SU(3) linear algebra primitives in the MILC code (Carleton)

see:   http://cliodhna.cop.uop.edu/~hetrick/milc/su3.html

The most critical are typically coded in assembly language.
These are the ones critical for Dslash:

Staggered Dslash:

   mult_adj_su3_mat_vec_4dir (does four su3 matrix x vector operations)
   mult_su3_mat_vec_sum_4dir
   sub_four_su3_vecs

Wilson Dslash:

   wp_shrink
   wp_grow
   mult_su3_mat_hwvec
   wp_grow_add

Also needed for conjugate gradient are

Staggered congrad:

  magsq_su3vec
  g_doublesum
  su3_rdot

and some specialized routines for staggered fermions:

  scalar_mult_latvec (performs scalar_mult_su3_vector with the same
   constant over the entire lattice)
  scalar_mult_add_latvec (same, but scalar_mult_add_su3_vector)
  copy_latvec (lattice global copy)
  clear_latvec (lattice global clear)

Wilson congrad:

  scalar_mult_add_wvec
  magsq_wvec
  g_doublesum
  copy_wvec
  wvec_dot
  c_scalar_mult_add_wvec
  g_dcomplexsum


For the more complicated improved actions with dynamical fermions, we
also have to worry about optimizing the calculation of the gauge force
term, since the time spent there is comparable to the time spent in
the fermion inversions.  Here are the primitives involved there:

  mult_su3_nn (by extension - other su3 matrix matrix multiplications)
  scalar_mult_add_su3_matrix
  mult_adj_su3_mat_hwvec
  mult_su3_mat_hwvec
  make_anti_hermitian
  uncompress_anti_hermetian
  su3_projector
  mult_adj_su3_mat_vec

Even with assembly-coded versions of complete inverters for standard
actions, need an optimized set of primitives, to develop new actions.