THG's substitution applier

[matthiasdiener]

Based on 61690f4.

Addresses #162 in part.

Needs:

• THG's substitution applier pymbolic#105

[matthiasdiener]

This currently fails in some of the codegen tests with

[...]
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 153, in __call__
    return method(expr, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pytato/pytato/target/loopy/codegen.py", line 645, in map_reduce
    inner_expr = self.rec(inner_expr, prstnt_ctx,
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 153, in __call__
    return method(expr, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 615, in map_substitution
    child = self.rec(expr.child, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 153, in __call__
    return method(expr, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pytato/pytato/target/loopy/codegen.py", line 586, in map_subscript
    return result.to_loopy_expression(self.rec(expr.index, prstnt_ctx,
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 151, in __call__
    return self.map_foreign(expr, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 206, in map_foreign
    return self.map_tuple(expr, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 583, in map_tuple
    children = tuple([self.rec(child, *args, **kwargs) for child in expr])
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 583, in <listcomp>
    children = tuple([self.rec(child, *args, **kwargs) for child in expr])
  File "/Users/mdiener/Work/emirge/pymbolic/pymbolic/mapper/__init__.py", line 153, in __call__
    return method(expr, *args, **kwargs)
  File "/Users/mdiener/Work/emirge/pytato/pytato/target/loopy/codegen.py", line 605, in map_variable
    array = local_ctx.lookup(expr.name)
  File "/Users/mdiener/Work/emirge/pytato/pytato/target/loopy/codegen.py", line 165, in lookup
    return self.local_namespace[name]
KeyError: '_r0'

[kaushikcfd]

Thanks for bringing this back to life! Also, I'm curious are we currently running into any cases where this could be profitable?

[kaushikcfd]

For context, on some large expression graphs implemented in drivers-y2-isolator, we obtain this profile -- https://gist.github.com/kaushikcfd/0867bd549acfff43781077e8f8b0ba56 spanning around 8 minutes for walltime, and I don't see substitute(..) accounting for a significant amount in the flamegraph.

[matthiasdiener]

I think illinois-ceesd/mirgecom#645 was a case where we saw extremely long times spent in substitute.

[kaushikcfd]

I think illinois-ceesd/mirgecom#645 was a case where we saw extremely long times spent in substitute.

Aah! (Thanks!)
After trying to run the example, it still seems to take a very long time to compile. At this point, it seems to me that there are some mappers that aren't using memorized variants. (c.f. inducer/loopy#625). Seems like there are more mappers like the one in loopy.realize_reduction that do not use the memoized variants. At this point I don't have enough data to blame the substitute here, if you already have some data, please do share.