PoC: multithreaded chunked model CPU computation

[illwieckz]

Chunked version of:

• PoC: multithreaded model CPU computation #1833

For unknown reasons, only one thread is spawned.

[illwieckz]

@slipher here is my chunked variant, but it doesn't spawn more than one thread.

Actually, it could be possible to write a custom dispatch function (I don't know how to do it).

[illwieckz]

I added a commit that uses standard threading functions instead of __gnu_parallel::for_each() or std::for_each( std::execution::par, …), that would allow us to not require libgomp or libtbb, but it still runs as if all was done sequentially.

I added a logger to check things and things look correct. I don't know what is missing.

[illwieckz]

Hmm, one drawback of doing it that way, is that unlike what does OpenMP, threads are not reused and then profilers like Orbit list thousands of threads and not only that amount is crazy to list, but profiling is just meh because computed statistics are for each thread separately. It would be cool to be able to reuse those threads.

[illwieckz]

But at least, if we could get the current implementation working that would be a start.

[illwieckz]

Hmm, actually, it seems to work, with my custom thread start. It's just so inefficient that performance drops like if nothing was done. When switching from 1 thread to 2 I see a performance difference. It's just so bad compared to OpenMP.

But now, I don't get why this chunked implementation doesn't work with OpenMP.

[illwieckz]

I got it working with OpenMP, I had to use another syntax (which in fact skips my useless vector trick). I now get 438fps with 16 threads, which is much faster! So yes my idea of controlling the way it is chunked was good!

[illwieckz]

With this implementation we are now as fast in the CPU code than when using the GPU code running on CPU with the llvmpipe software renderer. With LIBGL_ALWAYS_SOFTWARE=1 I get the exact same framerate wether I enable r_vboVertexSkinning or disable it. Before, the difference wasn't big, now I see none.

[illwieckz]

The experiment was a success. I close this and will submit a completed and cleaned-up branch later.