void optimized_dtrsm(const enum CBLAS_ORDER Order, const enum CBLAS_SIDE Side,
         const enum CBLAS_UPLO Uplo, const enum CBLAS_TRANSPOSE TransA,
         const enum CBLAS_DIAG Diag, const int M, const int N,
         const double alpha, const double *A, const int lda,
         double *B, const int ldb)
{
    int i, j, k;
    register double aux;
    register double *pbij, *pbkj;

    assert(Side == CblasLeft);
    assert(TransA == CblasNoTrans);
    assert(Uplo == CblasUpper);
    assert(Diag == CblasUnit);

    for (k = M - 1; k >= 0; --k) {
        for (i = 0; i < k; ++i) {
            aux = A[i * M + k];
            pbij = B + i * N;
            pbkj = B + k * N;
            for (j = 0; j < N; ++j) {
                *pbij++ -= *pbkj++ * aux;
            }
        }
    }

    for (i = 0; i < M; ++i) {
        pbij = B + i * N;
        for (j = i; j < N; ++j) {
            *pbij++ *= alpha;
        }
    }
}