← Back to team overview

dolfin team mailing list archive

[HG DOLFIN] Run MTL4 sandbox through dolfin::Assembler interface.

 

One or more new changesets pushed to the primary dolfin repository.
A short summary of the last three changesets is included below.

changeset:   4453:68f3a504544907e3635ff021a40b8bdeab0fb380
tag:         tip
user:        "Garth N. Wells <gnw20@xxxxxxxxx>"
date:        Fri Jul 18 12:55:37 2008 +0100
files:       dolfin/la/MTL4Matrix.cpp dolfin/la/MTL4Matrix.h dolfin/la/MTL4Vector.cpp dolfin/la/MTL4Vector.h dolfin/la/dolfin_la.h dolfin/la/mtl4.h sandbox/mtl4/backend_test/Poisson.form sandbox/mtl4/backend_test/Poisson.h sandbox/mtl4/backend_test/Poisson.py sandbox/mtl4/backend_test/SConstruct sandbox/mtl4/backend_test/main.cpp sandbox/mtl4/conventional_backend/DefaultFactory.cpp sandbox/mtl4/conventional_backend/MTL4Factory.cpp sandbox/mtl4/conventional_backend/MTL4Factory.h sandbox/mtl4/conventional_backend/MTL4Matrix.cpp sandbox/mtl4/conventional_backend/MTL4Matrix.h sandbox/mtl4/conventional_backend/MTL4SparsityPattern.cpp sandbox/mtl4/conventional_backend/MTL4SparsityPattern.h sandbox/mtl4/conventional_backend/MTL4Vector.cpp sandbox/mtl4/conventional_backend/MTL4Vector.h sandbox/mtl4/conventional_backend/SConstruct sandbox/mtl4/conventional_backend/dolfin_la.h sandbox/mtl4/conventional_backend/mtl4.h sandbox/mtl4/mat_vec_mult_bench/main.cpp sandbox/mtl4/nse_bench/main.cpp sandbox/mtl4/poisson_bench/main.cpp
description:
Run MTL4 sandbox through dolfin::Assembler interface.

Results are still very good, espcially for Navier-Stokes. MTL4 relies on specifying the number of non-zeroes per row. Without this, it is slower than uBLAS.

For Navier-Stokes (matrix size 681k x 681k), MTL4 with number of non zeroes specified is 3.2 times faster than uBLAS, and 2 times faster for the re-assmeble.

Comparison is not entirely fair because eventually some form of sparsity pattern will need to be computed for MTL4 to provide number of non-zeroes per row, although knowing the sparsity pattern may possible speed insertion.

MTL4 is 1.25 faster for the matrix-vector multiplication test.


changeset:   4452:6b9f22aa71a8330bda1070fb2cb829fc0a20a53a
user:        "Garth N. Wells <gnw20@xxxxxxxxx>"
date:        Fri Jul 18 11:25:24 2008 +0100
files:       sandbox/mtl4/mat_vec_mult_bench/SConstruct sandbox/mtl4/nse_bench/SConstruct sandbox/mtl4/poisson_bench/SConstruct
description:
Remove MTL4_DIR variable in sandbox. pkg-config now returns necessary path.


changeset:   4451:159cc501482bb19e353eab6132e60b229d73034b
user:        "Garth N. Wells <gnw20@xxxxxxxxx>"
date:        Fri Jul 18 11:06:13 2008 +0100
files:       SConstruct dolfin/la/DefaultFactory.cpp dolfin/la/mtl4.h
description:
Undo change to SConstruct.

----------------------------------------------------------------------
For more details, visit http://www.fenics.org/hg/dolfin