On Mon, Apr 25, 2011 at 21:09, Roy Stogner <roystgnr@ices.utexas.edu> wrote:
Hmm... I added long double support to libMesh because those extra
couple digits can be a useful last resort for verifying when your
convergence is bottoming out due to floating point error.

Yeah, it's handy to check what's going wrong with MFFD Jacobians or if you just have huge jumps in coefficients (e.g. subduction or structural mechanics). Penalty boundary conditions make it worse, of course.

I don't see anything useful about 34 digit precision, but it's just
too cool to *not* support.

I don't know where 34 came from. This is 16-byte. It's quad, not quad-double. The trouble is that "long double" is usually just 80 bits stored in 12 or 16 bytes, and arithmetic is with the x87 unit. That's pretty silly, __float128 works the way you want it to.
 I'm too busy to futz with petsc-dev right
now, but would you ping me when this makes it out into a release?

Sure, I'll ping you when 3.2 is released.