Skip to content

Decode decimal byte parts on the GPU at the canonical width - #11460

Merged
connortsui20 merged 1 commit into
developfrom
ct/dbp-gpu-canonical-width
Nov 17, 2026
Merged

connortsui20 merged 1 commit into
developfrom
ct/dbp-gpu-canonical-width

Conversation

@connortsui20

@connortsui20 connortsui20 commented Nov 13, 2026 •

Copy link
Copy Markdown
Member

Summary

After #11459, the GPU decoder still executes a Widen MSP to its logical width and uses that buffer as the decimal storage. Above precision 18 it then returns i64 values for a decimal whose canonical width is i128 or i256, which breaks the rule that #10701 needs before every DecimalArray can have the canonical width. The decoder now reads the child of the Widen and widens it once, straight to the canonical width, which also removes the copy to the logical width.

Changes

This is its own PR because the decoder test has not run yet: it needs a GPU, so it runs for the first time in GPU CI, and a decoder bug found there then does not block #11459. The test now asserts the canonical width. The four vortex-array doc sentences that name DecimalByteParts as a source of noncanonical storage go in this PR, because they stay true until it lands.

@connortsui20 connortsui20 added the changelog/performance A performance improvement label Nov 13, 2026
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from cad5c8f to ddb0198 Compare November 13, 2026 18:12
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch 6 times, most recently from fb675af to 1ea9b28 Compare November 16, 2026 19:18
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from 1ea9b28 to 7c949a2 Compare November 16, 2026 20:45
@connortsui20
connortsui20 removed this pull request from stack #11461 November 16, 2026 20:45
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from 7c949a2 to 2f54fc6 Compare November 16, 2026 21:57
@connortsui20
connortsui20 marked this pull request as ready for review November 16, 2026 22:26
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from 2f54fc6 to 2ad83bf Compare November 17, 2026 15:04
@codspeed

codspeed Bot commented Nov 17, 2026 •

Copy link
Copy Markdown

Merging this PR will not alter performance

2 warnings

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚠️ 2 benchmarks are dominated by system calls

System calls cannot be simulated consistently and are left out of the measure. The reported value understates the real cost. Switch to the Walltime instrument to measure them.

Learn more about system calls

See affected benchmarks

✅ 2022 untouched benchmarks
⏩ 6 skipped benchmarks1


Comparing ct/dbp-gpu-canonical-width (63fa0c7) with develop (51cd7d0)

Open in CodSpeed

Footnotes

  1. 6 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. ↩

@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from 2ad83bf to beda529 Compare November 17, 2026 16:28
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from beda529 to 5a54104 Compare November 17, 2026 18:32
@connortsui20 connortsui20 mentioned this pull request Nov 17, 2026
33 of 36 tasks
Base automatically changed from ct/dbp-logical-msp-width to develop November 17, 2026 20:15
Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
connortsui20 added a commit that referenced this pull request Nov 17, 2026
## Summary

- Tracking Issue: #10701
- Replaces #11354

#10701 needs one buffer width for each decimal precision, but
DecimalByteParts decoding and `narrowed_decimal` in the compressor still
make narrower decimal arrays on `develop`.

The MSP's logical dtype now depends on the precision alone: the signed
integer at the canonical width, capped at `i64`. The canonical width is
the smallest signed width that holds `10^p - 1`, so every value of a
narrower signed integer is a valid `Decimal(p)` value, and a narrow MSP
is an integer `Widen` that needs no precision label and no scan. The
closed #10905 decoded a narrow MSP into a decimal `Widen` instead, which
needed both, and CodSpeed reported decoding 4.7x to 8.4x slower with
that design.

## Changes

The wire stores the child of the `Widen`, not a new form, so old readers
read new files and the integer `Widen` wire form can wait for #10733.

Only files from earlier releases hold an MSP that is wider than its
logical dtype. The reader gives it a checked copy on the host instead of
a cast, because a cast keeps a null value that fits the width and skips
its check when the child has exact min and max statistics. That MSP
fails to read from device memory.

`WidenArray::from_stored_integer` and `WidenArray::to_stored_integer`
live in `vortex-array` instead of private helpers, because any encoding
that fixes the dtype of an integer child needs the same two steps.

Null slots no longer choose the lower-part count or the MSP type, which
is why `p38_large_null_slots` goes from one lower part to none. The GPU
decoder still returns `i64` storage above precision 18 until #11460. For
review, read the commits one at a time, because the second one is a pure
`git mv`.

### Known costs

- `decimal_wide_p40` doubles in size, because the canonical width of
precision 40 is `i256`.
- Decoding an `i64` MSP at precision 38 gains one copy, to `i128`.
- With lower parts, compare loses its whole-answer shortcut when the
constant's MSP word is outside the range of the narrow child. A
`TODO(connor)` in `compare.rs` names the fix.
- At precision 18 or below, the Arrow `Decimal128` export of a narrow
MSP makes two copies instead of one. The follow-up is an Arrow export
hook keyed by encoding.

## API Changes

- `DecimalByteParts::try_new`, `try_new_with_lower_parts`, and
`DecimalBytePartsData::validate` require the logical MSP dtype of the
precision and at most the number of lower parts that the precision
needs.
- `split_decimal` takes the validity mask and an explicit lower-part
count.
- `narrowed_decimal` is removed.
- A file whose wider MSP holds a valid value outside the precision no
longer reads.

<details>
<summary>Validation</summary>

On the rebased stack (`develop` at `26a12a670b`):

- `cargo +nightly-2026-11-13 fmt --all --check` passes.
- `cargo clippy --all-targets --all-features -- -D warnings` passes for
`vortex-decimal-byte-parts`, `vortex-btrblocks`, `vortex-array`,
`vortex-file`, `vortex-compat`, `vortex-arrow`, and `vortex-cuda`.
- `cargo nextest run --all-features` passes for
`vortex-decimal-byte-parts`, `vortex-btrblocks`, `vortex-compat`, and
`vortex-file`, and both golden tests pass with `INSTA_UPDATE=no`.

Before the rebase (`develop` at `d81bd4ce25`), the same checks also
passed for `vortex-duckdb` and `vortex`, with `cargo bench -- --test`
for the DBP, `to_arrow`, `decimal_export`, and
`common_encoding_tree_throughput` benches. `compat.py check --mode all`
read every fixture of the 28 published versions, including the wide
legacy MSP files of 0.91.0 to 0.93.0. Of the generated fixtures, only
`decimal_byte_parts_narrow_msp.vortex` changes.

The debug assert in the v1 scheme, the `WidenArray` stored integer
functions, and the rebase onto `develop` at `c58d6237e5` came after
these runs. They only passed the format check, and no local clippy or
test run covers them yet.

Not run: the CUDA tests (no GPU), CodSpeed, and the arm64 and Windows
jobs. `cargo clippy -p vortex-btrblocks --all-targets` with default
features fails on an unused import in `vortex-btrblocks/src/tests.rs`
that is already on `develop`.

</details>

---------

Signed-off-by: Connor Tsui <connor.tsui20@gmail.com>
@connortsui20
connortsui20 force-pushed the ct/dbp-gpu-canonical-width branch from 5a54104 to 63fa0c7 Compare November 17, 2026 20:15
@connortsui20
connortsui20 merged commit 411b033 into develop Nov 17, 2026
107 of 110 checks passed
@connortsui20
connortsui20 deleted the ct/dbp-gpu-canonical-width branch November 17, 2026 20:59
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

changelog/performance A performance improvement

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants