[Perf] [opengl] Support 'ti.thread_dim(x)' to change invocations per thread for OpenGL

[archibate]

Related issue = unrevert #1676

[Click here for the format server]

---

Now I realized I was totally misunderstood the term stride, I changed it to invocation now. Will do name nit in C++ iapr.

See #1691 for the new terms:

Taichi programs could be executed in parallel on GPU.
The level of structure of GPU is defined hierarchically.

From small to large, the computation units are:
invocation < thread < block < grid.

• invocation:
  Invocation is the body of a for-loop.
  Each invocation corresponding to a specific i value in for-loop.

• thread:
  Invocations are grouped into threads.
  Threads are the minimal unit that is parallelized.
  All invocations within a thread are executed in serial.
  We usually use 1 invocation per thread for maximizing parallel performance.

• block:
  Threads are grouped into blocks.
  All threads within a block are executed in parallel.
  Threads within the same block can share their block local storage.

• grid:
  Blocks are grouped into grids.
  Grid is the minimal unit that being launched from host.
  All blocks within a grid are executed in parallel.
  In Taichi, each parallelized for-loop is a grid.

[yuanming-hu]

Thanks. I believe you are using thread_dim as an unrolling factor for the for-loops? I'm not sure if this will help improve performance on GPUs.

I lean towards not introducing one more concept for people to learn.

[archibate]

Closing, we need come to agree with a systematic solution for all these stuffs, block_dim, grid_dim, thread_dim, and make DecoratorRecorder more easy-to-add. FF2 put this into v0.8.0 roadmap.