[async] Switch to llvm::SmallVector<llvm::SmallSet> for edges

[k-ye]

According to #1927, llvm::SmallVector<llvm::SmallSet> seems to be the most performant.

Here's the profiling data for async_mgpcg.py:

• Before this change: 13.418 s + 18.605 s + 20.120 s

     54.404  s synchronize                   [8 x   6.800  s]
         11.333  s 20.83%  optimize_listgen      [75 x 151.111 ms]
              ... ...
              8.523  s 75.21%  rebuild_graph         [44 x 193.713 ms]
                  6.988  s 81.98%  insert_task           [397621 x  17.574 us]
                      0.180  s  2.57%  get_task_meta         [397621 x 451.757 ns]
                      2.874  s 41.14%  insert_task meta->input_states [1149493 x   2.501 us]
                          2.069  s 71.96%  insert_edge           [1149493 x   1.800 us]
                          0.806  s 28.04%  [unaccounted]
                      2.053  s 29.38%  insert_task meta->output_states [409654 x   5.011 us]
                          1.286  s 62.63%  insert_edge           [1063555 x   1.209 us]
                          0.767  s 37.37%  [unaccounted]
                      1.007  s 14.41%  insert_task latest_state_readers_ [1149493 x 876.057 ns]
                      0.874  s 12.50%  [unaccounted]
                  0.005  s  0.06%  reid_nodes            [44 x 108.058 us]
                  0.002  s  0.03%  reid_pending_nodes    [44 x  53.108 us]
                  1.529  s 17.93%  [unaccounted]
              ... ...
          ... ...
          9.339  s 17.17%  demote_activation     [71 x 131.542 ms]
              ... ...
              8.576  s 91.82%  rebuild_graph         [40 x 214.396 ms]
                  7.055  s 82.27%  insert_task           [393579 x  17.925 us]
                      0.177  s  2.51%  get_task_meta         [393579 x 449.787 ns]
                      2.950  s 41.81%  insert_task meta->input_states [1178512 x   2.503 us]
                          2.123  s 71.96%  insert_edge           [1178512 x   1.801 us]
                          0.827  s 28.04%  [unaccounted]
                      2.018  s 28.61%  insert_task meta->output_states [406848 x   4.960 us]
                          1.379  s 68.33%  insert_edge           [1120853 x   1.230 us]
                          0.639  s 31.67%  [unaccounted]
                      1.031  s 14.61%  insert_task latest_state_readers_ [1178512 x 874.833 ns]
                      0.879  s 12.46%  [unaccounted]
                  0.004  s  0.05%  reid_nodes            [40 x 107.372 us]
                  0.002  s  0.02%  reid_pending_nodes    [40 x  49.353 us]
                  1.515  s 17.66%  [unaccounted]
              0.419  s  4.48%  [unaccounted]

• After this change: 11.009 s + 14.249 s + 14.891 s

42.394  s synchronize                   [8 x   5.299  s]
          7.644  s 18.03%  optimize_listgen      [75 x 101.920 ms]
              ... ...
              5.227  s 68.38%  rebuild_graph         [44 x 118.793 ms]
                  4.890  s 93.55%  insert_task           [397621 x  12.298 us]
                      0.185  s  3.79%  get_task_meta         [397621 x 466.267 ns]
                      1.459  s 29.84%  insert_task meta->input_states [1149493 x   1.269 us]
                          0.653  s 44.79%  insert_edge           [1149493 x 568.482 ns]
                          0.806  s 55.21%  [unaccounted]
                      1.285  s 26.27%  insert_task meta->output_states [409654 x   3.136 us]
                          0.494  s 38.47%  insert_edge           [1063555 x 464.718 ns]
                          0.790  s 61.53%  [unaccounted]
                      1.048  s 21.44%  insert_task latest_state_readers_ [1149493 x 912.139 ns]
                      0.912  s 18.65%  [unaccounted]
                  0.005  s  0.09%  reid_nodes            [44 x 108.621 us]
                  0.002  s  0.04%  reid_pending_nodes    [44 x  48.941 us]
                  0.330  s  6.31%  [unaccounted]
              ... ...
          ... ...
          6.117  s 14.43%  demote_activation     [73 x  83.791 ms]
              ... ...
              5.414  s 88.52%  rebuild_graph         [42 x 128.916 ms]
                  5.079  s 93.81%  insert_task           [415641 x  12.220 us]
                      0.189  s  3.72%  get_task_meta         [415641 x 454.920 ns]
                      1.564  s 30.80%  insert_task meta->input_states [1244758 x   1.257 us]
                          0.701  s 44.82%  insert_edge           [1244758 x 563.253 ns]
                          0.863  s 55.18%  [unaccounted]
                      1.232  s 24.26%  insert_task meta->output_states [429700 x   2.867 us]
                          0.550  s 44.61%  insert_edge           [1184969 x 463.830 ns]
                          0.682  s 55.39%  [unaccounted]
                      1.126  s 22.17%  insert_task latest_state_readers_ [1244758 x 904.753 ns]
                      0.968  s 19.05%  [unaccounted]
                  0.005  s  0.09%  reid_nodes            [42 x 121.593 us]
                  0.002  s  0.04%  reid_pending_nodes    [42 x  47.105 us]
                  0.328  s  6.06%  [unaccounted]
              0.337  s  5.52%  [unaccounted]

Unfortunately, it could be due to the recent macOS 10.15.7 upgrade, the GUI has become much slower on my end recently. I wasn't able to perceive improvements just by looking at some of the examples...

Related issue = #742

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1936 into master will increase coverage by 0.92%.
The diff coverage is 16.66%.

@@            Coverage Diff             @@
##           master    #1936      +/-   ##
==========================================
+ Coverage   42.80%   43.72%   +0.92%     
==========================================
  Files          45       45              
  Lines        6413     6207     -206     
  Branches     1101     1103       +2     
==========================================
- Hits         2745     2714      -31     
+ Misses       3498     3322     -176     
- Partials      170      171       +1     

Impacted Files	Coverage Δ
python/taichi/misc/gui.py	8.68% <16.66%> (ø)
...d/taichi-dev/taichi/python/taichi/misc/__init__.py
...ild/taichi-dev/taichi/python/taichi/code_format.py
.../taichi-dev/taichi/python/taichi/make_changelog.py
...build/taichi-dev/taichi/python/taichi/lang/meta.py
...build/taichi-dev/taichi/python/taichi/lang/expr.py
...aichi-dev/taichi/python/taichi/lang/transformer.py
...-dev/taichi/python/taichi/lang/kernel_arguments.py
.../build/taichi-dev/taichi/python/taichi/diagnose.py
...ld/taichi-dev/taichi/python/taichi/tools/np2ply.py
... and 81 more

---

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 4fa0d60...567b599. Read the comment docs.

[xumingkuan]

Great! Thank you so much!

Maybe the next step is to refactor

std::unordered_map<AsyncState, Node *> latest_state_owner_;
std::unordered_map<AsyncState, std::unordered_set<Node *>>
    latest_state_readers_;

-- I believe they take up most of the [unaccounted] time in insert_task.

[yuanming-hu]

LGTM! Thank you so much! It's nice to see insert_edge is now 3x faster :-)