meson build: fall back to dependancy if find_library fails

[ghost]

find_library only does a basic search for libs where dependency can use cmake and pkg-config. this is helpful on systems like nix where libraries are located over may uniquely named directories

NixOS/nixpkgs#268178

[bodono]

Thanks for doing this! It looks good to me, but I don't know anything about meson. Paging @enzbus (who is the expert here) to take a look.

[enzbus]

This mesonbuild/meson#6405 (comment) confirms OP statement. Probably all find_library can be switched. Does it pass the CI?

[enzbus]

It seems it fails on linux, we could try to do apt install libopenblas-dev since it doesn't pick the conda blas. Weird, looks like find_library is more robust.

[bodono]

It seems it fails on linux, we could try to do apt install libopenblas-dev since it doesn't pick the conda blas. Weird, looks like find_library is more robust.

Hmmmmm, that's strange. Not sure what the solution is here.

[enzbus]

find_library only does a basic search for libs where dependency can use cmake and pkg-config. this is helpful on systems like nix where libraries are located over may uniquely named directories

NixOS/nixpkgs#268178

@a-n-n-a-l-e-e do you have a clear usecase for this? it breaks our CI, you're right that what you're pointing is probably a more canonical way to do it but I'm not sure it's worth breaking what we have. This meson.build was adapted from numpy and scipy's.

[enzbus]

Also, you probably know this but this is only for building pip packages. scs proper has a separate build system, and repository.

[ghost]

It seems it fails on linux, we could try to do apt install libopenblas-dev since it doesn't pick the conda blas. Weird, looks like find_library is more robust.

does seem that way -- i would've expect it to fall back on find_library, seems like the openblas.pc. scipy seems to use
blas = dependency(['openblas', 'OpenBLAS']).

the use case i have to to work better on nix where the libraries are in unique directories.

[enzbus]

If the CI fails again (let's see) you could instead of replacing the find_library(), add an extra if else clause to also do dependency().

[ghost]

If the CI fails again (let's see) you could instead of replacing the find_library(), add an extra if else clause to also do dependency()

looking at the docs dependancy defaults to auto detect, which is pkg-config + cmake + OSX thing. but there is a way to get it to also use system, which i think is the same as find_library
https://mesonbuild.com/Dependencies.html#dependency-detection-method

let me see if i can get something that works. I appreciate you trying the CI.

[ghost]

If the CI fails again (let's see) you could instead of replacing the find_library(), add an extra if else clause to also do dependency().

yeah -- this seems like the best option. unfortunately, the method: arg for dependency only takes a string, not a list of options. will update soon -- but looks kind of gross. sorry.

[bodono]

All checks pass! Is this ready to merge?

[ghost]

All checks pass! Is this ready to merge?

yes

[ghost]

woot! thanks!

[ghost]

so -- it seems that openblas 0.3.22 renamed the pkg-config file from openblas.pc -> openblas64.pc. so when checking for the openblas module probably want to have dependancy(['openblas64', 'openblas', ...], ...). looking at the CI logs i can't tell what version of openblas it is building with, but if version 0.3.22 or later, could've been why my initial change was failing.

OpenMathLib/OpenBLAS#3791

[bodono]

I presume openblas64 is using 64 bit integers? If so when linking against it we need to pass the compilation flag BLAS64 = 1.

[ghost]

I presume openblas64 is using 64 bit integers? If so when linking against it we need to pass the compilation flag BLAS64 = 1.

thanks, i missed this. I don't think ILP64 is what we want. the nixos openblas build recently changed from LP64 to ILP64. investigating...

[ghost]

i was confused and thought the 64 indicated pointer width, then further confused with ILP64 that the 8 byte integers don't match the c. nixos didn't recently change from LP64 to ILP64, it is just that the meta data in the pkg-config file now indicates the 8 byte integer width (nixos provides both libraries, and i just happened to be working on fixing the build of a package that was erroneously using the ILP64 version).

for nixos -- the python-scs uses openblas LP64 (except for darwin, as that uses accelerate). that matches what numpy is using, unless there is a compelling reason to build for ILP64 matching numpy is simpler.

[enzbus]

Feel free to try with openblas64 via dependency, if it passes the CI we can drop the initial find_library clauses which would simplify the code a little.

[ghost]

I don't think there is any pkg-config or cmake configuration data in the CI -- at least from looking at the files in the openblas-devel fedora package. in contrast with debian, ubuntu and nix which provide the configuration files to pkg-config and cmake

lib/.../cmake/openblas/OpenBLASConfig.cmake
lib/.../cmake/openblas/OpenBLASConfigVersion.cmake
lib/.../pkgconfig/blas-openblas.pc
lib/.../pkgconfig/lapack-openblas.pc
lib/.../pkgconfig/openblas.pc

i was mistaken to suggest openblas64.pc or that is for ILP64. LP64 uses openblas.pc

i will keep my eye out for some way that the two functions can be combined. i feel like i made a bit of a mess out of that file.

[enzbus]

OK so we keep the current (master branch) version, let us know if things work on your side.