Hello. I used MKL 2024.0 in this test. I tried VBSR variant with mtype=2 with a matrix about 100'000 unknowns and default parameters. It worked fine with CSR3 format and VBSR with iparm[36]=-95. Then I specified iparm[36]=-1 just to see what happens and it also computed the solution with good relative residual 1e-13 but percentage statistics was broken:
...
1143 %
1149 %
1154 %
1156 %
100 %
However, the solver should be terminated in this case, because if it merges all rows with at least 1% similarity into supernodes, the matrix should be almost dense and solver should break with error like "non enough memory", which it did on some another matrix. The solver may actually behave the other way in this case, for example, specify block size -iparm[36] in VBSR, so this should be either documented or fixed.
I also have a question about MKL BSR3 format (iparm[36]>0). MKL 2024.2 documentation states that "Intel ® oneAPI Math Kernel Library (oneMKL) supports only the VBSR format for real and symmetric positive definite or indefinite matrices (mtype = 2 or mtype = -2).". Is that correct the user cannot use in PARDISO any symmetric BSR3 format to factorize a symmetric matrix? However, when I specify iparm[36]=1, it works fine. Do I have to convert the matrix to unsymmetric BSR3 format in order to use to use BSR3 solver even for an SPD matrix and solve with mtype=11?
