Inject LetStmt of variables that are used downstream modulo some constant. - #9428
Draft
mcourteaux wants to merge 6 commits into
Draft
Inject LetStmt of variables that are used downstream modulo some constant.#9428mcourteaux wants to merge 6 commits into
mcourteaux wants to merge 6 commits into
Conversation
… LetStmt Distilled from NeonRAW's BilinearDemosaicAlignedGenerator, which selects an interpolation per Bayer phase with a mux over (x - bayer_offset) % 2 and is scheduled with splits aligned to that same offset so the phase is constant per tile. Twelve muxes survive lowering there; this reduces it to two. The producer is computed inside the consumer's aligned tile and then split again to vectorize. That second split binds the reconstructed loop variable to a LetStmt whose value carries the "+ off" from the aligned min. The simplifier can see through a Let, but not through a LetStmt, so at the use site it only ever sees "- off" applied to an opaque name, and (x - off) % 2 never reduces to a function of the inner loop variable alone. Confirmed by simplifying the same expression both ways: behind a LetStmt neither phase folds, while inlining the value folds both to constants -- and it still folds if the value is inlined only after the mod-2 sign flip has already rewritten "- off" into "+ off", so the LetStmt is the sole cause. The test fails at present; it documents the missed simplification. The generated code is correct, only the muxes are not resolved. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JbXuRsMMLQDkqE3mwtEKMs
A vectorized loop over a repeating pattern only resolves its phases at compile time if it gets deinterleaved by the pattern's period, so that each resulting slice has a single phase. The stride search stopped at four, which covered a Bayer sensor's period of two but not an X-Trans sensor's six. It can't simply go to eight: deinterleaving a period of eight into eight slices displaces permutes that a target may do in one instruction, and three of simd_op_check_hvx's vdelta checks regress that way. Seven is the largest value that leaves those alone. A period of eight still resolves if the loop is unrolled rather than vectorized. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JbXuRsMMLQDkqE3mwtEKMs
Generalizes the test over the period of the pattern the mux selects on, and runs it for two (Bayer), three, and six (X-Trans) rather than two alone. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01JbXuRsMMLQDkqE3mwtEKMs
A value is often needed only modulo some constant: x % 5, (x + 3) % 5 and (x + 2*y) % 3 all need no more of x than x % 5. Reducing the value where it is defined frequently collapses it. A loop variable reconstructed by an aligned split looks like xo * 16 + offset, which modulo two is just offset. The simplifier can't discover that at the use site. A LetStmt hides the value behind a name, and it won't substitute a whole expression back in. It will substitute a variable, so this pass binds the reduced value to one: (x - offset) % 2 becomes (x.mod.2 - offset) % 2, and with x.mod.2 bound to offset that folds to zero. This is what lets a demosaicer scheduled with aligned splits resolve its Bayer phases at compile time; twelve muxes survived lowering in the distilled test before. Only reductions that collapse to a constant or a single variable are bound, so the pass can never grow the IR, and the re-simplification afterwards is skipped when it bound nothing. On a pipeline that benefits, compiling gets faster rather than slower, because everything downstream sees smaller IR: 0.83s against 0.96s for twenty compilations of a demosaicer. On one that doesn't, it is free: 0.354s against 0.349s for a plain blur. Reducing a value looks through the lets it is built from, to a bounded depth, so that terms which cancel get the chance to meet, and it takes the conditions of dominating asserts as assumptions -- a require() on an offset is often the only thing saying it is small enough to be its own remainder. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016WxFGPpnekN4bH6sMCsQBQ
The correctness test needs no entry: the Makefile globs test/correctness. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016WxFGPpnekN4bH6sMCsQBQ
mcourteaux
force-pushed
the
mcourteaux/inject-var-mod-letstmt
branch
from
September 4, 2026 20:10
89c9ac0 to
b3e0958
Compare
Codecov Report❌ Patch coverage is
Additional details and impacted files@@ Coverage Diff @@
## mcourteaux/aligned-split-clean #9428 +/- ##
==================================================================
+ Coverage 70.17% 70.25% +0.08%
==================================================================
Files 261 262 +1
Lines 79497 79707 +210
Branches 19379 19450 +71
==================================================================
+ Hits 55786 56000 +214
- Misses 17865 17873 +8
+ Partials 5846 5834 -12 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Continues on #9409. Makes aligned splits work on an aligned split, for which the inner loop is then vectorized with an additional split factor.
Idea for the lowering pass by @abadams
Breaking changes
Checklist