Memory requirements of PLASMA dense matrix solvers

[uhle89]

Hi to all,
before a I started to use PLASMA I called LAPACK_sgesv() (LAPACK_dgesv())
to decompose dense matrices and to solve problems for various rhs.
These routines work in place and hence one can invert 100k (70k) eq within 48GB.
If there is enough memory to store the matrix twice, I loved to call sgesvx (dgesvx)
because they refine the solution, calculate condition, forward and backward errors.

My experiments with PLASMA show that LU decomposition can be done in place and
in parallel by PLASMA_sgetrf() (PLASMA_dgetrf).
(I failed to decompose in place by PLASMA_?gesv, PLASMA_?gesv_incpiv, PLASMA_?gelqf...)

The corresponding PLASMA solvers PLASMA_?getrs() however allocate the matrix memory a second
time, which forces me to choose a serial in place solver after decomposition if I run out of memory.

This doesnt hurt too much, since the effort for solving LUx=Py is only O(N^2). I was just wondering if there
is a better choice of a PLASMA_solver() which can do it in place.
regards Stephan

[mateo70]

Stephan,

Normally by adding the following line after the PLASMA_Init call, it should solve your problem:

Code: Select all

PLASMA_Set(PLASMA_TRANSLATION_MODE, PLASMA_INPLACE)

It will specify to do the translation from LAPACK layout to Tile layout inplace (with a low memory overhead) instead of using the out-of-place translation which doubles the memory requirements.
Or if you can, use directly the Tile interface of PLASMA to avoid conversion from one layout to another and so win a lot of space and time.

Mathieu

[uhle89]

Good to know, thank you.