Already wrote on stackoverflow for help to get atleast the multiply256 function improved:
https://stackoverflow.com/questions/67667314/transform-native-c-matrix-multiplication-to-opencl-simd-matrix-multiplication
https://stackoverflow.com/questions/67667314/transform-native-c-matrix-multiplication-to-opencl-simd-matrix-multiplication
At least your question didn't get closed for "duplicate"
But I would change the matrix multiplication algorithm to O(n^2.
Strassen which is faster than standard O(n^3) matmul. Changes in the algorithm used generally make a much more noticeable performance impact.