Migrate to Jitify2

[Robadob]

• RTC and Execution (Works with CUDA12.0, Windows/Linux)
• Serialisation/Deserialisation (jitify2 pre-processed serialised objects are 2-3x larger)
• CUDA 12.3 support
  • Works with Jitify2 misc fixes NVIDIA/jitify#128, hence wait for merge.
  • Same branch deprecates the launch() method (used in CUDASimulation), says to replace with launch_raw().
• Investigate access violation from cuda().ModuleUnload() during sim shutdown, when CUDAAgent map is cleared by CUDASimulation destructor.
  • Has occurred on Windows & Linux, but its inconsistent.
  • gdb log: https://gist.github.com/Robadob/3cb0e93014d56f05587f4ea1ea581203 (with jitify2-misc-fixes2 branch)
• Reimplement jitify1 demangle used in curve_rtc.cpp
• Optimise serialisation load time
  • Jitify2 serialises pre-NVRTC, this is 50x slower and produces 2x larger serial blob
  • Waiting for clarity: (jitify2) Serialisation post-NVRTC NVIDIA/jitify#133

• Optimise compile time
  • Rework of old header hack? (would require loading headers from file)
    • Preloading fgpu headers only cuts agent fn time from 6.8s to 4.1s.
    • Loadings CUDA headers too makes a big difference, but these may not be particularly stable between version
  • Offline pre-process FLAMEGPU2 include hierarchy into a single header file? (using jitify tools?)
  • Wait for this PR to be merged?

• Visual Studio 2019 support (we may be able to drop this)
• ManyLinux2014 support

[Robadob]

Not sure if this is as narrow as it could be, but I wanted to make minimal changes to get it working.

The wildcard on libnvjitlink-dev can probably be removed.

Same issue in the ubuntu script

[Robadob]

Update regarding header pre-loading with Jitify2/CUDA 12.3

Windows/CUDA 12.0

No preload
Millis: 6822.000000
Millis: 6853.000000

Preloading FLAMEGPU headers
Millis: 4045.000000
Millis: 4277.000000

Preload FLAMEGPU + CUDA headers
Millis: 1296.000000
Millis: 1667.000000

Linux/CUDA 12.3

Jitify 2 from scratch (Waimu)
Millis: 25318.000000
Millis: 24143.000000

Preload FLAMEGPU + CUDA headers
Millis: 1376.000000
Millis: 2218.000000

CUDA 12.0 has ~30 CUDA headers to preload.
CUDA 12.3 has ~257 CUDA headers to preload. (List contains some dupes)

Not clear whether we would want to generalise this code, to better handle different CUDA versions, because we could be potentially needing to update it with each CUDA update.

Edit: Removed from-cache times, latest commit has these matching Jitify1.

[Robadob]

Current issue holding back the Jitify2 preprocesor branch is that it expects our flamegpu headers to be included as system header <> rather than " ". Waiting to here back from the dev (Ben) before I try to correct that on our side.

[Robadob]

Did three full test runs last night, all passed, however in those cases the cmake jitify dependency was pointing at the preprocess branch. Not currently using that here as it causes all windows CI to fail with WError.

Linux/CUDA12.3/Seatbelts ON/GLM ON/Release
Linux/CUDA12.3/Seatbelts OFF/GLM ON/Release
Windows/CUDA12.0/Seatbelts ON/GLM OFF/Debug

In release builds kernels are taking ~1 second to compile each. As Jitify is now doing the pre-processing, this is closer to 2.5 seconds under Debug builds.

[ptheywood]

The updated jitify2 branch uses c++ features which error on windows with CUDA 11.0 and 11.1, but 11.2+ is fine.

This was a planned change anyway for CCCL support (see #1155), so will just need dealing with here too/instead.

Another option (to discuss) would be to drop 11.x more widely (although I'd rather not yet) as the plan is to only support 2 major cuda versions and I expect 13 to turn up this year, but would rather not do that yet.

For wheel production on github actions, we do need to maintain vs 2019 support, with the recent change for vs2022 updates requireing 12.4, even if we drop cuda 11.0.
(when the vs2019 image dissapears we will have to find a workaround for that, or just tell windows users to use WSL). Mainstream window server 2019 support has ended, but extended support is until 2029. windows server 2026 was dropped by github actions ~2 months after mainstream supported ended, but that was before Microsoft purchased github.