CoinAbcHelperFunctions
SIMD-optimized scatter/gather operations for ABC factorization
SIMD-optimized scatter/gather operations for ABC factorization
High-performance kernels for sparse matrix operations using:
- Manual loop unrolling (UNROLL_SCATTER, UNROLL_GATHER)
- AVX/AVX2 intrinsics when available
- Cilk parallel loops with configurable grainsize
- Prefetching hints (coin_prefetch macros)
Key operations:
- CoinAbcScatterUpdate: region[index[j]] -= value[j] * pivot
- CoinAbcGatherUpdate: dot product over sparse indices
- Memory utilities: CoinAbcMemcpyLong, CoinAbcMemset0Long
Includes specialized ScatterUpdateN functions for N=1..8 and 4N variants using function pointer dispatch (scatterStruct) for optimal performance.
Algorithm
Function Pointer Dispatch (scatterStruct): Precomputed table of specialized ScatterUpdate{1..8,4N} functions avoids branch misprediction in small-column cases.
Complexity: $O(nnz)$ per scatter/gather operation With SIMD: ~4x speedup on vectorizable sections With Cilk: ~Kx speedup on K cores for large columns
See Also
- CoinAbcFactorization which uses these kernels
- AbcSimplex which uses these for FTRAN/BTRAN
Source
Header file: `src/CoinAbcHelperFunctions.hpp`