[OpenGL] [perf] Support TLS to improve reduction performance

[archibate]

Related PR = #1336

[Click here for the format server]

---

Running on my poor Intel card:
fem99: 4.6 -> 40 (!!!!)
fem128: 18 -> 26
mpm_lagrangian_forces: 3.8 -> 15.7
misc/benchmark_reduction.py: GL_OUT_OF_MEMORY :(

No performance difference observed on the NV card, seems NV's compiler is already capable of doing TLS optimization for us?

(Partial) OFT:
[cli] [opengl] [metal] TI_MAKE_THREAD_LOCAL=0 to disable TLS on OpenGL/Metal.

[yuanming-hu]

Thank you! LGTM. My only concerns is that thls seems to be used in place of tls in some places. Is that intentional?

[codecov]

Codecov Report

Merging #1574 into master will increase coverage by 0.05%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #1574      +/-   ##
==========================================
+ Coverage   67.47%   67.53%   +0.05%     
==========================================
  Files          40       40              
  Lines        5624     5630       +6     
  Branches      982      982              
==========================================
+ Hits         3795     3802       +7     
+ Misses       1661     1660       -1     
  Partials      168      168              

Impacted Files	Coverage Δ
python/taichi/lang/__init__.py	80.20% <ø> (-0.01%)	⬇️
python/taichi/lang/ops.py	93.33% <0.00%> (+0.81%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update f528377...05ade2f. Read the comment docs.

[archibate]

Thank you! LGTM. My only concerns is that thls seems to be used in place of tls in some places. Is that intentional?

Sorry, that's intentional. For now all the buffers, root, args, gtmp, ..., are 4 chars.
But I think it's OK to make tls to be 3 chars since it's not a global buffer, but a per-thread buffer. Thank for pointing out.