Page 1 of 1


Posted: Tue Dec 05, 2017 6:58 am
by mrader1248

I stumbled over the paper "A HIGH PERFORMANCE QDWH-SVD SOLVER USING HARDWARE ACCELERATORS" by D. Sukkari et al. [1], where it is shown that SVDs can be performed much faster than it is done within MAGMA by first performing a polar decomposition (via QR decompositions) and then a hermitian EVD. To me it seems relatively easy to implement this using the routines already present in MAGMA. Therefore I want to ask if there are plans to implement this within MAGMA (in the near future)?

Thank you already in advance for your answer!

[1] ... mitted.pdf


Posted: Wed Dec 06, 2017 12:37 am
by mgates3
We don't have any immediate plans to include it, but will investigate its potential.

We are working on including an SVD solver with a 2-stage reduction to bidiagonal, which is considerably faster than the SVD with a classical 1-stage reduction to bidiagonal. Here is our recent paper:
Accelerating the SVD two stage bidiagonal reduction and divide and conquer using GPUs



Posted: Wed Dec 06, 2017 4:42 am
by mrader1248
Thank you for your reply!

This paper is not included in the MAGMA publications site?

Is there also a preprint available somewhere? (Unfortunately, my university has no subscription for this journal.)

The highlights section of your publication says:
Achieves over three times speedup compared to CPU-only two stage SVD, and 2.6 times speedup compared to GPU-accelerated one stage SVD.
For which case do you compare the speedup? How large is the matrix and is it a well-/ill-conditioned matrix?