20140203
When doing IV, the first stage will remove NAs with na.action
   indpendently from each other and the second stage. This may create
   havoc since the number of observations may differ. A workaround is
   to figure out which observations to use, and specify a subset to felm().

20110331
Look into writing OpenCL versions of the demeaning and Kaczmarz-code.  This could be a good
alternative to threading.  This is fairly simple I guess, just need a good autoconf-macro for figuring
out whether there's a working OpenCL installed.