cuda-data-operators

GPU-native data operators for SQL-like operations directly on GPU memory.

Quick Start (Google Colab)

!git clone https://github.com/xsfa/cuda-data-operators.git
%cd cuda-data-operators
!pip install --quiet cupy-cuda12x
!python test_runner.py --setup
!python test_runner.py

Requirements: GPU runtime (Runtime → Change runtime type → T4 GPU)

Local Development (with uv)

uv pip install -e ".[gpu]"
uv run python test_runner.py --setup
uv run python test_runner.py

Operators

Operator	Status	Description
Filter	✅	Predicate evaluation + stream compaction
SUM	✅	Parallel reduction
COUNT	✅	Parallel reduction
MIN/MAX	✅	Parallel reduction
GROUP BY	🔄	Hash-based grouping
Hash Join	🔄	Build + probe
Sort	🔄	Radix sort

Project Structure

src/
├── memory_pool.cuh      # Arena allocator for GPU memory
├── column.cuh           # Typed columnar arrays
├── primitives/
│   └── prefix_scan.cuh  # Blelloch scan algorithm
└── operators/
    ├── filter.cuh       # Stream compaction
    └── aggregate.cuh    # SUM, COUNT, MIN, MAX

Running Individual Tests

python test_runner.py --list
python test_runner.py --test filter
python test_runner.py --test sum_large