2008-06-26 00:30:21 BLAS/LAPACK
Andrew Willis (UNITED STATES)
Ported the BLAS / CLAPACK libraries (really just needed to change the makefiles) to blackfin linux.
These libraries provide linear algebra functions, BLAS concentrates on simple scalar/vector vector/vector matrix/vector and matrix/matrix arithmetic operations. CLAPACK concentrates on more sophisticated linear algebra functions, e.g., eigenvalues/eigenvectors solving linear systems and singular value decomposition.
If there's interest in the ported libraries, I'm happy to provide them.
2008-06-26 00:56:48 Re: BLAS/LAPACK
Mike Frysinger (UNITED STATES)
you can just post the patches as attachments so people who search in the future can find them
2008-06-26 09:30:50 Re: BLAS/LAPACK
Robin Getz (UNITED STATES)
I'm curious - why use CLAPACK? that is for systems which normally have no fortran compiler? (bfin-xxx-gfortran exists, and seems to work for all the built in tests, and other applications I have tried it on).
2008-06-26 20:30:05 Re: BLAS/LAPACK
Andrew Willis (UNITED STATES)
Short story: We're writing client C code.
Long story: I'm somewhat dubious regarding the depth of compiler optimizations and the overall future of FORTRAN which, for the time being, loads of mathematicians continue to use for numerical accuracy issues. While the gfortran compiler will likely do great for BLAS, i.e., outperform CBLAS, I can't be confident that will be the case for LAPACK and I'd rather no t mix-and-match languages for what I am using as a common codebase, i.e., having BLAS in FORTRAN and CLAPACK (in C of course).
CLAPACK is a solution to what I see as a changing situation, I'm hoping that nextgen mathematics algorithms will be developed in either a C++ or Java (bytecode run-time environment for the blackfin?). Also of notable consideration is the ScaLAPACK (for sparse matrices) and LAPACK++. While I can't predict the future, I'm hedging my development bets as you see above.
If one cares, you can get a quick snapshot of linear algebra source development trends through the work of Roldan Pozo from NIST (National Institute of Standards and Technology).
We're doing 3d reconstruction from a pair of cameras that connect to a BF537-STAMP through the FPGA-EZKIT. An abstract on the project is here. Some wiki stuff we've been compiling is here as we are just becoming accustomed to using the blackfin-uclinux distribution.
Sparse matrix inversion, singular value decomposition, and other linear algebra algorithms are needed to get good results from 3d stereo these days.
2008-07-25 03:16:06 Re: BLAS/LAPACK
Pei Chen (CHINA)
Hi, So far as I know, most of LAPACK's algorithms require floating point operation and dynamic memory allocate which is not so fast on blackfin. I believe some of them could be implemented by fractional point operation after normaliztion, but it is a tough job and require hundreds of hours' hard working. Maybe LAPACK on SHARC is an more feasible job.
If there is an open project for blackfin's LAPACK, I will try to donate something . I am working on linuxrobot right now, we need
matrix math algorithms for the invert dynamic control of walking. Maybe sharc processor could be my first choice, and blackfin's LAPACK could be the second.
2008-07-25 04:45:08 Re: BLAS/LAPACK
Yi Li (CHINA)
As Mike says, if you have any patch for linux on blackfin, you can just post it here or add to the patch tracker.