[Octopus-devel] cvs commit: marques/octopus/src casida.F90 eigen_evolution.F90 linalg_basic.F90 linalg_basic_blas.F90 math.F90

[cvsusers at tddft.org]

acastro     2005/07/18 11:43:22 WEST

  Modified files:
    octopus/src          casida.F90 eigen_evolution.F90 
                         linalg_basic.F90 linalg_basic_blas.F90 
                         math.F90 
  Log:
  * I implemented the a different orthonormalization scheme for the
    "evolution in imaginary time" eigensolver method, as suggested by
    Aichinger and Krotscheck (ref in a comment). It permits faster
    convergence (not spectacular, but noticeable).
  
    Still poor convergence for curvilinear coordinates; however it converges
    with a little bit of patience, if you want to try. I already did
    calculations for molecules such as CO2. Did not try the time evolution,
    though (it crashes).
  
    I think we now probably are now in a state similar to the state that
    Gygi achieved, unless he had a better diagonalizer (I wrote him an
    email, no response yet). Maybe after implementing a multigrid algorithm
    we will be even better.
  
  * Polished the sorting algorithm. I changed the casida routine so that
    now it uses the sort in math.
  
  * Extended the interface to the BLAS X(gemm) routine, so that it can
    be called fro three-dimensional matrices (the first two dimensions are
    in fact only one inside the routine).
  
  Revision  Changes    Path
  1.15      +15 -24    marques/octopus/src/casida.F90
  1.5       +29 -9     marques/octopus/src/eigen_evolution.F90
  1.8       +9 -5      marques/octopus/src/linalg_basic.F90
  1.8       +14 -3     marques/octopus/src/linalg_basic_blas.F90
  1.40      +38 -3     marques/octopus/src/math.F90