Concepts

These pages explain the cross-cutting ideas that show up across the GLASS API: how the glass::nvidia:: wrappers decide between cuBLASDx and pure-SIMT, what the TRAILING_SYNC template parameter controls, how to autotune the dispatch for your own hardware, the batched-1D GEMM APIs designed for kernels with a single 1D thread block, the block-tridiagonal layout used by the glass::bdmv / glass::pcg solvers, and the conventions behind the robotics-specialized operator families.

• Backend Dispatch
• Contraction-parallel ops (the *_reduced family)
• Namespaces, suffixes, and flags
• Trailing Synchronization (TRAILING_SYNC)
• Tuning for Your Hardware
• Batching models
• Batched-1D GEMM APIs
• Block-tridiagonal Solves
• Robotics operators & conventions
• Testing, oracles & receipts