MAGMA 1.0

[Stan Tomov]

MAGMA 1.0 RC1 is now available. This release includes the MAGMA sources. MAGMA 1.0 RC1 is intended for a single CUDA enabled NVIDIA GPU. It extends version 0.2 by adding support for Fermi GPUs (see the sample performances for LU, QR, and Cholesky).

See the Software section for a download link.

[mtacconi]

It seems that some important (at least for me :) ) routines of the 0.2 release (xGEHRD for example) haven't been included in the 1.0RC1.
Do you plan to include again the xGEHRD in the final 1.0 release?
I am also looking forward to experiment with the symmetric eigensolver, I wonder if we have any chance to see the xSYTRD released in the final 1.0 as well.

[Stan Tomov]

We plan to release the reductions to Hessenberg (xGEHRD), bidiagonal (xGEBRD), and tridiagonal (xHETRD/xSYTRD) forms, as well as the eigen- and singular-value solver routines that use them (xGEEV, xGESVD, and xHEEVD/xSYEVD).
These and other routines will be added to MAGMA 1.0 over the weekend (RC2) and next week (RC3).

[davidyu]

Download link seems break? will it be fixed soon? Thank you.

[mtacconi]

Good news, really!
It is going to be a happy number-crunching holyday season :)

[Stan Tomov]

MAGMA RC2 is now available. It includes more functions (mostly related to QR) and also fixes a bug in RC1 related to sgemm for Fermi. The sample Fermi performances of the basic one-sided factorizations (LU, QR, and Cholesky) are updated to include the performance in complex arithmetic.

[Boxed Cylon]

Thanks for the new software toys - its also nice to have MAGMA working in matlab!

Some assorted comments, some of which may be sorted out in the formal release...or mentioned in the documentation.

It develops that calls to magma_sgesv_gpu seem to need matrices to be padded with zeros to sizes that are multiples of 32 - anything else and I seem to get the wrong answers. This is the fermi version of this code.

Although the MAGMA 1.0/RC2 tarball came with the source code for sgemv for fermi, the Makefile does not seem to want to compile it - at least it is not set up for it. I'm using the cublas version for now. Perhaps this is an oversight, perhaps sgemv is still a work in progress.

The efficiency of computation seems to depend critically on the host BLAS/LAPACK employed. The Intel MKL seem to run fastest, with AMD's ACML a close second. The native BLAS with my Suse 11.1 system and a downloaded+compiled ATLAS performed poorly, as did Sun's, I mean Oracle's, sunperf.a library. These latter libraries cause my test sgesv routine to run some 30-45% slower than when this routine is used with the Intel or AMD libraries. I find this sensitivity to be a little odd, inasmuch as the GPU is supposed to be employed, but no doubt the technical details/requirements of these routines escape me.

Prompted by a recent post, I just tried GotoBLAS2 which indeed is the fastest of all the BLAS libraries so far. http://www.tacc.utexas.edu/tacc-projects/gotoblas2/

Comparison of the performance of sgesv using various versions of BLAS

Magma_Sgesv_Performance.gif (18.17 KiB) Viewed 6624 times

Here is a comparison of what I get for the various BLAS versions for a call to SGESV, NRHS=5000. The CPU is an AMD Phenom II X4 965, and Tcpu is calculated using matlab R2010b, single processor. The GPU is GTX 480, "fermi". The 3.8.3 version of ATLAS was downloaded and compiled by me, following directions on the Atlas website. Suse 11.1 refers to the generic version of BLAS (probably atlas) available by rpm for that version of linux. MKL is version 10.2.5.035. ACML is version 4.4.0, ifort. GotoBLAS is version 1.13, compiled by me using gfortran. Sunperf is from version 12.2 of Oracle/Sun solstudio. (I gather intel i7's perform quite a bit faster in matlab than the AMD processors.)

In for a penny, in for a pound, here is the result of MAGMA's testing_sgetrf using GotoBLAS2 and a single processor (export OMP_NUM_THREADS=1) to give a benchmark for the AMD 965:

Code: Select all

./testing_sgetrf
device 0: GeForce GTX 480, 1401.0 MHz clock, 1535.7 MB memory
  M     N   CPU GFlop/s    GPU GFlop/s   ||PA-LU||/(||A||*N)
============================================================
 1024  1024   19.17          39.09         2.251748e-09
 2048  2048   20.77         115.83         1.963081e-09
 3072  3072   21.59         197.02         1.827424e-09
 4032  4032   22.05         254.75         2.083052e-09
 5184  5184   22.41         338.94         1.983713e-09
 6016  6016   22.64         389.84         1.922557e-09
 7040  7040   22.82         439.81         1.863838e-09
 8064  8064   22.97         478.04         1.977259e-09
 9088  9088   23.11         511.14         2.192004e-09


[jimmysu]

When is RC3 coming out? I am anxiously waiting for its release, in particular for the hybrid SVD implementation. I don't think there is a comparable hybrid SVD implementation out there yet. Please let me know if there is a way to get an early peek at the SVD implementation. Thanks.

[Stan Tomov]

MAGMA RC3 is now available. This release candidate adds the two-sided factorizations (Hessenberg, bidiagonal and tridiagonal reductions) and eigen/singular-value problem solvers. Added is also one-sided LAPACK factorizations optimized for just multicore architectures (no GPU). For more details on the functionality available see the MAGMA 1.0 presentation.

See the Software section for a download link.

[fletchjp]

Stan

Is there a date yet for RC4 and will there be some release notes or other documentation of what has been changed?

Thanks

John