Reproduce “Endeavor: Efficient PairHMM for Detection of DNA Variants in Genome-Scale Datasets” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Miguel Graça, Aleksandar Ilic
Endeavor reformulates PairHMM likelihood computation by removing the explicit deletion matrix and replacing anti-diagonal traversal with row-wise parallel processing. Implementations for multicore CPUs and NVIDIA and AMD GPUs preserve support for FP32 and FP64 arithmetic and long sequences, reaching up to 2.14× the throughput of GKL on CPUs, 2.85 TCUPS on an H100, and up to 2.05× the throughput of gpuPairHMM on reported real-data workloads.
Key insights
- Row-wise processing exposes independent work across each haplotype row, reducing the sequential depth and avoiding the uneven anti-diagonal workloads of conventional PairHMM implementations.
- Removing the deletion matrix lowers storage requirements and produces a more regular computation, but the weighted prefix operation introduces an important trade-off involving scans, register usage, and memory pressure.
- The algorithmic reformulation, rather than a device-specific optimization alone, enables portability across vendors and scales to sequence lengths up to 131,072 base pairs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.