[async] Add allocator async state

[k-ye]

taichi/taichi/lang_util.cpp

Lines 156 to 158 in ceea597

	bool is_gc_able(SNodeType t) {
	return (t == SNodeType::pointer || t == SNodeType::dynamic);
	}

Before this PR, test_block_async() would actually fail. Here's the SFG. Note that all GC tasks are floating (as expected):

With this PR, SFG:

Highlighting the allocator state:

Related issue = #742

[Click here for the format server]

---

[codecov]

Codecov Report

Merging #1973 into master will increase coverage by 0.96%.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #1973      +/-   ##
==========================================
+ Coverage   42.76%   43.72%   +0.96%     
==========================================
  Files          45       45              
  Lines        6431     6221     -210     
  Branches     1105     1105              
==========================================
- Hits         2750     2720      -30     
+ Misses       3507     3328     -179     
+ Partials      174      173       -1     

Impacted Files	Coverage Δ
python/taichi/lang/ast_checker.py	70.58% <0.00%> (-1.64%)	⬇️
python/taichi/testing.py	75.00% <0.00%> (-0.72%)	⬇️
python/taichi/lang/linalg.py	89.33% <0.00%> (-0.67%)	⬇️
python/taichi/lang/__init__.py	40.45% <0.00%> (-0.33%)	⬇️
python/taichi/misc/util.py	20.54% <0.00%> (-0.21%)	⬇️
python/taichi/misc/task.py	0.00% <0.00%> (ø)
python/taichi/tools/patterns.py	0.00% <0.00%> (ø)
python/taichi/lang/kernel.py	71.39% <0.00%> (+0.13%)	⬆️
python/taichi/misc/gui.py	8.89% <0.00%> (+0.18%)	⬆️
python/taichi/main.py	22.95% <0.00%> (+0.35%)	⬆️
... and 8 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 0298bca...03e7f3d. Read the comment docs.

[xumingkuan]

LGTM in general! I think we should also change some code in task fusion:

taichi/taichi/program/state_flow_graph.cpp

Lines 455 to 458 in b8a1bc4

	if (state.first.type != AsyncState::Type::value) {
	// No need to check mask/list states as there must be value states.
	continue;
	}

IIUC allocator states should be treated like value states here: if A->B is an allocator edge on SNode s, only if both A and B are element-wise (or loop-unique on the same value in the future), can A and B be fused. For example, these two tasks cannot be fused:

@ti.kernel()
def foo():
  for i in x:
    ti.activate(y, i + 1)

@ti.kernel()
def bar():
  for i in x:
    ti.deactivate(y, i)

BTW, do these SNodeOpStmts change mask states?

[k-ye]

LGTM in general! I think we should also change some code in task fusion:

taichi/taichi/program/state_flow_graph.cpp

Lines 455 to 458 in b8a1bc4

	if (state.first.type != AsyncState::Type::value) {
	// No need to check mask/list states as there must be value states.
	continue;
	}

IIUC allocator states should be treated like value states here: if A->B is an allocator edge on SNode s, only if both A and B are element-wise (or loop-unique on the same value in the future), can A and B be fused. For example, these two tasks cannot be fused:

@ti.kernel()
def foo():
  for i in x:
    ti.activate(y, i + 1)

@ti.kernel()
def bar():
  for i in x:
    ti.deactivate(y, i)

Great catch! I've updated the code accordingly. However, note that I checked for mask state instead of allocator. IMHO, mask is directly related to sparsity, element_wise, etc. WDYT?

BTW, do these SNodeOpStmts change mask states?

Yes, thanks!

[k-ye]

I've confirmed this with the dumped dot graph. But task fusion is currently not recorded in the stats yet.

[xumingkuan]

Cool, thanks!