DGEEV change

[riclarsson]

Hi,

I reran my CI tests this week. The CI tests a code that does eigenvalue diagonalization. I have two solutions to this problem. A custom dgeev-like solver (for purely real eigenvalues) and one calling dgeev. The former works. The latter fails after the latest update to OpenBLAS, and I am fairly certain that the update is the cause of the failure because what's mentioned is the only difference between the two tests.

The code calling dgeev now fails in one of my tests but the code calling custom code clears it.

This brings me here with questions:

1. Was there any serious optimization work done that would affect the real-eigenvalues only path of dgeev in the latest release?
2. What is the numerical drift you had for dgeev this release?

(If it matters, in the tests I am comparing model result output with externally computed values. A relative difference of less than 1e-6 is considered fine. The difference was never zero because the external solver is from python using lots of unoptimized but easy to maintain and clear to understand code, so it is likely a tipping point in numerical drift that I am seeing. But I am seeing it.)

[riclarsson]

Sorry, the packaging had changed. CMake no longer found your project it seems. It is fixed on our side. The error was caused by finding libblas instead of your code. Again, sorry, I don't have access to a Linux machine and as the error was in dgeev and you had an update, I expected that that was the cause (it was, but not for the reasons I thought).