Adapt amg4psblas to the communication_v2 PSBLAS interfaces, which removed
the work argument from all vector routines (psb_*_vect, the base prec
apply_vect, map_U2V_v/map_V2U_v).
Remove work from the whole vector apply chain:
- prec apply2_vect/apply1_vect and the amg_*precaply2/1_vect implementations
(dropped the local work_ buffer);
- mlprec_aply_vect and its inner recursive routines
(inner_ml_aply, inner_add/mult/k_cycle, inneritkcycle);
- the smoother/solver *_apply_vect implementations and their interface
declarations, dropping the now-dead aux/ww scratch buffers;
- AS smoother restr_v/prol_v (psb_halo/psb_ovrl on vectors);
- onelev map_rstr_v/map_prol_v (psb_map_U2V/V2U on vectors);
- poly_smoother_bld power-iteration apply_v calls.
Array routines keep work. mumps/slu/sludist/umf vector applies now pass a
zero-size local buffer to the underlying array apply (work optional there).
Regenerated for s/d/c/z; the library builds and the pdegen samples converge.
Stack-1
sfilippone