$Id$

In some cases, the current implementation has been left with some
holes to fill in for generality.

General
o Approximations other than V, especially asymmetric, not well tested.
o useful test technique: jobs with number of virtuals less than number
  of occupieds.
o Are all integrals produced for gradients needed? (ap|D) integrals
  maybe doesn't need (ay|D), for example?

driver_xf:
o No dynamic block sizing
o xform_3cint should be modified to write out both 2-index and 3-index
  transformed integrals to speed the asymmetric split of the SVS and
  V approximations.
o Schwarz screening for other than V approximation?

driver_e:
o dynamic block sizing out of date
o either p2ij or p2ab will work only if entire amplitude tensor is in core
o NMO probably not needed in routine
o Don't pass in both NFit and FitBas

t_stats:
o t statistics should get only unique indices

driver_p & driver_g:
o l2 uses ga_matmul_patch and is therefore not scalable
o l2(ab) block is possibly transposed?  Does it matter?
o density contributions to W2 are backwards from derivation????
o p2am.F requires only Vx(C+A) block of L.  Irrelevant AxC block added
  in mk_l.F.  Probably can compute less of L1-4, but CHECK NEEDS FOR
  P2(iy) and W2 (all in mk_l.F)

nonsep2
o g_V should probably be distributed on shell blocks

Helpful debugging tips:
o Run jobs for which number of number of virtuals < occupied
o Run jobs for which number of frozen core > actives
o Run jobs for which number of frozen core > number of virtuals
  (i.e. Ne2 in 3-21G basis with 9 frozen core, 1 occupied, and 8
  virtuals) 
