AOVs in the above render processes, thus they are written out to PV, Team Render, etc.
The IPR window has an option to render all active AOVs or only the currently selected AOV (Render menu > Render all AOVs). If you don't want an AOV to render out, but only look at it in the IPR window, don't assign it to the display driver, and select it from the Inactive group of the AOV combo box.

---

Also the driver_c4d_display AOVs seem to significantly impact Picture Viewer rendering but only when using GPU.

---

There are way more data update calls during a GPU render compared to CPU, and probably that's where the performance suffers. Hopefully this is something we can fix in an upcoming release. Thank you for reporting the issue.

[peterDTE7K]

This is great info Peter, many thanks. Now I can finally fully switch to using just GPU.

[peter_horvath]

The slow down should be fixed in C4DtoA 4.6.5. If you still have problems in this version, please let me know.

[peterDTE7K]

Back to full speed, also IPR seems to start faster. Free panning IPR, spacebar on/off, composition overlays....this update is golden!

[no188328]

Can you share your GPU sampling setup with us please to get a clean render? No matter what I do the GPU renders looks very noisy compared to CPU, even tho with my adaptive threshold is set to 0.001 and the AA smaple Max is 100 with camera AA set to 6 or 7. I still get a noisy results with simple scenes.

[peterDTE7K]

For production renders I use CPU. GPU is still often unreliable with some weird artifacts however it is much faster so I use it for previews, tweaking scene and materials. I rarely go above 30 MAX AA with default threshold. 100 MAX AA is huge. Are you using some heavy DOF or motion blur? CPU is often having less problems with those.

If you want you can send me direct message with some scene and I will check it.

[no188328]

I use it for previews as well but I thought you figured out a way to use it for production renders when you said "Now I can finally fully switch to using just GPU" that's why I was curious. My GPU is stronger than CPU that's why I was searching for a way to relay on it more but I think I'll have to stick with the CPU for final renders for now. Thanks a lot for your notes and proposition I'll definitely take it into consideration, really appreciated. Have a great day.

[peterDTE7K]

If I see no problematic difference between CPU and GPU I will use GPU for production especially for animations since it is multiple times faster. If you have hair you need to increase specular ray depth and if you have transmissive objects like glass you need to increase both specular and transmission rays to get same results as on CPU. Also light samples matter. 

[peter_horvath]

Can you tell us what parts of your scene are not cleaned up properly on the GPU? It would help us focusing on those areas. Would denoising help?