Not possible to link Blis and Lapack statically into the same executable
Opened this issue · 12 comments
I would like to link Blis and Lapack statically into the same executable. However, this is not possible since several symbols are multiply defined between Blis and Lapack. These are lsame
, xerbla
, xerbla_array
, crot
, czymv
, csyr
, zrot
, zsymv
, zsyr
.
I noticed an earlier bug report #776 in which some of these were discussed and then included into Blis. Would it be possible to have an option to disable this in the future?
The background is that I am integrating them into a Rust project. Rust prefers static linkage and does not handle rpath well for shared libraries.
We also just handled a similar request via #812. Your request, @tbetcke, has me wondering about a general-purpose solution that combines the compatibility accommodations for both 812 and your request into a single option, say, --disable-symbols=<comma-separated-list-of-symbols>
. Then the user could omit whichever symbols they wish. (Of course, we'd have to guard the eligible symbols with cpp macros, but that's not difficult and it's something we can do incrementally as these offending symbols are brought to our attention.)
What do you think, @devinamatthews?
Yes, that's an interesting idea but it would be a lot of work to make it work "seamlessly" for end-users (e.g. they can disable any symbols within certain classes), and would be surprising to users if the list of symbols which can be disabled is small, irregularly defined, and/or constantly changing.
As an author of applications that use BLIS, LAPACK, and ScaLAPACK together, I'm generally in favor of improving interoperability. Beyond resolving multiply defined symbols, what I'd really like is for BLIS implementations to override those in Sca/LAPACK. A partial solution would be for BLIS to name-mangle select routines with __wrap_
, so I can use --wrap
to prefer BLIS implementations. I believe that this solution is only partial because I don't see how it could hijack the libcalls inside the Sca/LAPACK binaries; achieving that I think would require build-system modifications in both BLIS and Sca/LAPACK. I don't know the best way forward to request such cross-project functionality.
I think the point Nick is making: the implementation in a BLAS library is likely to be higher performing than the reference implementation in LAPACK or ScaLAPACK (or some other library). Hence, it is such a library that should selectively remove rather than, say, BLIS.
On the other hand, XERBLAS may be a replacement for the one in the BLAS specifically to do more meaningful error handling. In that case, it is the one in LAPACK/ScaLAPACK that should overwrite.
This is, I guess, an argument for vertical integration of such libraries... A project we intend to pursue next.
For the computational routines, one can access the BLIS version using the BLIS object or typed APIs. For stuff like lsame
and xerbla
there is no really good solution (esp. for xerbla
since that is supposed to be overridden by user applications.
Is it possible to mitigate the problem with __attribute__((weak))
?
@tbetcke does linking with -Wl,-z,muldefs
work?
@tbetcke one thing that we are concerned about in particular is xerbla
. Is a custom version of xerbla
(not the one in BLIS or LAPACK) also being included? The "default" xerbla
may exit on error which I assume is not what you want in a Rust crate.
I am told the most recent version of LAPACK marks xerbla
as a weak symbol so we could do that too, also for lsame
and xerbla_array
. For the computational routines (crot
etc.), we'll make sure these are covered by --enable-scalapack-compat
.
Or, we could add an --enable-lapack-compat
for the other symbols.
Apologies the long absence from this discussion. @devinamatthews would something like Openblas work in principle, where one can build Lapack as part of OpenBLAS? This would solve the multiple libraries issue for static archives as well.
I'm currently implementing the following configure
option:
--omit-symbols=LIST
Omit a custom set of compatibility symbols when building
BLIS. When given, LIST is parsed as a comma-separated
list of symbol names (excluding any trailing underscore).
This option is useful when (1) the user is planning to
link BLIS with another library that provides conflicting
symbols, and (2) the user wishes the symbols in this
other library to prevail at link time without relying on
weak/strong symbol semantics. Note that currently only
the following symbols are supported for omission:
crot zrot lsame
csymv zsymv xerbla
csyr zsyr xerbla_array
csyr2 zsyr2
Did I miss any of the symbols germane to this or related Issues?