Expand description
Rust/CUDA acceleration primitives for spiking neural networks (SNN) and
neuromorphic workloads: safe Rust wrappers around first-party CUDA kernels
covering spiking-network dynamics, routing, SAT search, and packed ternary
GEMV/GEMM. The cuda feature targets sm_120 (NVIDIA Blackwell); the
default build is CPU-only and needs no CUDA toolkit or device.
This is a low-level compute layer, not a complete SNN framework or a general model-compatibility layer.
Re-exports§
pub use gpu_stub as gpu;pub use capability::Backend;pub use capability::CapabilityFacts;pub use capability::CapabilityReport;pub use capability::ComputeCapability;pub use capability::ExecutionPolicy;pub use capability::FallbackReason;pub use capability::FallbackRecord;pub use capability::KernelAvailability;pub use capability::evaluate_capabilities;pub use capability::probe_capabilities;pub use capability::sanitize_diagnostic;pub use gpu_stub::GpuAccelerator;pub use gpu_stub::GpuBuffer;pub use gpu_stub::GpuContext;pub use gpu_stub::GpuError;pub use gpu_stub::KernelModule;
Modules§
- bench
- Reproducible benchmark manifests and opt-in regression budgets.
- bitpacking
- Binary and ternary bitpacking utilities.
- capability
- Structured CUDA capability probes and fallback reason codes.
- gpu_
stub - oracle
- Independent scalar CPU oracles for the public CUDA kernel paths.