Changeset [462474199d6a88082d4d590cfc535eb3cd31a472] by Jed Brown

Preliminary implementation of p-Laplacian solver

• Some fixes in gradient evaluation
• Boundary conditions not yet implemented
• Q1 Jacobian matches FD exactly for first-order elements

• Still need MatShell Jacobian to replace MFFD

• Profiling results valgrind --tool=callgrind --dump-instr=yes --simulate-cache=yes ./ellip -dmesh_in dblock6.h5m -const_BDeg 4 -exact 1 -snes_mf_operator -snes_monitor -ksp_converged_reason -ksp_constant_null_space -pc_type hypre -pc_hypre_type euclid -ksp_monitor

• Jp assembly is not cheap (23% of total runtime)

  • 10% in dFSMatSetValues (0.4% overhead)
  • 6% in pointwise evaluation (can
• 50% spent in TensorMultHex
  • less than 400 MFLOPS, need to unroll some loops

Signed-off-by: Jed Brown jed@59A2.org http://github.com/jedbrown/dohp/...