Force CHOLMOD to use an optimiced BLAS library

[Víctor de Buen Remiro]

CHOLMOD efficiency decays for dense and block-dense matrices, and it's almost sure due to it is using a non optimiced BLAS library.

It's neccessary to force CHOLMOD to use ATLAS or other optimiced libraries.

[Víctor de Buen Remiro]

A new analysis of Cholesky performance for semi-sparse matrices has been codified in TOL file
​trunk/tol_tests/tol/LinearAlgebra/vmatrix/test_0014/performance.tol

Results are stored in ​trunk/tol_tests/tol/LinearAlgebra/vmatrix/test_0014/performance.ods

Main conclusion is that efficience decays specially when CholeskiFactor of X'X is calculated directly over a non square matrix X that have a density of non null zeros not very low. Then it's a better option to calculate explicitly X'X and then to apply CholeskiFactor. Obtained simulation results show that density frontier is above 13%.

Using BLAS optimized libraries can help in generic terms but it doesn't seem very probable that could solve the problem by itself.

[Víctor de Buen Remiro]

In windows is posible to force CHOLMOD to use ATLAS libraries instead of non optimized ones (libblas32.dll and liblapack32.dll), by mean of this environment variables

set BLAPI_BLAS_LIB=libcblas.dll
set BLAPI_LAPACK_LIB=libclapack.dll
set BLAPI_VERBOSE=1

However, it causes some problems in release mode that cannot be reproduce in debug mode compiling trunk/tol_tests/tol/LinearAlgebra/vmatrix/test_0012/test.tol