Keep GPU FSST decompression on device - #9414

Draft

joseph-isaacs wants to merge 23 commits into

joe/gpu-decompress-on-device

Commits on Aug 3, 2026

Add a GPU Parquet decompression backend to the compression benchmark
claude
committed
37af5b0
Verify GPU Vortex decode against a separate host scan
claude
committed
5eb7a6d
Isolate the verification scans and report how a GPU decode differs
claude
committed
e7a6b41
fix(cuda): add the frame of reference to bit-packed patch values
claude
committed
335d14a
Survey every dataset in a GPU verification pass
claude
committed
5f297d4
CI: publish the GPU verification matrix to the PR
claude
committed
b0eda65
CI: install DuckDB for the GPU compression benchmark
claude
committed
66c75a5

Commits on Aug 13, 2026

Commits on Aug 14, 2026

CI: measure a CPU decompression baseline on the same GPU runner
claude
committed
b9fed6a
Keep GPU FSST decompression on device
joseph-isaacs
committed
f3bceed
Remove FSST scalar waits and support unsigned datetime parts
joseph-isaacs
committed
b9b9526
Decode nullable RunEnd arrays on CUDA
joseph-isaacs
committed
f081914
Remove FSST metadata upload callbacks
joseph-isaacs
committed
25818d7
Add single-stream CUDA benchmark fast path
joseph-isaacs
committed
3893bb7
Add reusable GPU Vortex profiling metrics
joseph-isaacs
committed
a214287
Keep GPU patch indices on device
joseph-isaacs
committed
fccded9
Avoid synchronizing fused GPU patch decodes
joseph-isaacs
committed
5b7e3a3