Building GINKGO on Windows - Any binaries ?

[shock-pro]

Hello Everyone,
I read the Installation section and it is not clear to me whether
Ginkco builds on Windows with CMAKE ?
If yes, what toolset can be used, VS ?, MinGW?, etc.
Any prebuilt binaries available ?
Thanks in advance!

[shock-pro]

Sorry, I just found https://github.com/ginkgo-project/ginkgo/releases/tag/v1.3.0
where it clearly states that GINKGO can be built on Windows.

Still, I would like to know if there are any prebuilt binaries available ?
Thanks

[upsj]

You can use MSVC cl, MinGW or Cygwin gcc. MSVC clang might also work, but we haven't tested that yet. The normal pattern

mkdir build
cd build
cmake ..
cmake . --build --config Release

should work, though you may want to add -j8 or something depending on the number of your cores to speed up the build.
While we do support Windows builds, we do not test them extensively. We only support 2 of our 5 backends (Reference and CUDA) with MSVC due to their outdated OpenMP support.
If you want a quick start for working with Ginkgo on Windows, I would suggest using vcpkg or conan, which already have version 1.3.0 available, and conan might even provide prebuilt binaries? Though I personally prefer vcpkg, since their CMake integration is much nicer.
The thing with prebuilt binaries is that they require us to go all-or-nothing - Ginkgo has a number of optional features, and if we were to release binaries, they would need to all be enabled. That includes builds for all CUDA architectures (some of which may be deprecated or unavailable in newer or older CUDA versions, respectively), ROCm (which is unavailable on Windows), OpenMP (again, currently unavailable) and DPCPP (which should be available, but we haven't tested it yet)

[shock-pro]

Hello Tobias, Thank you so much for your advice/recommendations. You are absolutely right on MSVC's inferior OpenMP implementation. Now that oneAPI is free, can we build on Windows with icl for all code containing OpenMP, and leave cl just for nvcc (I don't know if you use nvcc)? Thanks again!

…

On Mon, Jul 5, 2021 at 12:52 AM Tobias Ribizel ***@***.***> wrote: You can use MSVC cl, MinGW or Cygwin gcc. MSVC clang might also work, but we haven't tested that yet. The normal pattern mkdir buildcd build cmake .. cmake . --build --config Release should work, though you may want to add -j8 or something depending on the number of your cores to speed up the build. While we do support Windows builds, we do not test them extensively. We only support 2 of our 5 backends (Reference and CUDA) with MSVC due to their outdated OpenMP support. If you want a quick start for working with Ginkgo on Windows, I would suggest using vcpkg or conan, which already have version 1.3.0 available, and conan might even provide prebuilt binaries? Though I personally prefer vcpkg, since their CMake integration is much nicer. The thing with prebuilt binaries is that they require us to go all-or-nothing - Ginkgo has a number of optional features, and if we were to release binaries, they would need to all be enabled. That includes builds for all CUDA architectures (some of which may be deprecated or unavailable in newer or older CUDA versions, respectively), ROCm (which is unavailable on Windows), OpenMP (again, currently unavailable) and DPCPP (which should be available, but we haven't tested it yet) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#814 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ASNY4BOWOT2P7WC3VN4257DTWFCB7ANCNFSM47ZYSGOQ> .

[upsj]

I think that might be difficult, as CMake does not support different C++ compilers at the same time. The only solution I could imagine would be using cl as nvcc host compiler and icl as regular C++ compiler. That does sound very complex though and almost like a recipe for binary incompatibility, but it might be worth a try? You can use MinGW gcc, which gives you OpenMP support, or use MSVC and get CUDA support, but both doesn't seem possible at the moment. Luckily, the situation is better under Linux (I cannot recommend WSL2 with GPU passthrough enough!)

[shock-pro]

I am using icl for all cpp code except .cu files, as nvcc only supports cl, and the binaries are 100% compatible, that was always the idea behind icl. However, I write my own Makefiles, so I can pick what compiler to use, namely icl for all cpp, and nvcc for .cu, and nvcc automatically picks up cl because that's all that it knows. I would have thought that cmake could do the same when writing the VS projects, but based on your statement, it can't. Thanks again for your help. Cheers,

…

On Mon, Jul 5, 2021 at 3:56 PM Tobias Ribizel ***@***.***> wrote: I think that might be difficult, as CMake does not support different C++ compilers. The only solution I could imagine would be using cl as nvcc host compiler and icl as regular C++ compiler. That does sound very complex though and almost like a recipe for binary incompatibility, but it might be worth a try? You can use MinGW gcc, which gives you OpenMP support, or use MSVC and get CUDA support, but both doesn't seem possible at the moment. Luckily, the situation is better under Linux (I cannot recommend WSL2 with GPU passthrough enough!) — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#814 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ASNY4BPYJNJQYSSFYIUNJS3TWIMADANCNFSM47ZYSGOQ> .