Author: Joungmin Lee Description: If an alternative cblas library built with march optimization is used, test fails in LU decomposition with values slightly higher than the tolerance in test_luc.c. When I looked into the test code, it appears to have some faults when multiplying a coefficient with the matrix size, even though the code pass the test with its internal cblas. This patch slightly relieves tolerances in the test function argument, by multiplying max(size1, size2) instead of faulty n (being 51 after iteration above, absolutely irrelevant to the rect size). This issue must be discussed by the upstream authors. --- a/linalg/test_luc.c 2021-12-01 14:43:39.000000000 +0900 +++ b/linalg/test_luc.c 2026-06-10 15:02:39.969239165 +0900 @@ -166,21 +166,21 @@ { gsl_matrix_complex * m = gsl_matrix_complex_alloc(100, 50); create_random_complex_matrix(m, r); - test_LUc_decomp_eps(m, 256.0 * n * GSL_DBL_EPSILON, "complex_LU_decomp rect1"); + test_LUc_decomp_eps(m, 256.0 * 100 * GSL_DBL_EPSILON, "complex_LU_decomp rect1"); gsl_matrix_complex_free(m); } { gsl_matrix_complex * m = gsl_matrix_complex_alloc(50, 100); create_random_complex_matrix(m, r); - test_LUc_decomp_eps(m, 1.0e3 * n * GSL_DBL_EPSILON, "complex_LU_decomp rect2"); + test_LUc_decomp_eps(m, 1.0e3 * 100 * GSL_DBL_EPSILON, "complex_LU_decomp rect2"); gsl_matrix_complex_free(m); } { gsl_matrix_complex * m = gsl_matrix_complex_alloc(80, 100); create_random_complex_matrix(m, r); - test_LUc_decomp_eps(m, 1.0e3 * n * GSL_DBL_EPSILON, "complex_LU_decomp rect3"); + test_LUc_decomp_eps(m, 1.0e3 * 100 * GSL_DBL_EPSILON, "complex_LU_decomp rect3"); gsl_matrix_complex_free(m); }