The other memcpys C into a local array and then accumulates onto that before memcpying it back at the end of the kernel. The latter is thus 100% equivalent to the std::simd kernel, except that the compiler needs to turn the innermost loop into the SIMD FMA that std::simd encodes directly.

This is ~3–4x slower.

TBH, I expected less of a difference.
But anyway, if you want to express data-parallelism don't write a loop, use std::simd. It helps.

2/2

#stdsimd

Two more results. This time without using std::simd. One uses a plain loop over C[i, j] += A[i, k] * B[k, j] (in the inner kernel—it is still blocked over all levels of the cache hierarchy).

This is ~10–30x slower.

1/2

#stdsimd #cpp26 #simd

mdspan rocks! A simple switch to go from layout_right to layout_right_padded and performance for larger matrices goes 📈 up! (e.g. 4096×4096 from 76GFLOP/s to 100GFLOP/s) I introduced padding of one cache line between rows to avoid cache associativity virtually reducing cache sizes.
For small matrices the extra padding is counterproductive, though. But mdspan abstracts it all away. The matrix-mul function is unchanged.

#stdsimd #mdspan #cpp26 #optimization