Add new BLAS and LAPACK symbols in quadruple precision #145
Conversation
amontoison
commented
Nov 9, 2024
•
amontoison
commented
- Add BLAS and LAPACK symbols in quadruple precision
- Convention for symbols in quadruple precision Reference-LAPACK/lapack#1072
amontoison
force-pushed
the
quad_symbols
branch
2 times, most recently
from
ecd9f06 to
341e7e3
Compare
There was a problem hiding this comment.
This is going to need a version bump. Can you increase the version number here? I think increasing the patch number should be sufficient, as this is totally backwards-compatible.
ext/gensymbol/gensymbol
Outdated
| @@ -481,7 +495,7 @@ zpotri, | |||
| dtptrs, | |||
| dtrcon, dtrevc, dtrexc, dtrrfs, dtrsen, dtrsna, dtrsyl, | |||
| dtrtrs, dtzrzf, dstemr, | |||
| dsgesv, dsposv, dlag2s, slag2d, dlat2s, | |||
| dsgesv, dsposv, dlag2s, dlag2d, dlat2s, | |||
| @@ -566,6 +581,69 @@ zpotri, | |||
| zgeqrt, zgeqrt2, zgeqrt3, zgemqrt, | |||
| ztpqrt, ztpqrt2, ztpmqrt, ztprfb, | |||
| ); | |||
|
|
|||
| @exlapackobjs2 = ( | |||
There was a problem hiding this comment.
This is quite a lot of symbols being added; what happened here? Were we just missing these?
There was a problem hiding this comment.
The idea is to support quadruple-precision BLAS/LAPACK with LBT.
The symbols are not standard in quadruple precision; OpenBLAS uses the prefixes q* or x*, but only for the BLAS symbols and routines with real inputs.
I would like to add a similar paradigm for LAPACK symbols, but I am open to suggestions!
Integrating Fortran code compiled in quadruple precision is straightforward in Julia using the Float128 data type from Quadmath.jl.
However, we need an efficient way to compile our library with BLAS/LAPACK libraries that have dedicated symbols for quadruple precision.
LBT is optimal because we can add a dedicated backend compiled in quadruple precision at runtime.
There was a problem hiding this comment.
So you're planning to have LBT redirect back to Julia code for the backing algorithms for this quadruple precision?
There was a problem hiding this comment.
No, I was planning to compile Reference-LAPACK in quadruple precision.
A simple way to do this is to use the compiler-dependent flag -freal-8-real-16 with gfortran and some macros for the symbols.
The goal is to use it in libHSL, which is more or less similar to SuiteSparse and has some of the most robust linear solvers, such as MA27 and MA57, available in quadruple precision.
I can easily call them later in Julia with wrappers like this.
CUTEst doesn’t have any dependencies, unlike libHSL, which requires BLAS / LAPACK.
There was a problem hiding this comment.
I see. And so you’re trying to start a new convention with this naming scheme, or is there already a naming scheme for this and you’re conforming to that?
There was a problem hiding this comment.
For the q* symbols, I followed what was initially established by GotoBLAS, so I’m not creating a new convention here; instead, I’m following one that already exists informally. I proposed splitting this PR into two parts, one of which would include only the quadruple symbols.
I’ll contact Julien Langou, one of the developers of Reference-LAPACK, to confirm this informal convention, and then we can decide how to proceed.
amontoison
changed the title
amontoison
force-pushed
the
quad_symbols
branch
from
d5f56df to
6fb30b3
Compare
amontoison
force-pushed
the
quad_symbols
branch
from
6fb30b3 to
2607335
Compare
