TODO

Features
========

Twisted mass operator.
Multi-GPU operation.
Multigrid solver.
More Krylov solver types.

Structure
=========

Eliminate in_parity in favor of matpc_type?

The use of QUDA_REFERENCE_CREATE to create ColorSpinorFields that
refer to others is far from perfect.  In particular, a
ColorSpinorField should know when it is a reference, and this should
be a distinct case from when a ColorSpinorField is not initialized.

Optimizations
=============

Optimized domain wall fermions (gauge field load sharing across Ls).
Dslash auto-tuning (block sizes, padding).
Enable the BLAS autotuning through the interface.