Auto-generated explicit template instantiations

[Sam Preston]

I've been playing with some auto-generation craziness, but it's working well so I created a feature branch and am setting up this issue just to discuss it. If we have a source file that needs a bunch of explicit template instantiations in it (for instance FOpers.cxx), we create a file FOpers_inst.tpl that might look something like this:

def AllBG = [
BACKGROUND_STRATEGY_CLAMP,
BACKGROUND_STRATEGY_WRAP,
BACKGROUND_STRATEGY_ZERO,
BACKGROUND_STRATEGY_PARTIAL_ZERO,
BACKGROUND_STRATEGY_ID,
BACKGROUND_STRATEGY_PARTIAL_ID]

def BoolArg = [true, false]

template void Opers::
ApplyH<${AllBG}>(Field3D&, const Field3D&, const Field3D&, const float&, StreamT);

template void Opers::
Resample<${AllBG}, ${BoolArg}>(Field3D&, const Field3D&, StreamT);
.
.
.

During compilation FOpers_inst.tpl will be parsed into the expanded instantiations in FOpers_inst.cxx, and in FOpers.cxx we just add the line #include "FOpers_inst.cxx. In the CMakeLists.txt we just need to set up a list of .tpl files to process (as we do for source files) and I've written some code that correctly sets up the dependencies and generation to turn FOpers_inst.tpl into FOpers_inst.cxx (it's just a simple python script that does the parsing).

I've also set up the generation so that the file is first parsed by the C preprocessor and all definitions passed to the C compiler are also passed to the preprocessing of the .tpl files, so we can do things like:

#ifdef CUDA_ENABLED
def ExecMode = [EXEC_CPU, EXEC_GPU]
#else
def ExecMode = [EXEC_CPU]
#endif

which will help us as we work on issue #12.

[Jacob Hinkle]

Wow. Awesome. So I assume we can also keep various other lists, like

def AllDefBG = [
BACKGROUND_STRATEGY_WRAP,
BACKGROUND_STRATEGY_ID,
BACKGROUND_STRATEGY_PARTIAL_ID]

or similar, and one for vfields, and use those in the same way? This looks great. Since these are just lists of strings we can do lists of ints too right?

[Sam Preston]

Yeah, have a look at Code/Cxx/inc/math/FieldOpers_inst.tpl -- that's the best example with multiple replacement lists and some #ifdefs. Resample, for instance, used to take 24 explicit template instantiations to cover all combinations of parameters, and now it's just written once. Edit: added link to FieldOpers_inst.tpl so that you don't have to go searching for it

[Jacob Hinkle]

Wow. that's awesome. I notice that the version of Resample for instance that takes in bg and rescaleVector as arguments still has a manual series of ifs. Could that be handled the same way? I'd imagine the CMake magic could also just generate that function definition as well.

[Sam Preston]

Yeah, I'm sure it could be set up -- the actual parsing is just done by a python script I wrote. Parsing the template arguments and inserting them as function arguments is just more complicated.

[Sam Preston]

Merged in AutoTemplateInstantiationTest (pull request #3)

Merge AutoTemplateInstantiationTest into master. This closes issue #14

→ <<cset ade675840f97>>