Helix StudioGenome IDE
← Docs
Helix CLI docs
Browse Helix CLI docs

CUDA Backend Plan

The CRISPR engine surface is frozen as (G, L) × (N, L) → (G, N). The first CUDA milestone should:

  1. Implement a score_pairs_encoded kernel in CUDA (guides[G,L], windows[N,L], scores[G,N]).
  2. Mirror the existing C++ CPU semantics exactly (same encoding, penalties, scoring version).
  3. Expose it via backend="gpu" in create_crispr_physics.
  4. Validate against the golden fixtures (tests/data/engine_crispr/).
  5. Benchmark with benchmarks/engine_crispr_perf.py --backend gpu ....

Keep GPU-specific padding or alignment hidden inside the backend; Python and the simulator should never see it.

Building with CUDA

src/helix_engine/CMakeLists.txt exposes HELIX_ENGINE_ENABLE_CUDA=ON. When set, CMake enables the CUDA language and compiles src/physics_cuda.cu (currently just a stub). Example:

cmake -S src/helix_engine -B build/helix_engine -DHELIX_ENGINE_ENABLE_CUDA=ON

Future patches will replace the stub with a real kernel and plumb the Python backend selection through backend="gpu".