Minimal broadcasting engine for c++

[serge-sans-paille]

This is somehow simpler (but probably less generic) than @wolfv solution. Probably a good starter before eating wolv's one :-)

[oxinabox]

Yes,
I had big hopes for templating.

[SimonDanisch]

Pretty cool :) I'll look into it in more detail!

[SimonDanisch]

Maybe we can develop an interface for the C++ solutions:
benchmark.cpp:

#include "solution.cpp"
extern "C" 
{
  void benchmark_broadcast(){
    auto op = make_lazy([](auto& out, auto x, auto y, auto z) { out = x + y - std::sin(z); });
    materialize(op)
  }
  void benchmark_sum(){
    ...
  }
}

Not sure if we should generate the elements in benchmark, have globals, or pass them to the function. The advantage would be, that we very easily could benchmark this from Julia with the exact same benchmarking library ;)
cc @wolfv

[SimonDanisch]

And I would have access to the benchmark code in a more orderly manner ;)

[wolfv]

Well, that will generate a bunch of questions though.

For example, if we pass pointers to the interface, C++ can't be sure that they don't alias so not all optimizations can be enabled an these kind of things. However, benchmarking the creation as well will require me to write more code (as we'll need to use a uninitialized vector type) and this kind of stuff.

We might be able to develop an interface which initializes a couple of global C++ arrays, and then calls functions without parameters to bench.

Also, I think this might be a bit overkill. I think the minimal time for one iteration gives a good indication of the quality of the generated code.

[SimonDanisch]

We might be able to develop an interface which initializes a couple of global C++ arrays, and then calls functions without parameters to bench.

Sounds good ;)

Also, I think this might be a bit overkill. I think the minimal time for one iteration gives a good indication of the quality of the generated code.

Yeah I guess ... I may, or may not, want to use the minimal C interface to further play around with it, though :P
Anyways, it was just an idea to organize the code for me to immediately see what get's benchmarked... So putting the different benchmarks in their own function might be a good organization anyways.

[serge-sans-paille]

@SimonDanisch please tell me updated if you need more things in that example. I didn't add benchmark because there is no native 2D array in C++...