QR Factorization Discrepancy
- by KyleSum
I'm trying to get a feel for the Intel MKL library with a simple back-solve (A*x = b) using a QR factorization and comparing my MKL answer to the answer of a known working solution. When my answers didn't come up correct I printed a diff between the Q and R matrices of the known working and the MKL test code.
I know MKL/lapack uses "elementary reflectors" to store the values of both the Q and R matrices. So, I'm wondering if these differences (mostly +/-) are by design or the result of some bug.
I'm using DGEQRF, DORMQR, and DTRSM routines to solve the system and DORGQR (for debugging) to get the Q matrix shown in the diff.
diff with 6x6 matrix (top known, bottom mkl): http://pastebin.com/4uwcME0J