2 ms·

matmul is old and useful -- there's a lot of hardware on a chip that makes it run much faster (prefetch, vectorization, instruction parallelism) and some of the

• yxhuvud 3y ago
  And some languages that perform array bounds check on each array deindex operation, which basically stops some of the above even if the optimizer can do all of it. Crystal is an example of that, where it would be possible and quite straightforward to write a specialized matrix class instead of the very generic dynamic array implementation.
• gnufx 3y ago
  Those will not simply give you close to optimal performance (>~90% of peak), even, say, POWER10's 4x4 matmul instruction. You need a structure matching the micro-architecture -- the cache and register structure [1]. That's not a triply-nested loop. Any remotely decent compiler will unroll and vectorize the micro-kernel appropriately, but you may still have to resort to assembler-level prefetch fiddling for the last 10s of percent performance (specifically on avx512). Compilers may recognize the loop structure and replace it with an optimal-ish implementation. I've an idea that FORTRAN H extended did that, but Polly does it in clang. 1. https://www.cs.utexas.edu/users/flame/pubs/blis1_toms_rev3.pdf https://www.cs.utexas.edu/users/flame/pubs/blis1_toms_rev3.p...