Vector Extension Type by connortsui20 · Pull Request #6964 · vortex-data/vortex

connortsui20 · 2026-03-13T21:57:09Z

Summary

Tracking Issue: #6865

Adds a Vector extension type and a new L2Norm expression.

Additionally adds a AnyTensor type that can be matched on for any kind of tensor we want.

Right now the code assumes that everything is built on top of FixedSizeList, but in the future that might change.

Additionally make some touchups to the vortex-tensor crate in general.

API Changes

The new Vector and L2Norm types.

Testing

Some basic tests.

gatesn · 2026-03-16T18:58:16Z

vortex-array/src/scalar_fn/vtable.rs

        Ok(args.to_vec())
    }

+    // TODO(connor): This needs a precondition for the number of args it has, or all implementations


There should already be a check using the arity of the function

I tried backtracing through the code myself but I couldn't find where that is called

vortex-tensor/src/scalar_fns/cosine_similarity.rs

vortex-tensor/src/scalar_fns/l2_norm.rs

joseph-isaacs · 2026-03-17T09:55:34Z

vortex-tensor/src/scalar_fns/l2_norm.rs

+    fn is_fallible(&self, _options: &Self::Options) -> bool {
+        // Canonicalization of the storage array can fail.
+        true


This is not the meaning of this

Its asks if the operation can fail liked checked_add (overflow). Can this happen here

/// Returns whether this expression itself is fallible. Conservatively default to *true*. /// /// An expression is runtime fallible is there is an input set that causes the expression to /// panic or return an error, for example checked_add is fallible if there is overflow. /// /// Note: this is only applicable to expressions that pass type-checking /// [`ScalarFnVTable::return_dtype`].

In that case this doc comment needs to be more detailed since that was not clear. What is a better description here? How do we distinguish between an execution failure vs a logical failure when we get a VortexResult back regardless?

codspeed-hq · 2026-03-18T19:18:48Z

Merging this PR will not alter performance

✅ 1009 untouched benchmarks
⏩ 1515 skipped benchmarks¹

_{Comparing ct/ext-vec (fe41f12) with develop (683ba3a)}

1515 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

vortex-tensor/src/scalar_fns/cosine_similarity.rs

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

AdamGS · 2026-03-19T13:33:03Z

vortex-tensor/src/vector/vtable.rs

+    // TODO(connor): This is just a placeholder for now.
+    type NativeValue<'a> = &'a ScalarValue;


It's mainly because we are blocked on #6717 and have to figure some things out on that first

In reality we don't actually care about this now, but once we start using these extension types in the compressor then we do care because we would want a more efficient representation of a vector when (for example) encoded as a ConstantArray

AdamGS · 2026-03-19T13:34:31Z

vortex-tensor/src/scalar_fns/cosine_similarity.rs

        )?;

-        // Row 0: identical → 1.0, row 1: orthogonal → 0.0.
+        // Row 0: identical -> 1.0, row 1: orthogonal -> 0.0.


AdamGS · 2026-03-19T13:36:47Z

vortex-tensor/src/scalar_fns/utils.rs

@@ -0,0 +1,236 @@
+// SPDX-License-Identifier: Apache-2.0


just a personal taste thing - but these hyper specialized utils just crate fragmentation in the codebase, its more layers and they only exist in this corner

Not really sure where else to put this though? We need it for both of the scalar fns in here (and will need it for more in the future)

you can just inline them? they seem to be mostly very short and easy to quickly read through, even if it introduces some code duplication

Hmmm I'm not sure I agree with this, especially since that would mean duplicating the FlatElements struct I introduced which feels wrong.

AdamGS · 2026-03-19T13:41:24Z

vortex-tensor/src/scalar_fns/cosine_similarity.rs

-/// For [`FixedShapeTensor`], computes `dot(a, b) / (||a|| * ||b||)` over the flat backing buffer of
-/// each tensor. The shape and permutation do not affect the result because cosine similarity only
-/// depends on the element values, not their logical arrangement.
+/// Computes `dot(a, b) / (||a|| * ||b||)` over the flat backing buffer of each tensor or vector.


Not part of this PR, but I think we also want to do expr(array, scalar) for this sort of thing

yeah that is also part of why we want #6717

AdamGS

This seems very reasonable to me, some pieces depend on future work, like Variant we probably want a standard way to express these sort of stability/maturity guarantees

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

danking · 2026-03-19T13:44:52Z

vortex-tensor/src/fixed_shape/vtable.rs


    fn id(&self) -> ExtId {
-        ExtId::new_ref("vortex.fixed_shape_tensor")
+        ExtId::new_ref("vortex.tensor.fixed_shape_tensor")


Are we certain no one has written a vortex.fixed_shape_tensor?

We're pretty sure, this is a very recent addition

danking · 2026-03-19T13:48:13Z

vortex-tensor/src/scalar_fns/l2_norm.rs

+                .map(|i| l2_norm_row(flat.row::<T>(i)))
+                .collect();
+
+            Ok(result.into_array())


Do we intend to eventually replace this with a call to BLAS? It seems like extract_flat_elements produces a BLAS compatible matrix.

what is BLAS?

https://www.netlib.org/blas/

Basic Linear Algebra Subprograms. It's the way to do fast linear algebra. There's also LAPACK which mostly deals with matrix transformations e.g. QR factorization.

There's many implementations of the BLAS interface. Perhaps the most interesting is GotoBLAS which this guy hand wrote in assembly. If you're on Intel processors you really want to link against the Intel Matrix Kernel Library (MKL). There's also OpenBLAS which is arch independent and cuBLAS which is for GPUs. Generally, one expects these libraries to be installed on the machine and you dynamically link against them.

is there a standard rust crate that does this? If we can pull that in and it operates over flat elements then that is a trivial thing to add. I looked online and found a few things but am not familiar with this space

vortex-tensor/src/scalar_fns/utils.rs

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

connortsui20 added the changelog/feature A new feature label Mar 13, 2026

connortsui20 force-pushed the ct/ext-vec branch 2 times, most recently from 33409b3 to 5a198e8 Compare March 16, 2026 15:01

connortsui20 marked this pull request as ready for review March 16, 2026 15:01

connortsui20 requested a review from gatesn March 16, 2026 15:01

connortsui20 mentioned this pull request Mar 16, 2026

Tracking Issue: Tensor-related Extension Types #6865

Open

13 tasks

gatesn reviewed Mar 16, 2026

View reviewed changes

vortex-tensor/src/scalar_fns/cosine_similarity.rs Outdated Show resolved Hide resolved

gatesn reviewed Mar 16, 2026

View reviewed changes

vortex-tensor/src/scalar_fns/l2_norm.rs Outdated Show resolved Hide resolved

joseph-isaacs reviewed Mar 17, 2026

View reviewed changes

connortsui20 requested a review from gatesn March 17, 2026 11:39

connortsui20 force-pushed the ct/ext-vec branch 4 times, most recently from 615b78d to 2da0149 Compare March 18, 2026 19:58

connortsui20 requested a review from joseph-isaacs March 18, 2026 21:11

AdamGS reviewed Mar 19, 2026

View reviewed changes

vortex-tensor/src/scalar_fns/cosine_similarity.rs Outdated Show resolved Hide resolved

connortsui20 added 8 commits March 19, 2026 09:30

vector type first draft

903eab9

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

add AnyTensor matcher and impl cosine similarity for vector

a2ec5b4

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

add l2 norm scalar fn

b47bd61

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

clean up

e86b0db

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

lockfile

d20e54c

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

clean up

2ce76b5

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

address comments

c25e331

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

make not fallible

fcf3e67

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

connortsui20 force-pushed the ct/ext-vec branch from 2da0149 to 01ea8b1 Compare March 19, 2026 13:31

AdamGS reviewed Mar 19, 2026

View reviewed changes

AdamGS approved these changes Mar 19, 2026

View reviewed changes

clean up

30c4258

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

connortsui20 force-pushed the ct/ext-vec branch from 01ea8b1 to 30c4258 Compare March 19, 2026 13:44

danking reviewed Mar 19, 2026

View reviewed changes

slice better

fe41f12

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>

connortsui20 merged commit 8efe1dc into develop Mar 19, 2026
57 checks passed

connortsui20 deleted the ct/ext-vec branch March 19, 2026 14:13

		// TODO(connor): This is just a placeholder for now.
		type NativeValue<'a> = &'a ScalarValue;

Conversation

connortsui20 commented Mar 13, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

API Changes

Testing

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

joseph-isaacs Mar 17, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codspeed-hq bot commented Mar 18, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Merging this PR will not alter performance

Footnotes

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

AdamGS left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

connortsui20 commented Mar 13, 2026 •

edited

Loading

joseph-isaacs Mar 17, 2026 •

edited

Loading

codspeed-hq bot commented Mar 18, 2026 •

edited

Loading