DGELSD faster than DGELSY?

[mengxr]

Hi,

I tried to solve a least squares problem of size 1e6-by-1e3. The machine I used has 12 Xeon cores. DGELSD took 152.8s to solve the problem. DGELSY took 223.8s. I tested ATLAS+LAPACK (3.3.0) and MATLAB's BLAS/LAPACK. DGELSD is faster than DGELSY in both cases. However, from the LAPACK Benchmark (http://www.netlib.org/lapack/lug/node71.html), DGELSY is almost as fast as DGELS and is significantly faster than DGELSD. Is it because DGELSD has better multi-threading support? or because of later improvement of DGELSD?

Thanks,
Xiangrui

[admin]

Yes that could be, the benchmark dates from before the multicore area.
But actually the multi-threading in within the BLAS, not LAPACK.
To test if the improvement comes from using a multithreaded BLAS, just link with the Reference BLAS (that is not multi-threaded)
Also some routines called beneath were updated like DLARFT that is called by DGELQF and DGEQRF.
See 9.4. High Performance QR and Householder Reflections on Low-Profile Matrices (see http://www.netlib.org/lapack/lapack-3.2.html)
Julie