[Faiss] Fix dependencies

[bobuc-havok]

Describe the pull request

• What does your PR fix?

  Fixes [faiss] The target files on x64-linux contain absolute paths #18776

• Which triplets are supported/not supported? Have you updated the CI baseline?

  all previously supported, No

• Does your PR follow the maintainer guide?

  Yes

• If you have added/updated a port: Have you run ./vcpkg x-add-version --all and committed the result?

  Yes

If you are still working on the PR, open it as a Draft: https://github.blog/2019-02-14-introducing-draft-pull-requests/

[NancyLi1013]

faiss failed on x64-windows with the following errors:

/pdb:faiss\faiss.pdb /dll /version:0.0 /machine:x64 /nologo /debug /INCREMENTAL /DEF:faiss\CMakeFiles\faiss.dir\.\exports.def lapack.lib D:\installed\x64-windows\debug\lib\openblas.lib kernel32.lib user32.lib gdi32.lib winspool.lib shell32.lib ole32.lib oleaut32.lib uuid.lib comdlg32.lib advapi32.lib /MANIFEST /MANIFESTFILE:faiss\CMakeFiles\faiss.dir/intermediate.manifest faiss\CMakeFiles\faiss.dir/manifest.res" failed (exit code 1104) with the following output:
LINK : fatal error LNK1104: cannot open file 'lapack.lib'
ninja: build stopped: subcommand failed.

[bobuc-havok]

@NancyLi1013 yup, saw that build break, it went green now :)

[NancyLi1013]

LGTM now, have you tested the feature gpu?

[bobuc-havok]

@NancyLi1013 good point, I added a small fix to the port file for gpu. That said faiss[gpu] does not build for me locally, regardless of any changes. I'm getting nvcc compiler errors such as:

faiss/gpu/utils/Tensor.cuh(366): error: argument list for template "T::canUseIndexType" is missing
faiss/gpu/utils/PtxUtils.cuh(25): error: asm operand type size(4) does not match type/size implied by constraint 'l'

There are others. In terms of CMake targets, the gpu feature should be fine, as the dependency is declared as:

find_package(CUDAToolkit REQUIRED)
target_link_libraries(faiss PRIVATE CUDA::cudart CUDA::cublas)

so it should not generate absolute paths. I'm not sure if the CI builds test faiss[gpu], but in case the failure to compile the cuda code is more general than on my machine, I think it should be split out from this PR / issue.

[NancyLi1013]

gpu is not default feature, so it will not be tested on CI. Thanks for your fix and information. Now can you build faiss[gpu]:x64-windows on your local machine?

[bobuc-havok]

@NancyLi1013 no I cannot build faiss[gpu]:x64-windows locally, I'm getting the nvcc compiler errors mentioned above. It might be because I don't have an NVidia graphics card on my machine, not sure.

[NancyLi1013]

I was blocked by the failure on my local with master branch.

CMake Error at CMakeLists.txt:26 (enable_language):
  No CMAKE_CUDA_COMPILER could be found.

Have you tried with master branch on your machine? Does it work? If you have the same problem with the changes on this PR, it might not be the new issue introduced by this PR.

[bobuc-havok]

@NancyLi1013 yup, I am getting the same cuda compiler not found error on master. Seems to be common, as other ports have fixes for it. It looks like you can't get to compiling cuda on master until you make changes and fix the cuda compiler like I have in the portfile in this PR. I don't think my changes broke the faiss[cuda] - I think it was simply broken from the start and not tested on CI.

[NancyLi1013]

Thanks for your efforts and patience. @bucurb

[vicroms]

Thanks for the PR!