Highlights
This release substantially optimizes the parity-focused CPU engine without changing the public workflow.
- Groups pathways by size and scores them through batched, parallel cumulative-ES kernels.
- Reduces temporary allocations and reuses sampling state in permutation-heavy execution.
- Accelerates ordering and comparison operations while retaining exact overflow-safe fallbacks.
- Adds realistic Criterion benchmarks, heavy-workload profiles, and thread-scaling controls.
- Expands real-data performance and memory validation.
On the documented release benchmark, single-worker Rust execution was approximately 3× faster in simple mode and 4.1× faster in multilevel mode than R fgsea. The committed real-data validation used approximately 4.1× less peak memory.
CPU parity remained at floating-point-noise scale.
Full changelog: v0.3.3...v0.3.4