GPU¶

A binary built with --features gpu can score on a CUDA GPU: one GPU thread per (read, reference) pair, a whole --batch-size batch per kernel launch. Only the scores come from the device. The tie set, the tracebacks, MD/NM and the records are the CPU path's own code, so --device gpu and --device cpu write the same records byte for byte.

Placement¶

Reads over 4,096 nt are scored on the CPU inside the same run, and a panel with a reference over 3,072 nt cannot use the GPU (auto then stays on the CPU and says so).

Build and runtime¶

Nothing CUDA is needed at build time, because cudarc dlopens at run time:

cargo build --release --features gpu -p eschalign-cli

At run time the GPU path needs the CUDA driver (from the node) and NVRTC (libnvrtc.so.12, which compiles the score kernel). Create the gpu pixi environment once, on a networked node:

pixi run install-gpu     # CONDA_OVERRIDE_CUDA=12 pixi install -e gpu

Then run on a GPU node:

srun -p gpu -A gpu_rbi -c 16 --gres=gpu:1 \
    pixi run -e gpu ./target/release/eschalign reads.ubam -r trna.fa -o out.bam --device gpu

On Alpine use aa100 (A100) or al40 (L40); ami100 is AMD/ROCm and cannot run the CUDA feature. See Development.

Tests¶

srun -p gpu -A gpu_rbi -c 16 --gres=gpu:1 \
    pixi run -e dev-gpu cargo nextest run --features gpu -p eschalign --test gpu_parity

Use -e dev-gpu, not -e gpu: cargo-nextest lives in the dev feature.

Performance¶

On one A30 (-c 16), the 3.3 M-read sample takes ~70 s and ~700 CPU-seconds with GPU scoring against 184 s and ~2,900 CPU-seconds without it on the same 16 cores. See Benchmarks for the arms and caveats.