Handle reductions in get_insn_access_map

[majosm]

Currently, get_insn_access_map only passes the inames from within_inames in the call to get_access_map. As a result, if reductions are present get_access_map will may (edit: it only happens if the reduction domain is separate from the element domain, otherwise knl.get_inames_domain() picks up the reduction) fail, e.g.:

unable to determine access range of subscript: [_pt_temp_12_dim0, _pt_sum_r0_8] (encountered ExpressionToAffineConversionError: could not convert expression '_pt_sum_r0_8' to affine representation: UnknownVariableError: cannot find '_pt_sum_r0_8' in context (pass it in on evaluation))

This causes _compute_isinfusible_via_access_map to return True whenever reductions are present (code), which subsequently prevents get_kennedy_unweighted_fusion_candidates from fusing loops that it potentially could.

A real world example of this can be seen in the generated code for a DG operator here. I've annotated with # DESCRIPTION: what DG operations are being done in each loop in the device code. The first 3 loop nests are performing face-local work on the interior faces (the iel loop with 5641 iterations). Since the operations are face-local, the iel loops should be able to be fused, but the reductions over the idof axes are preventing that from happening due to this issue.

This PR changes get_insn_access_map to get_insn_access_maps. It now computes separate access maps for accesses outside of reductions + each different reduction level present, by traversing the instruction and updating the domain for reductions accordingly (AFAIK, they cannot all be unioned together due to the different spaces involved). It additionally modifies compute_isinfusible_via_access_map to do the unioning after projection once it's safe to do so.

With this change, I'm now seeing fusion of the element loops as expected: code.

[majosm]

I can't request review on this repo, so @kaushikcfd @inducer this is probably ready for a glance.

[majosm]

I should mention: this is more or less my first time doing anything nontrivial with isl, so extra scrutiny is likely warranted. 🙂

One thing I'm not clear on is whether I should be using project_out_except after get_inames_domain (e.g., in _InstructionAccessMapCollector.map_reduction).

[Copilot]

Pull request overview

This PR updates loop-fusion dependency analysis so access maps can be computed for instructions containing reductions whose domains are separate from the surrounding inames.

Changes:

• Replaces single instruction access-map collection with per-reduction-level access maps.
• Updates loop fusion to project and union multiple access maps before dependence checks.
• Adds a regression test covering fusion with an inner reduction in a separate domain.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
loopy/kernel/tools.py	Adds reduction-aware instruction access-map collection and exposes map unioning.
loopy/transform/loop_fusion.py	Consumes multiple access maps and unions them after projection.
test/test_loop_fusion.py	Adds regression coverage for fusing loops when a dependent read occurs inside a reduction.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.