Re: Matrix multiplication performance
Michael Lehn <[email protected]>
| Newsgroups | gmane.comp.lib.boost.ublas |
|---|---|
| Message-ID | <[email protected]> |
On 27 Jan 2016, at 17:15, Nasos Iliopoulos <[email protected]> wrote: > > > On 01/27/2016 10:52 AM, Michael Lehn wrote: >>> 5. ublas is one of the fastest libraries for small matrices. Because I think the most prevalent use of dense multiplication is on small matrices ( geometric transformations etc.) we need to make sure this stays that way. >> Depends on the domain of usage. In my domain dense matrix multiplication is the key to both, fast dense linear algebra (LAPACK etc.) and >> sparse direct or iterative solvers. But sure, we should have both. >> >> I will finish a prototype of BLAS Level 3 the next days but that might require some help on uBLAS. In particular I want to write a BLAS F77 and >> CBLAS interface so that one can build a BLAS library that can be linked against LAPACK. Because that’s the message: First BLAS was written >> in Fortran, then in C and they provided a Fortran interface. Today one writes a BLAS implementation in C++ and provides a C and Fortran interface. > I agree with that. We may want to provide a C and fortran interface eventually No, not eventually :-) I will do it as soon as I have figured out how to create a uBLAS matrix/vector from a raw C-array. Once I know that its a matter or hours. This also allows us to use the LAPACK test suite as *one* test for the BLAS implementation. Also we can use the ATLAS benchmark suite: It compares the performance of ATLAS with a F77 BLAS implementation. Personally I never trust any benchmarks that can not be reproduced this way.