Since now some models are comverted to tensorrt engines. Wouldnt it be better if we had the option to build the engine locally instead of us all using the shipped “generic” engines? Tensorrt benefits massively from building engines per config instead of using generic ones and can further improve perf for the people with Rtx gpus
If I remember correctly, Blackmagic Design introduced this with Davinci Resolve 18.6 for the AI parts of the program. After the installation of the new major version, launching it for the first time had an “Optimize DaVinci Neural Engine” stage.
This stage took some time, but it’s been done once.
I believe it’s the same thing. And the results were as expected. Everything using Blackmagic Design’s “Neural Engine” got way faster. Like their Depth Map OFX, or Magic Mask.
It could be cool to have this option. And for the fallback, it would just use the optimized version from Topaz. Have a setting to build the tensor engine for a particular model.
As you can see, in Davinci Resolve, we can activate or deactivate this particular optimization based on someone’s unique system and fallback on how Davinci Resolve uses out of the box:
1 Like
