#ifndef _matmul_n_
#define _matmul_n_


#include "./utils/matrix.h"


namespace numerics{
	
	template <typename T>
	utils::Matrix<T> matmul(const utils::Matrix<T>& A, const utils::Matrix<T>& B){

		if(A.cols() != B.rows()){
			throw std::runtime_error("matmul: dimension mismatch");
		}

		const uint64_t m = A.rows();
		const uint64_t n = A.cols(); // also B.rows()
		const uint64_t p = B.cols();
		T tmp;

		utils::Matrix<T> C(m, n, T{0});

		//#pragma omp parallel for collapse(2) schedule(static)
		#pragma omp parallel for
		for (uint64_t i = 0; i < m; ++i){
			for (uint64_t j = 0; j < n; ++j){
				tmp = A(i,j);
				for (uint64_t k = 0; k < p; ++k){
					C(i,k) += tmp * B(j,k);
				}
			}
		}
		return C;
	}




} // namespace numerics

#endif // _matmul_n_