aa on gpu render on cpu?

[losbellos73]

Have you ever thought rendering the image and running the aa pass on the gpu? Even several times, for better result. GPU could have very fast AA even on high res. There are a lots of programmable filters there. Maybe it would work.. Lots of rendering times can be saved.
What do you guys think, is it possible?

[vlado]

It might be possible for some things, certainly... will have to think about it.

Best regards,
Vlado

[Mark Thompson]

Proper antialiasing isn't just a blur filter, it actually involves rendering more (sub-pixel) samples for edge pixels. In other words, to get proper antialiasing you would have to upload the image data to the GPU, which would be very slow with most current graphics cards.

I think that's the same reason why V-Ray RT gives you the option between CPU and GPU, but cannot use both CPU and GPU at the same time.

However, using AMD Fusion chips, the connection between CPU and GPU is a lot faster, so maybe a hybrid renderer would be possible.

[ghiom3d]

Proper antialiasing isn't just a blur filter, it actually involves rendering more (sub-pixel) samples for edge pixels. In other words, to get proper antialiasing you would have to upload the image data to the GPU, which would be very slow with most current graphics cards.

I think that's the same reason why V-Ray RT gives you the option between CPU and GPU, but cannot use both CPU and GPU at the same time.

However, using AMD Fusion chips, the connection between CPU and GPU is a lot faster, so maybe a hybrid renderer would be possible.

Hybrid rendering already exist . CPU+GPU renderers are working very great.
You can modifying to your seetings to make VRAYRT hybrid ( CPU+GPU)
I thing chaosgroup have decide to split this to way of rendering to make a CPU version with more features than it's possible with the GPU at this time.

[Mark Thompson]

Hybrid renderers need to either duplicate a lot of data (which results in a waste of CPU/GPU cycles and memory) or send data back and forth between the CPU and GPU (which can take very long, for complex scenes with large textures). Neither of those solutions is optimal.

Integrating CPU, GPU and memory controller on the same chip (like AMD is doing with Fusion APU chips and Intel will also be doing with their new product families) will make it possible to share data between CPU and GPU, as well as improve bandwidth and latency, since they won't need to go through the PCIe bus. In 3-4 years I expect most renderers will take advantage of that architecture and we'll see big improvements even in the "production" versions of V-Ray Advanced / Mental Ray / Maxwell, etc., not just a simplified "RT" version running on discrete GPUs.