kobold-cuda

GPU/CUDA-accelerated batch decoding of COBOL numeric fields (COMP-3 packed-decimal and zoned-decimal DISPLAY) for high-throughput legacy-data ingest -- the GPU counterpart to the CPU/SIMD kernels in kobold-simd.

Part of KOBOLD -- independently-authored forensic/throughput tooling for legacy COBOL estates. Contains no GnuCOBOL source. Apache-2.0.

Architecture

• gnucobol-rs (separate crate) = the oracle-proven semantic primitive layer.
• kobold-* = the forensic/throughput layer; kobold-* MAY depend on gnucobol-rs, never the reverse.

Status

Scaffold -- the name is parked. The CUDA kernels (one record per thread over a fixed stride) and the host bindings land later; a portable scalar fallback keeps the crate usable in the meantime.

License

Apache-2.0 (see LICENSE).