Remove CUDA 11.0/11.1 support and upgrade to CCCL 2.2.0+

[ptheywood]

Drops CUDA 11.0 and 11.1 support, and replaces CUB/Thrust with CCCL for future support.

nvidia/CCCL is the new combined home of CUB, thrust and libcudacxx as of CUDA 12.2.

This switches to use CCCL for better support moving forwards, CUB/Thrust 2.x api and to get rid of a number of cub/thrust CMake workarounds. Also implicitly adds libcudacxx as a dependency via cub.

CCCL does not support 11.0, hence the need for removal.
CUDA 11.1 support is being removed to simplify the build matrix for python, due to 11.2+ abi stability.

Closes #1021

---

Todo

• Wait for CCCL v2.3.0 2.3.2/2.4.0, require 2.3.2/2.4.0, test CUDA 11.0/11.1/11.2 again.
  • This might also allow us to remove the workaround for CCCL v2.2.0's double find_package issue that has also been fixed in branch/2.3.x but not in 2.3.0/2.3.1 tagged commits.
• Tests: Linux, CUDA 11.0
  • passed, so our current usage doesn't hit the compiler bugs, though we do have a warning about a known compiler issue that we worked around internally
• Tests: Linux, CUDA 11.1
• Tests: Linux, CUDA 11.2
• Tests: Linux, CUDA 12.0
• Tests: Linux, CUDA >= 12.2
• Tests: Windows, CUDA 11.1/11.2
• Tests: Windows, CUDA 11.7 VS 2019 (what I had installed locally)
• Tests: Windows, CUDA 12.0
• Tests: Windows, CUDA 12.3, VS 2022

[ptheywood]

using CCCL with 11.0 is DOA:

For instance, CCCL requires a minimum supported version of 11.1 from the 11.x series due to an unavoidable compiler issue present in CTK 11.0.

To be honest, I don't strictly mind dropping 11.0 support (or 11.1, so we can just rely on 11.2 +'s stability promises).

I could make 11.0 work by using old cub/thrust from their respective locations, but eventually that will break. Will discuss this before I do anything else.

[Robadob]

so we can just rely on 11.2 +'s stability promises

This seems like the best plan, assuming you're not aware of any HPC with only super old CUDA available.

[ptheywood]

so we can just rely on 11.2 +'s stability promises

This seems like the best plan, assuming you're not aware of any HPC with only super old CUDA available.

Bessemer's central install is 11.0 or 11.1 iirc, but the driver is 12.2/3 compatible, so we can always grab the toolkit from conda (assuming cmake agrees...) (or open a ticket).
Stanage has at least one 12.x release, as does Bede. Unsure on JADE, but I'd be surpised/worried if the driver wasn't 12.x compatible.

Google colab is 11.2+ as well (the reason we were producing 11.0 wheels previously).

11.2 was released December 2020, 11.1 September 2020, and 11.0 June 2020.

---

Turns out 11.0 on linux compiles fine too, just windows it doesn't, so even if we don't "support" it, it currently works.

so it could be "11.2+ is supported, 11.0 & 11.1 may work under linux, but are not supported"

[Robadob]

Your choice tbh, you manage it. 11.2+ only feels simpler.

[ptheywood]

During meeting concluded that dropping CUDA 11.0 support in favour of future CCCL support is worthwhile / CUDA 11.2 is old enough to be a minimum.

I'll adjust CI to reflect this and test windows on 11.1+ at some point.

We don't need to drop 11.1, but probably simpler to just say 11.2+ as then our python wheels will be consistent with what we support.

[ptheywood]

Looks like the previous windows cuda 11.0 CI errors were actually a github actions windows-2019 vs windows-2022 difference.

i.e. visual studio 2019 issue?

2023-12-01T16:10:13.7088440Z ##[error]     1>D:\a\FLAMEGPU2\FLAMEGPU2\build\_deps\cccl-src\libcudacxx\include\cuda\std\detail\libcxx\include\__type_traits/is_constant_evaluated.h(27): error : identifier "__builtin_is_constant_evaluated" is undefined [D:\a\FLAMEGPU2\FLAMEGPU2\build\FLAMEGPU\flamegpu.vcxproj]

[Robadob]

i.e. visual studio 2019 issue?

I have that on my home desktop, can try fighting it at some point. Or we just sack off VS2019, given it doesnt built current jitify2-preprocess branch either.

[ptheywood]

It's still supported by CUDA, which libcucxx/CCCL claim to support the same platforms as so afiak it should work / be supported.
I had a quick look into the error message but didn't find any one with the particular error.

I think I've got vs 2019 installed too, just need to spend the time in windows at some point.

---

Looking at the full draft-release CI log, there is a windows-2019 CUDA 11.8 job which did pass, so its an older CUDA + vs2019 thing when doing things with libcu++'s type_traits (via cub should be our only inclusion of libcu++ headers currently).

Might be worth trying cccl on its own to see if we can repro it on vs2019 as well when testing locally (if the main issue can be repro'd locally, otherwise it'll be CI fun).

[ptheywood]

I do have visual studio 2019 installed, but only cuda 11.7 and 12.0 on windows currently.

Can select the cuda toolkit via -T from the command line, i.e. using 11.7 with vs 2019:

$ cmake .. -A x64 -G "Visual Studio 16 2019" -T cuda=11.7 -DCMAKE_CUDA_ARCHITECTURES=86 -DFLAMEGPU_BUILD_TESTS=ON
-- Selecting Windows SDK version 10.0.19041.0 to target Windows 10.0.22621.
... 
-- The CXX compiler identification is MSVC 19.29.30139.0
... 
-- Looking for a CUDA compiler - C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.7/bin/nvcc.exe
...

And building with:

$ cmake --build . --target flamegpu tests -j 8 --config Release

This correctly configured, and built the current state of the CCCL branch (though an env var may be needed for RTC if the cuda on path older than the one selected)

Installing 11.2 to try and repro the issue, but unlikely to actualyl dig into any failures this evening if it does repro the error.

[ptheywood]

11.2 + visual studio 2019 has reproduced the error locally.

A quick attempt to configure cccl standalone failed as I'm missing a test suite dependency.

[ptheywood]

Reproduced the error using the CCCL example locally with visual studio 2019.

git clone [email protected]:nvidia/CCCL
cd CCCL/examples/example_project
mkdir build
cd build
cmake 
cmake .. -A x64 -G "Visual Studio 16 2019" -T cuda=11.2 -DCMAKE_CUDA_ARCHITECTURES=86
cmake --build .

$ cmake --build .
Microsoft (R) Build Engine version 16.11.2+f32259642 for .NET Framework
Copyright (C) Microsoft Corporation. All rights reserved.

  Compiling CUDA source file ..\example.cu...

  C:\Users\ptheywood\code\cccl\examples\example_project\build>"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2
  \bin\nvcc.exe" -gencode=arch=compute_86,code=\"compute_86,compute_86\" -gencode=arch=compute_86,code=\"sm_86,compute_
  86\" --use-local-env -ccbin "C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\
  bin\HostX64\x64" -x cu   -I"C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\thrust\thrust\
  cmake\..\.." -I"C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\libcudacxx\lib\cmake\libcu
  dacxx\..\..\..\include" -I"C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\cub\cub\cmake\.
  .\.." -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\include"     --keep-dir x64\Debug  -maxrregcount=0
   --machine 64 --compile -cudart static -Xcompiler="/EHsc -Zi -Ob0" -g  -D_WINDOWS -DTHRUST_HOST_SYSTEM=THRUST_HOST_SY
  STEM_CPP -DTHRUST_DEVICE_SYSTEM=THRUST_DEVICE_SYSTEM_CUDA -D"CMAKE_INTDIR=\"Debug\"" -D_MBCS -D"CMAKE_INTDIR=\"Debug\
  "" -Xcompiler "/EHsc /W3 /nologo /Od /Fdexample_project.dir\Debug\vc142.pdb /FS /Zi /RTC1 /MDd /GR" -o example_projec
  t.dir\Debug\example.obj "C:\Users\ptheywood\code\cccl\examples\example_project\example.cu"
C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\libcudacxx\include\cuda\std\detail\libcxx\in
clude\__type_traits/is_constant_evaluated.h(31): error : identifier "__builtin_is_constant_evaluated" is undefined [C:\
Users\ptheywood\code\cccl\examples\example_project\build\example_project.vcxproj]

C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\libcudacxx\include\cuda\std\detail\libcxx\in
clude\__type_traits/is_constant_evaluated.h(36): error : identifier "__builtin_is_constant_evaluated" is undefined [C:\
Users\ptheywood\code\cccl\examples\example_project\build\example_project.vcxproj]

  2 errors detected in the compilation of "C:/Users/ptheywood/code/cccl/examples/example_project/example.cu".
  example.cu
C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\MSBuild\Microsoft\VC\v160\BuildCustomizations\CUDA 11.2.t
argets(785,9): error MSB3721: The command ""C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.2\bin\nvcc.exe" -gen
code=arch=compute_86,code=\"compute_86,compute_86\" -gencode=arch=compute_86,code=\"sm_86,compute_86\" --use-local-env
-ccbin "C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\HostX64\x64" -x cu
  -I"C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\thrust\thrust\cmake\..\.." -I"C:\Users\
ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\libcudacxx\lib\cmake\libcudacxx\..\..\..\include" -I"
C:\Users\ptheywood\code\cccl\examples\example_project\build\_deps\cccl-src\cub\cub\cmake\..\.." -I"C:\Program Files\NVI
DIA GPU Computing Toolkit\CUDA\v11.2\include"     --keep-dir x64\Debug  -maxrregcount=0  --machine 64 --compile -cudart
 static -Xcompiler="/EHsc -Zi -Ob0" -g  -D_WINDOWS -DTHRUST_HOST_SYSTEM=THRUST_HOST_SYSTEM_CPP -DTHRUST_DEVICE_SYSTEM=T
HRUST_DEVICE_SYSTEM_CUDA -D"CMAKE_INTDIR=\"Debug\"" -D_MBCS -D"CMAKE_INTDIR=\"Debug\"" -Xcompiler "/EHsc /W3 /nologo /O
d /Fdexample_project.dir\Debug\vc142.pdb /FS /Zi /RTC1 /MDd /GR" -o example_project.dir\Debug\example.obj "C:\Users\pth
eywood\code\cccl\examples\example_project\example.cu"" exited with code 1. [C:\Users\ptheywood\code\cccl\examples\examp
le_project\build\example_project.vcxproj]

I'll set up a quick reproducer CI repo to pin down the affected CUDA versions and report it upstream.

[ptheywood]

CUDA 12.3 vs 2022 tests pass.

CI sweep pinning down vs2019 + CUDA 11.x versions which exhibit the libcu++ compilation error, will report upstream tomorrow once versions are known. 11.7 works, 11.2 doesn't.

CCCL/libcu++ includes some msvc conditions for using the offending symbol for an older vs 2019 version(s) + other combos, so it could be that a more recent vs 2019 sub-update (locally have 1929, 1924 was the prev version with a workaround) so that might be relevant.

https://github.com/ptheywood/cccl-is-constant-evaluated-mwe/actions/runs/7105395494/job/19342497821

[ptheywood]

CUDA 11.3+ is fine with visual studio 2019, so its just 11.2 (and 11.1) which breaks for us. This would prevent us from producing 11.2 wheels.

We can't drop 11.2 yet, as its the version installed on google colab iirc (and I'd rather not either).

I've reported this upstream: NVIDIA/cccl#1179

[ptheywood]

Upstream has a PR in to fix this.

The simplest way to incorporate this into the cmake logic would be to just make our minimum the next release post merge, but I'm not sure when that would be.
If we want it sooner we could pin agaisnt the post merge hash pre tag, and just adjust our min version to what it should be (might be

Additionally, if we depend on something newer than 2.2.0 CCCL treats itself as system headers, even when not using isystem which will be good for our relatively strong warning levels.

[ptheywood]

CCCL 2.3.0 has been released on github: https://github.com/NVIDIA/cccl/releases/tag/v2.3.0

This should include the fixes we require, so making our min CCCL 2.3.0 and fetching newer if not found should be ok, but worth checking that both required fixes made it into this release.

[ptheywood]

The v2.3.0 tagged commit does not include the cmake fix or msvc fixes, although they were backported to the branch/v2.3.x branch.

There's a v2.3.1 tagged commit which also does not include these fixes, so presumably we need to wait for 2.3.2 or 2.4.0

We'll probably just need to keep the first find_package(COMPONENTS) workaround in place anyway, as CUDA 12.3 includes CCCL 2.2 for instance just to avoid any problems.

[ptheywood]

CUDA 12.4 has been released, which includes CCCL 2.3.1 according to the release notes.
Checking /usr/local/cuda-12.4/lib64/cmake/cccl/cccl-config.cmake this still does not contain the cmake fix we require (i.e. it lines up with the v2.3.1 tag on github) so we still need to wait for 2.3.2 or 2.4 before as our minimum cccl that will work on windows and not cause errors on re-finding in CMake.

    if (TARGET Thrust::Thrust AND NOT CCCL::Thrust)

[ptheywood]

CCCL v2.3.2 has just been tagged / released, which does include the 2 fixes we need, so it should now be possible to switch to this / this shouldn't be blocked any more.

https://github.com/NVIDIA/cccl/releases/tag/v2.3.2

[ptheywood]

CCCL 2.5.0 has been released on github, mostly fixes but also some potentially interesting additions (but not yet safe to use).

Shouldn't be a need to bump our minimum/fetched version to this in the PR though, 2.3.2 should be fine still (unless i've misssed something)

---

This PR is more or less good to go, just wanting to re-run windows testing with it requiring 2.3.2 just in case (Though I believe it would be fine). And a rebase would prolly be worthwhile.

Merging this prior to a non pre-release would be best, due to dropping CUDA 11.0/11.1 support.