vortex-data/vortex - CodSpeed

vortex

Blog Docs Changelog

Performance History

Latest Results

feat[gpu]: multi-stage dynamic dispatch with plan builder Evolve the dynamic dispatch kernel from a single-pipeline design (one source op + scalar ops) to a multi-stage plan. Input stages decode child arrays (dictionary values, run-end endpoints) into shared memory; the output stage references them and writes directly to global memory. Add `build_plan`, which walks a Vortex encoding tree and automatically produces a `DynamicDispatchPlan`. Supported encodings: BitPacked, FoR, ZigZag, ALP (f32), Dict, RunEnd, and Primitive. Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

ad/dyn-dispatch

5 hours ago

feat[gpu]: multi-stage dynamic dispatch with plan builder Evolve the dynamic dispatch kernel from a single-pipeline design (one source op + scalar ops) to a multi-stage plan. Input stages decode child arrays (dictionary values, run-end endpoints) into shared memory; the output stage references them and writes directly to global memory. Add `build_plan`, which walks a Vortex encoding tree and automatically produces a `DynamicDispatchPlan`. Supported encodings: BitPacked, FoR, ZigZag, ALP (f32), Dict, RunEnd, and Primitive. Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

ad/dyn-dispatch

5 hours ago

feat[gpu]: multi-stage dynamic dispatch with plan builder Evolve the dynamic dispatch kernel from a single-pipeline design (one source op + scalar ops) to a multi-stage plan. Input stages decode child arrays (dictionary values, run-end endpoints) into shared memory; the output stage references them and writes directly to global memory. Add `build_plan`, which walks a Vortex encoding tree and automatically produces a `DynamicDispatchPlan`. Supported encodings: BitPacked, FoR, ZigZag, ALP (f32), Dict, RunEnd, and Primitive. Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

ad/dyn-dispatch

6 hours ago

feat[gpu]: multi-stage dynamic dispatch with plan builder Evolve the dynamic dispatch kernel from a single-pipeline design (one source op + scalar ops) to a multi-stage plan. Input stages decode child arrays (dictionary values, run-end endpoints) into shared memory; the output stage references them and writes directly to global memory. Add `build_plan`, which walks a Vortex encoding tree and automatically produces a `DynamicDispatchPlan`. Supported encodings: BitPacked, FoR, ZigZag, ALP (f32), Dict, RunEnd, and Primitive. Signed-off-by: Alexander Droste <alexander.droste@protonmail.com>

ad/dyn-dispatch

6 hours ago

only a read lock for current_array_ref Signed-off-by: Daniel King <dan@spiraldb.com>

dk/buffer-pool

1 day ago

only a read lock for current_array_ref (#6631) ## Summary SharedArray uses a Mutex as of #6453. This interacts particularly badly with `scalar_at` by way of `is_valid`. In particular, all calls to `scalar_at` on the given array are serialized. It seems reasonable to allow multiple threads to simultaneously read the validity. This PR permits that by using an RwLock. ## Testing I have a Vortex-based system which recently pulled in a version of Vortex that includes #6453. With that version, my profile looks like this: https://share.firefox.dev/4tO94Ki <img width="1920" height="1080" alt="Screenshot 2026-02-20 at 4 26 28 PM" src="https://github.com/user-attachments/assets/77f24c44-268e-4dd8-8bc7-ec2c193443b6" /> With this PR applied (i.e. with an RwLock in place of the Mutex), it looks like this: https://share.firefox.dev/4aNPZzt <img width="1920" height="1080" alt="Screenshot 2026-02-20 at 4 27 25 PM" src="https://github.com/user-attachments/assets/dcf2fde3-e35a-413b-a6e6-2b4053e05653" /> CPU utilization is high, and runtime is ~7s. By comparison, with a Mutex, CPU utilization is low and runtime is ~18s. Signed-off-by: Daniel King <dan@spiraldb.com>

develop

1 day ago

GlobalData should not be passed as exclusive reference to thread-local functions Signed-off-by: Nicholas Gates <nick@nickgates.com>

ngates/duckdb-global-mut

1 day ago

chore(deps): update dependency ray to v2.54.0 [security]

renovate/pypi-ray-vulnerability

1 day ago

Active Branches

feat[gpu]: multi-stage dynamic dispatch with plan builder

last run

5 hours ago

0%

last run

1 day ago

-65%

Convert &mut to & for DuckDB GlobalData

last run

1 day ago

-27%

© 2026 CodSpeed Technology

Home Terms Privacy Docs