Skip to content

compiler: Linearize Indexeds nested in other Indexeds' indices #3065

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We鈥檒l occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
ggorman wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

compiler: Linearize Indexeds nested in other Indexeds' indices #3065

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions devito/passes/iet/linearization.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
)
from devito.passes.iet.engine import iet_pass
from devito.passes.iet.parpragma import PragmaIteration
from devito.symbolics import retrieve_indexed, uxreplace
from devito.tools import filter_ordered, flatten, prod
from devito.types import Array, Bundle, FIndexed, Symbol, Wildcard
from devito.types.dense import DiscreteFunction
Expand Down Expand Up @@ -219,6 +220,12 @@ def linearize_accesses(iet, key0, tracker=None):
if v is not None:
subs[i] = v

# Uxreplace doesn't recurse into a replacement, so the Indexeds nested in
# another's indices are substituted here, innermost first. E.g.,
# `T[0, I[0, 0, 0], 0] -> TL0(0, IL0(0, 0, 0), 0)`
for i in sorted(subs, key=lambda i: len(retrieve_indexed(i.indices, deep=True))):

@mloubout mloubout Oct 19, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

that's the wrong fix, Uxreplace should be fixed instead.

@ggorman ggorman Oct 19, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed the root cause is in uxreplace: _uxreplace (symbolics/manipulation.py:80-83) returns a matched key's replacement verbatim, like xreplace, so with T[0, I[0,0,0], 0] -> TL0(0, I[0,0,0], 0) the inner I[0,0,0] (also a key) is never visited.

The catch is that any fix there turns the mapper from a simultaneous substitution into a transitive one somewhere, and uxreplace has ~86 call sites here and ~60 in PRO, so I'd like your view before picking one. I prototyped two (with the call-site resolution in this PR reverted, both fix the reported case and 3-level nesting, and the same set of suites passes, ~1230 tests):

  • A. General: on a match, keep substituting inside the replacement's args (the matched key masked so {x: x+1} still terminates). Changes simultaneous semantics, e.g. {x: y+1, y: x+1} on a[x,y,z] gives a[x+2, y+2, z] instead of a[y+1, x+1, z]; ~7% slower on a uxreplace micro-benchmark.
  • B. Indexed only: register _uxreplace for Indexed, and only when key and replacement are both Indexed substitute inside the replacement's indices. Everything else unchanged; the one behaviour change is {a[x,y,z]: b[x,y,z], x: y} on a[x,y,z] now giving b[y,y,z] (was b[x,y,z]). No measurable cost.
  • C. Opt-in: A's mechanics behind Uxreplace(subs, nested=True) (Transformer already has a nested kwarg), used only by linearization. No change for anyone else, a bit more plumbing.

I'd lean to B, but you know the callers better. Which would you prefer?

subs[i] = uxreplace(subs[i], subs)

iet = Uxreplace(subs).visit(iet)

# 2) What `iet` *offers*
Expand Down
7 changes: 6 additions & 1 deletion devito/passes/iet/misc.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,7 +15,8 @@
from devito.passes.iet.engine import iet_pass
from devito.passes.iet.languages.C import CPrinter
from devito.symbolics import (
Cast, RoundUp, ValueLimit, evalrel, has_integer_args, limits_mapper
Cast, RoundUp, ValueLimit, evalrel, has_integer_args, limits_mapper, retrieve_indexed,
uxreplace
)
from devito.tools import Bunch, as_mapper, as_tuple, filter_ordered, split
from devito.types import FIndexed
Expand Down Expand Up @@ -201,6 +202,10 @@ def _generate_macros_findexeds(iet, sregistry=None, tracker=None, **kwargs):
subs[i] = v
tracker[i.base] = Bunch(header=header, v=v)

# Nested FIndexeds, e.g. `TL0(0, IL0(0, 0, 0), 0)`, innermost first
for i in sorted(subs, key=lambda i: len(retrieve_indexed(i.indices, deep=True))):
subs[i] = uxreplace(subs[i], subs)

iet = Uxreplace(subs).visit(iet)

return iet
Expand Down
31 changes: 29 additions & 2 deletions tests/test_linearize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,8 +3,8 @@
import scipy.sparse

from devito import (
Eq, Function, Grid, Inc, MatrixSparseTimeFunction, Operator, SparseTimeFunction,
TimeFunction, configuration, sin, switchconfig
ConditionalDimension, Eq, Function, Grid, Inc, MatrixSparseTimeFunction, Operator,
SparseTimeFunction, TimeFunction, configuration, sin, switchconfig
)
from devito.ir import Call, Callable, DummyExpr, Expression, FindNodes, SymbolRegistry
from devito.passes import Graph, generate_macros, linearize
Expand Down Expand Up @@ -737,3 +737,30 @@ def test_bundle_derived_dim_stride():
assert bundle.symbolic_shape[p] is not bundle.symbolic_shape[rp]
assert bundle.symbolic_shape[rp] == 2
assert bundle.symbolic_shape.dindex(rp) == 1


@pytest.mark.parametrize('opt', ['noop', ('advanced', {'linearize': True})])
def test_nested_indexeds_in_guard(opt):
grid = Grid(shape=(4, 4))
time = grid.time_dim

T = TimeFunction(name='T', grid=grid, time_order=0, space_order=0)
ind = TimeFunction(name='I', grid=grid, time_order=0, space_order=0,
dtype=np.int32)
f = TimeFunction(name='f', grid=grid, time_order=0, space_order=0, save=4)

# The guard reads `T` through an indirection
cd = ConditionalDimension(name='cd', parent=time,
condition=T[0, ind[0, 0, 0], 0] < 1.5)
eqns = [Eq(T, time), Eq(f, T + 1, implicit_dims=cd)]

op = Operator(eqns, opt=opt, name='NestedGuard')

if opt != 'noop':
assert 'TL0(0, IL0(0, 0, 0), 0) < 1.5' in str(op)

ind.data[:] = 2
expected = np.array([1, 2, 0, 0])[:, None, None]

op.apply(time_M=3)
assert np.all(f.data == expected)
Loading