Hi all,
Anyone has an idea if there will be some implementation of automatic tuning
(something similar do ATLAS) for GPUs?
Is there (or there will be) any similar mechanism in Magma?
thanks
Gianluca
BLAS autotuning
-
Stan Tomov
- Posts: 283
- Joined: Fri Aug 21, 2009 10:39 pm
Re: BLAS autotuning
Yes, our kernels are "auto-tuned" (not totally automatic yet). Some of the
our work on auto-tuning MAGMA BLAS kernels is described in
Li, Y., Dongarra, J., Tomov, S. " A note on auto-tuning GEMM for GPUs,"
Proc. of ICCS'09, Baton Rouge, LA, UT-CS-09-635, May 25-27, 2009.
Nath, R., Tomov, S., Dongarra, J. "Accelerating GPU Kernels for Dense Linear Algebra,"
Proc. of VECPAR'10, Berkeley, CA, June 22-25, 2010.
We released the sources for our new SGEMM and DGEMMs for Fermi. The algorithms had
to be extended so the search space for the "automatic" search for new algorithms was also
expanded (e.g., this is how the sgemm was derived from the dgemm implementation).
Note that these kernels can be used to speedup most of the other Lavel 3 BLAS by expressing
them in terms of GEMMs, or third party efforts to develop auto-tuned BLAS.
Stan
our work on auto-tuning MAGMA BLAS kernels is described in
Li, Y., Dongarra, J., Tomov, S. " A note on auto-tuning GEMM for GPUs,"
Proc. of ICCS'09, Baton Rouge, LA, UT-CS-09-635, May 25-27, 2009.
Nath, R., Tomov, S., Dongarra, J. "Accelerating GPU Kernels for Dense Linear Algebra,"
Proc. of VECPAR'10, Berkeley, CA, June 22-25, 2010.
We released the sources for our new SGEMM and DGEMMs for Fermi. The algorithms had
to be extended so the search space for the "automatic" search for new algorithms was also
expanded (e.g., this is how the sgemm was derived from the dgemm implementation).
Note that these kernels can be used to speedup most of the other Lavel 3 BLAS by expressing
them in terms of GEMMs, or third party efforts to develop auto-tuned BLAS.
Stan