DaVinci Resolve 18.6 and TensorRT in the container

[fat-tire]

(I'm opening this discussion to discuss the TensorRT support in 18.6.)

Among the new features in this 18.6 release is the following:

• Up to 2x faster Neural Engine performance with Nvidia TensorRT.

Up to 2x faster sounds like a big win!

When starting up 18.6 for the first time in the container, a dialog popped up that said:

Optimize DaVinci Neural Engines

There are one more new DaVinci Neural Engines that need to be optimized for optimal peformance on NVIDIA GPUs. This process might take several minutes.

(the options are Disable/Skip/Optimize)

I don't think TensorRT is installed currently, so I tried adding it manually.

./resolve.sh /bin/bash

This starts the container and puts you in the shell. I then did this:

pip3 install --upgrade pip
pip3 install --extra-index-url https://pypi.nvidia.com tensorrt-libs

This seemed to install the tensorrt library as well as cudnn as seen here:

$ pip3 list
Package                  Version
------------------------ --------
ffmpeg-python            0.2.0
future                   0.18.2
nvidia-cublas-cu12       12.2.5.6
nvidia-cuda-nvrtc-cu12   12.2.140
nvidia-cuda-runtime-cu12 12.2.140
nvidia-cudnn-cu12        8.9.4.25
pip                      23.2.1
setuptools               50.3.2
tensorrt-libs            8.6.1

The install persisted between reloads of the container.

Not sure if this made any difference while doing the optimization, but for what it's worth you can verify that the GPU Optimizations are turned on by selecting:

DaVinci Resolve (top menu) -> Preferences -> Memory and GPU item -> Use neural optimization on NVIDIA (checked)

For fun I also added:

pip3 install --extra-index-url https://pypi.nvidia.com tensorrt

At which point I could access tensorrt in python:

bash-4.4$ python3
Python 3.9.16 (main, Jul  3 2023, 20:07:32) 
[GCC 8.5.0 20210514 (Red Hat 8.5.0-18)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import tensorrt 
>>> print(tensorrt.__version__)
8.6.1

Did a quick magic mask and the neural stuff still works. Can't tell if it's faster or if it's using tensorrt or not...

Questions: Should TensorRT be added automatically by build.sh? Only via a $RESOLVE_FLAG? Does it make any difference whether it's here or not? Was it installed correctly via pip3 or would an .rpm have been the right way to go?

Your thoughts are much appreciated!