Make L2Denorm an array encoding instead of a scalar function - #9138
Open
connortsui20 wants to merge 1 commit into
Open
Make L2Denorm an array encoding instead of a scalar function#9138connortsui20 wants to merge 1 commit into
connortsui20 wants to merge 1 commit into
Conversation
This was referenced Aug 2, 2026
connortsui20
force-pushed
the
ct/l2-denorm-encoding
branch
from
August 2, 2026 18:03
80daa53 to
63e459b
Compare
connortsui20
marked this pull request as ready for review
August 3, 2026 15:11
Member
Author
|
There are things I need to touch up here but would appreciate review on the goal of this PR (making this scalar function an encoding instead) |
connortsui20
force-pushed
the
ct/l2-denorm-encoding
branch
2 times, most recently
from
August 3, 2026 15:29
d62d2a9 to
8598018
Compare
Merging this PR will improve performance by 10.2%
Performance Changes
Tip Curious why this is faster? Comment Comparing 
Footnotes
|
connortsui20
commented
Aug 3, 2026
Comment on lines
+286
to
+287
| &ConstantArray::new(normalized, 1).into_array(), | ||
| &ConstantArray::new(norms, 1).into_array(), |
Member
Author
There was a problem hiding this comment.
what happened to the conversation about having a special class of scalars that are just length-1 arrays?
`L2Denorm` was registered as a `ScalarFnVTable`, but it never behaved like one. Its constructor took an `ExecutionCtx` and scanned both children to enforce a unit-norm invariant, `L2Norm` read its stored norms instead of recomputing, `CosineSimilarity` and `InnerProduct` reached into its physical children, and the compressor scheme named it as a produced encoding. Those are all properties of a physical decomposition, not of an operation over arbitrary well-typed values. Moves it to `vortex-tensor/src/encodings/l2_denorm/` as a real `VTable` with two slots (`normalized`, `norms`). Structural validation runs on construction and on deserialization, `try_new` additionally scans for the exact unit-norm invariant, and `try_new_trusted` skips that scan for lossy normalized children whose stored norms stay authoritative. Neither constructor is `unsafe`, since violating the contract produces wrong answers rather than undefined behavior. The encoding keeps the `vortex.tensor.l2_denorm` array ID and the same two-field metadata message, so the wire format is unchanged. Slice and filter now push down into both children through `reduce_parent`. The generic `ScalarFnArray` filter rule only fired when at most one child was non-constant, which for this encoding was almost never. Also makes `L2DenormScheme` cascade its two children like `TemporalScheme` does, which lets both `HACK TO SUPPORT L2 DENORMALIZATION` special cases come out of `CascadingCompressor`. The scheme now competes on measured size like every other scheme. Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
connortsui20
force-pushed
the
ct/l2-denorm-encoding
branch
from
August 3, 2026 15:42
8598018 to
780d4d0
Compare
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
2 participants
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.
Rationale for this change
(semi-related) Tracking Issue: #9129
I have a new proposed idiom (cc @gatesn @joseph-isaacs @robert3005): encodings may hold invariants over their children, scalar functions may not.
I think that a scalar function should just be an operation over any well-typed input. Our encoding, however, can "own" their physical decomposition, so they get validation in their constructors and can enforce whatever invariants it needs to between its children.
FoRis the one edge case that I can think of? Its child is an offset from the reference, which is an invariant, but nothing enforces it andencoded + referencewould work as a plain scalar function over any integer column.Anyways,
L2Denormwas always a very strange scalar function, as I had to match against it if I ever wanted to guarantee that a vector array was normalized, and it was just very awkward. This PR changesL2Denorminto an encoding instead of a scalar function.I am also happy to change the name to something else (maybe just "Normalized"?).
What changes are included in this PR?
The unfortunate thing about this is that a lot of mechanical things have to be moved, but the slightly-hairy logic does not. So the large diff here is not a real.
L2Denormbecomes aVTableinvortex-tensor/src/encodings/l2_denorm/with slotsnormalizedandnorms.try_newscans for the exact invariants,try_new_trustedskips the scan for lossy normalized children. Neither isunsafe, since breaking the contract gives wrong answers rather than undefined behavior.ScalarFnArrayrule only fires when at most one child is non-constant.L2DenormSchemecascades its children likeTemporalScheme, which removes bothHACK TO SUPPORT L2 DENORMALIZATIONcases fromCascadingCompressor.What APIs are changed? Are there any user-facing changes?
Details
scalar_fns::l2_denorm::*encodings::l2_denorm::*L2Denorm::try_new_array(n, s, ctx)L2Denorm::try_new(n, s, ctx)unsafe L2Denorm::new_array_unchecked(n, s)L2Denorm::try_new_trusted(n, s)L2Denorm::new()session.scalar_fns().register(L2Denorm)session.arrays().register(L2Denorm), unconditionalScalarFnArrayPlugin::new(L2Denorm)VX_SCALAR_FN_ARRAY_TENSOR_PLUGINno longer gatesL2Denorm, since the compressor can emit it. The array ID andL2DenormMetadataare unchanged, so the on-disk bytes match.L2DenormSchemestays out ofALL_SCHEMES.