Warn on potential data races

[luca-patrignani]

This PR aims to raise a warning if two or more memlets are writing to the same access node and their ranges are overlapping. It also adds a new config entry to activate this check.
This PR tries to close issue #424.

Needs confirmation:

• Are the new tests overcomplicated?
• Do we need some unit tests for are_intersecting?

[luca-patrignani]

@tbennun when you got some free time can you review this? Thanks

[tbennun]

Thank you for the contribution! Overall this is a good start, but it will not cover the majority of cases of data races that may occur in dace programs (see comments). Also, please use the subsets.intersects API.

[tbennun]

This is much improved! The data race check still needs one more change to be correct (see comment)

[tbennun]

A few notes on this test:

1. You should keep the actual access nodes in your dictionary. If a map that reads from A[i] also writes to A[i], that would trigger a validation warning incorrectly. Instead you need to check if:

• The node is the same (in that case, write-write conflicts may still happen)
• If the nodes are not the same, whether there is a path from the first to the second (use has_path to check that). If there is a path, then there is no conflict (read/write or write/write). This is another reason for this check being expensive. There may be some analysis passes in dace.transformation.passes that give you the paths between dependent access nodes, but creating a dictionary would be an optimization. To be safe you can use has_path for now and leave a comment that this could be optimized.

2. Write-conflict resolution memlets (memlet.wcr is not None) are exempt from write-write conflicts (maybe not read-write)

Please also add an elementwise map test (A->mapentry-A[i]->tasklet-A[i]->mapexit->A) just to make sure.

[luca-patrignani]

I tried to follow your notes but I felt kind of lost, especially for point 2. Consider this two test cases:

def test1():
    sdfg = dace.SDFG('memlet_overlap_with_wcr')
    state = sdfg.add_state()
    sdfg.add_array("A", (20,), dace.int32)
    sdfg.add_array("B", (1,), dace.int32)
    A = state.add_read("A")
    B = state.add_write("B")

    state.add_mapped_tasklet(
        name="first_reduction",
        code="b = a",
        inputs={"a": dace.Memlet(data="A", subset="k")},
        outputs={"b": dace.Memlet(data="B", subset="0", wcr="lambda old, new: old + new")},
        map_ranges={"k": "0:20"},
        external_edges=True,
        input_nodes={"A": A},
        output_nodes={"B": B}
    )

    state.add_mapped_tasklet(
        name="second_reduction",
        code="b = a",
        inputs={"a": dace.Memlet(data="A", subset="k")},
        outputs={"b": dace.Memlet(data="B", subset="0", wcr="lambda old, new: old * new")},
        map_ranges={"k": "0:20"},
        external_edges=True,
        input_nodes={"A": A},
        output_nodes={"B": B}
    )

    with warnings.catch_warnings():
        warnings.simplefilter("error", UserWarning)
        with dace.config.set_temporary('experimental', 'check_race_conditions', value=True):
            sdfg.validate()

def test2():
    sdfg = dace.SDFG('memlet_overlap_with_wcr')
    state = sdfg.add_state()
    sdfg.add_array("A", (20,), dace.int32)
    sdfg.add_array("B", (1,), dace.int32)
    A = state.add_read("A")
    B = state.add_write("B")

    state.add_mapped_tasklet(
        name="first_reduction",
        code="b = a",
        inputs={"a": dace.Memlet(data="A", subset="k")},
        outputs={"b": dace.Memlet(data="B", subset="0", wcr="lambda old, new: old + new")},
        map_ranges={"k": "0:20"},
        external_edges=True,
        input_nodes={"A": A},
        output_nodes={"B": B}
    )

    state.add_mapped_tasklet(
        name="second_reduction",
        code="b = a",
        inputs={"a": dace.Memlet(data="A", subset="k")},
        outputs={"b": dace.Memlet(data="B", subset="0")},
        map_ranges={"k": "0:20"},
        external_edges=True,
        input_nodes={"A": A},
        output_nodes={"B": B}
    )

    with warnings.catch_warnings():
        warnings.simplefilter("error", UserWarning)
        with dace.config.set_temporary('experimental', 'check_race_conditions', value=True):
            sdfg.validate()

Are they considered data races? At this moment my code raises no warnings.
I'm probably missing something important, sorry for this long PR review.

[luca-patrignani]

@tbennun when I got some free time can you review the last changes? I left a reply on your last comment.

[phschaad]

Pinging @tbennun , I assume this was burried

[tbennun]

Thank you for making the corrections, LGTM!

[tbennun]

@luca-patrignani @phschaad if possible, can we revive this so we can merge it? Thanks!

[phschaad]

@tbennun The PR is revived in #1712 :)