Understanding loop_priority

[inducer]

Discussed in #889

^{Originally posted by connorjward November 28, 2024}
I am trying to tackle a large number of warnings that loopy is now producing. Specifically we see a lot of:

/home/connor/Code/firedrake-dev3/src/loopy/loopy/schedule/__init__.py:2247: LoopyWarning: in kernel
 loopy_kernel: Falling back to a slow scheduler implementation due to: cannot schedule kernels with
 priority dependencies between sibling loop nests (add 'v1_scheduler_fallback' to silenced_warnings
 kernel attribute to disable)
  warn_with_kernel(

So far my attempts to address this have not worked. I was wondering if you could explain to me why the following produces the warnings:

import loopy as lp


knl = lp.make_kernel(
    [
        "{ [i]: 0 <= i < 5}",
        "{ [j]: 0 <= j < 5}",
        "{ [k]: 0 <= k < 5}",
    ],
    [
        lp.Assignment("x[i]", "x[i] + y[j]", within_inames=frozenset({"i", "j"}), id="id0"),
        lp.Assignment("x[i]", "x[i] + y[k]", within_inames=frozenset({"i", "k"}), depends_on=frozenset({"id0"})),
    ],
    [lp.GlobalArg("x", dtype=int), lp.GlobalArg("y", dtype=int)],
    loop_priority=frozenset({("i", "j"), ("i", "k")}),
    lang_version=(2018,2),
)

code = lp.generate_code_v2(knl)

AIUI the loop_priority I set here states that loop j must be inside i and that k must also be inside i. To me there is no 'priority dependency between sibling loop nests'.

Any help would be appreciated.

[connorjward]

@inducer @kaushikcfd what are your thoughts? If the fix for this will take some time then I will temporarily remove loop priorities from Firedrake/PyOP2 as they are producing thousands of warnings during our CI. Actually I will just disable the warning. Though the improved codegen performance would be nice to have.