Maybe try it on Linux and then think about the whole Windows build tool SNAFU.

Some encoding hacks on the nvidia branch. Only seeing ~ 7fps on NVIDIA GeForce GTX 980 TI (with arrietty); maybe I'm doing something wrong.

Current showstopper is it doesn't seem to support custom precinct sizes; setting enable_custom_precincts to 1 gives a NVJPEG2K_STATUS_JPEG_NOT_SUPPORTED (I think I'm using 0.5.0 of nvjpeg2k).

It seems that there's more optimisation (pipelining) in the decoder than the encoder at the moment. Perhaps the encoder will get better.

Fixed a few things to make it actually work and the FPS is closer to 10.5.

Translating the PSNR to our bitrate still needs doing.

Decoding seems not to work with 4K (nvjpeg2kStreamParse gives NVJPEG2K_STATUS_INTERNAL_ERROR)

2K playback seems OK, about 1 core being used on i7-7700 with GeForce GTX 980 Ti to play back a DCI flat DCP (BTaER).

I put a message on nvidia's foruma about precinct sizes:

https://forums.developer.nvidia.com/t/nvjpeg2k-enable-custom-precincts/215482

https://forums.developer.nvidia.com/t/nvjpeg2kstreamparse-error-on-4k-jpeg2000-stream/202551 says they're already on the 4K problem.
Fixed apparently in 0.6.0.

Got a sponsorship offer to implement this for playback on Linux.

It seems that as of v0.8.0 of nvjpeg2000 custom-precinct sizes are now supported, perhaps a new test is in order.

Was able to use mingw in Linux to compile a simple win64 exe test program that initializes nvjpeg2k library and queries it for basic information. Now working to test encoding a some video frames.

N:\dev\jp2ktest>jp2ktest.exe
nVidia JPEG2000 Test...
nvjpeg2000 library initialized successfully!
nvjpeg2000 version: 0.9.0
Library cleanup complete.

Was able to encode raw frame image files to (hopefully) DCP compliant jpeg2000 frames using nvjpeg2000, passed along example source code to Carl. There is a lot of work done to be here to not only try and make this cross-platform for Linux/Windows/Mac using the same cuda binary compiled function, and even more work probably to get the GPU threading done right so it encodes faster than a CPU. Ideally the target would be at least 4x~8x faster than CPU encoding.