DAV1D Experimenting With Vulkan & OpenGL ES GPU Offloading

[phoronix]

Phoronix: DAV1D Experimenting With Vulkan & OpenGL ES GPU Offloading

There isn't any AV1 video decode/encode built into the video engines of today's GPUs, but the DAV1D project CPU-based AV1 decoder is experimenting with offloading some aspects of the process to current generation hardware with OpenGL ES and Vulkan...

http://www.phoronix.com/scan.php?page=news_item&px=DAV1D-Vulkan-GLES-Experiment

[FireBurn]

I've got a horrible feeling this won't be more efficient - would love to see some stats showing the benefit though

[bug77]

I've got a horrible feeling this won't be more efficient - would love to see some stats showing the benefit though

Don't worry, it's in a branch. Easily discarded if it doesn't deliver

[tildearrow]

I wonder if x264 also has full offloading. Yeah, I know it does, but it is only for specific steps and does not help at all.

[pal666]

i hope it works out, because for several years majority of players in use will have vulkan/gles but no av1 decode

[andreano]

Another supposedly GPU offloadable AV1 feature is film grain, though I haven't seen that in use yet.

CDEF offlading would be great, as cdef_filter_block_c was no stranger in the performance reports (if you didn't have AVX2 or SSSE, I suppose, which I could also reproduce).

[dwagner]

Another supposedly GPU offloadable AV1 feature is film grain, though I haven't seen that in use yet.

That is because not only it saves CPU not to add "film grain", it also improves the picture :-)

There are crazy people out there who even add "film grain" to computer generated game graphics, as if an analog camera flaw was something great. And then, only after players complain, make it optional, as just recently happened to "Control" - https://controlgame.com/control-comm...pdate-30-8-19/

[coder]

#version 310 es

Because they're using compute shaders, rather than fragment shader hacks. Of course, that rules out Raspberry Pi 4 support, for the time being.

The initial patch just accelerates a few simple neighborhood filters. It's not obvious how much benefit it'll provide, but I guess it's a starting point.