Refactor FEInterface to speed up dof_indices?

[libmesh]

Refactor FEInterface to speed up dof_indices?

[roystgnr]

On Fri, 26 Apr 2013, Benjamin S. Kirk wrote:

I was looking into the DofMap::dof_indices()function to see what I might do to speed it up.

Sounds great! It'll be interesting to figure out how to benchmark the
change well, though...

What I see is that it becomes a double switch which dispatches to another double switch, and it
gets called obviously for each node on the element.

I was thinking about making the current n_dof_at_node() a private member of FEInterface, and
using it to build up static arrays, say

_n_dofs_at_node[DIM][FETYPE][ELEM_TYPE][NODENUM];

Isn't "turn switch statements into branch tables" a common compiler
optimization anyway? Although that's a compiler optimization we may
be inadvertently disabling with our sparse enum. It'd be interesting
to start by just packing the enum and seeing if the code gets any
faster with g++ or icpc.

but to do this right, it will probably involve cleaning up our
FE_FAMILY enum so it is packed, which means writing a string or
something instead of a number to the restart files.

@roystgnr, @jwpeterson, @pbauman, @friedmud - Thoughts on this

I want to support old xdr/xda restart files indefinitely, so that's
going to require maintaining an "old number to new-number/string"
table somewhere.

I'd like us to move to something more extensible for a restart file
format in the future. I was originally thinking hdf5-based; now that
that describes exodus/nemesis too I'm wondering if it might be
possible to add extensions there. This might be too much work for us
to bite off right now, but it might be a reason to just use the old
numbering via translation in XDR/XDA I/O for now, then switch to a
less ambiguous string identifier at the same time we're changing the
rest of the format too.

or on the broader issue of FEInterface altogether?

I despised the mass of switches in FEInterface until roughly last
week, when I discovered myself having to recommend basically the exact
same design pattern to someone to make up for the inability to do

templated virtual functions in C++.

Roy

[roystgnr]

Yaagh. I thought maybe the email-mangling that you guys had noticed had something to do with HTML email, but the quote characters and paragraph breaks in my alpine output just got shredded there.

[benkirk]

Isn't "turn switch statements into branch tables" a common compiler optimization anyway? Although that's a compiler optimization we may be inadvertently disabling with our sparse enum. It'd be interesting to start by just packing the enum and seeing if the code gets any faster with g++ or icpc.

It is, but when we have a double switch that then makes a function call to another double switch in a different translation unit and don't usually do inter-procedural optimization I can't imagine a compiler in the world that could do that too efficiently...

As for the restart bit - yes, probably worth its own thread. In the meantime I can have a simple int->string mapping to handle old restart files.

[friedmud]

On Fri, Apr 26, 2013 at 1:20 PM, roystgnr [email protected] wrote:

Sounds great! It'll be interesting to figure out how to benchmark the
change well, though...

I will be able to tell you how much of a speed up it is... we spend a
decent amount of time in dof_indices()....