Cannot build for CUDA 13

[ludvigak]

I'm trying to get my build script for cufinufft to work for CUDA 13, and I'm getting warnings about CUDA_SDK not having a mapping for the platform:

...
  [6cbf2f2e] + CUDA_SDK_jll v13.0.1+1
→ [8d8e3561] + CUDA_SDK_static_jll v13.0.1+1
  [e98f9f5b] + NVTX_jll v3.2.2+0
        Info packages marked with → not downloaded, use `instantiate` to download
...
  [6cbf2f2e] + CUDA_SDK_jll v13.0.1+1
→ [8d8e3561] + CUDA_SDK_static_jll v13.0.1+1
...
┌ Warning: Dependency CUDA_SDK_static_jll does not have a mapping for artifact CUDA_SDK_static for platform x86_64-linux-gnu-libgfortran5-cxx03-cuda+13.0
└ @ BinaryBuilderBase ~/.julia/packages/BinaryBuilderBase/SOM6i/src/Prefix.jl:855
┌ Warning: Dependency CUDA_SDK_jll does not have a mapping for artifact CUDA_SDK for platform x86_64-linux-gnu-libgfortran5-cxx03-cuda+13.0
└ @ BinaryBuilderBase ~/.julia/packages/BinaryBuilderBase/SOM6i/src/Prefix.jl:855

The build later fails because /workspace/destdir/cuda is missing.

I don't know what to do here, have I configured wrong or is there a bug in the build pipeline?

[ptiede]

I was playing around with this and #12021 made me realize that in CUDA 13 you need to update the platform so that the version is just the major 13 and doesn't include the minor version numbers.

full_platforms = Platform[]
for platform in platforms
    augmented_platform = deepcopy(platform)
    augmented_platform["cuda"] = "$(version.major)"
    push!(full_platforms, augmented_platform)
end

So if you update the build_tarballs to utilize this, I was able to get rid of the warning. However, when I tried to continue the build, I got a new error

cmake .. -DCMAKE_PREFIX_PATH="${prefix}" -DCMAKE_INSTALL_PREFIX="${prefix}" -DCMAKE_TOOLCHAIN_FILE="${CMAKE_TARGET_TOOLCHAIN}" -DCMAKE_BUILD_TYPE=Release -DFINUFFT_FFTW_SUFFIX="" -DFINUFFT_USE_CPU=OFF -DFINUFFT_USE_CUDA=ON -DCUDA_PATH=$prefix/cuda -DCUDAToolkit_ROOT=$prefix/cuda -DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHS}" -DFINUFFT_STATIC_LINKING="OFF"
Re-run cmake no build system arguments
-- LTO is supported and enabled.
-- Performing Test funrollloops
-- Performing Test funrollloops - Success
-- Performing Test ffpcontract_fast
-- Performing Test ffpcontract_fast - Success
-- Performing Test fnomatherrno
-- Performing Test fnomatherrno - Success
-- Performing Test fnosignedzeros
-- Performing Test fnosignedzeros - Success
-- Performing Test fnotrappingmath
-- Performing Test fnotrappingmath - Success
-- Performing Test fassociativemath
-- Performing Test fassociativemath - Success
-- Performing Test freciprocalmath
-- Performing Test freciprocalmath - Success
-- Performing Test fmergeallconstants
-- Performing Test fmergeallconstants - Success
-- Performing Test ftreevectorize
-- Performing Test ftreevectorize - Success
-- Performing Test fimplicitconstexpr
-- Performing Test fimplicitconstexpr - Failed
-- Flag -fimplicit-constexpr is not supported
-- Performing Test fcxlimitedrange
-- Performing Test fcxlimitedrange - Success
-- Performing Test O3
-- Performing Test O3 - Success
-- Performing Test /Ox
-- Performing Test /Ox - Failed
-- Flag /Ox is not supported
-- Performing Test /fp:contract
-- Performing Test /fp:contract - Failed
-- Flag /fp:contract is not supported
-- Performing Test /fp:except
-- Performing Test /fp:except - Failed
-- Flag /fp:except- is not supported
-- Performing Test /GF
-- Performing Test /GF - Failed
-- Flag /GF is not supported
-- Performing Test /GY
-- Performing Test /GY - Failed
-- Flag /GY is not supported
-- Performing Test /GS
-- Performing Test /GS - Failed
-- Flag /GS- is not supported
-- Performing Test /Ob
-- Performing Test /Ob - Failed
-- Flag /Ob is not supported
-- Performing Test /Oi
-- Performing Test /Oi - Failed
-- Flag /Oi is not supported
-- Performing Test /Ot
-- Performing Test /Ot - Failed
-- Flag /Ot is not supported
-- Performing Test /Oy
-- Performing Test /Oy - Failed
-- Flag /Oy is not supported
-- FINUFFT Release flags: -funroll-loops;-ffp-contract=fast;-fno-math-errno;-fno-signed-zeros;-fno-trapping-math;-fassociative-math;-freciprocal-math;-fmerge-all-constants;-ftree-vectorize;-fcx-limited-range;-O3
CMake is NOT being executed inside a GitHub workflow
-- ENV{GITHUB_ACTIONS}: 
-- Performing Test g
-- Performing Test g - Success
-- Performing Test g3
-- Performing Test g3 - Success
-- Performing Test ggdb
-- Performing Test ggdb - Success
-- Performing Test ggdb3
-- Performing Test ggdb3 - Success
-- Performing Test Wall
-- Performing Test Wall - Success
-- Performing Test Wextra
-- Performing Test Wextra - Success
-- Performing Test Wpedantic
-- Performing Test Wpedantic - Success
-- Performing Test Wnounknownpragmas
-- Performing Test Wnounknownpragmas - Success
-- FINUFFT Debug flags: -g;-g3;-ggdb;-ggdb3;-Wall;-Wextra;-Wpedantic;-Wno-unknown-pragmas
-- FINUFFT RelWithDebInfo flags: -funroll-loops;-ffp-contract=fast;-fno-math-errno;-fno-signed-zeros;-fno-trapping-math;-fassociative-math;-freciprocal-math;-fmerge-all-constants;-ftree-vectorize;-fcx-limited-range;-O3;-funroll-loops;-ffp-contract=fast;-fno-math-errno;-fno-signed-zeros;-fno-trapping-math;-fassociative-math;-freciprocal-math;-fmerge-all-constants;-ftree-vectorize;-fcx-limited-range;-O3;-g;-g3;-ggdb;-ggdb3;-Wall;-Wextra;-Wpedantic;-Wno-unknown-pragmas
CMake Error at /usr/share/cmake/Modules/CMakeDetermineCUDACompiler.cmake:212 (message):
  Couldn't find CUDA library root.
Call Stack (most recent call first):
  CMakeLists.txt:295 (enable_language)

I'm not sure why this is happening, and CUDA linking is not something I have a ton of experience with.

[ludvigak]

Huh, yes I can confirm that the CUDA_SDK gets included correctly when only specifying the major version. But this feels broken, since the build system is supposed to build for all minor versions.

Meanwhile, the CUDA 13 build (only specifying major version) now breaks because nvvm seems to be missing from the jll's.

[maleadt]

Meanwhile, the CUDA 13 build (only specifying major version) now breaks because nvvm seems to be missing from the jll's.

#12358

[ludvigak]

#12358

@maleadt when will this change make its way into a new jll? It seems a binary build has not been triggered yet.

[maleadt]

My mistake; opened a new PR in #12367.

[ludvigak]

Thanks! Closing this as building now works for CUDA 13. Minor version still cannot be specified, but that can be a later problem.