Page 1 of 1

xLARTG vs LAWN 148 discrepancy for F=0 and (G .ne. 0)

PostPosted: Tue Apr 25, 2017 11:54 pm
by jip
When F=0 and (G .ne. 0), the LAWN 148 proposes to set CS=0 and SN=sign(conj(G)) (Algorithm 1 on page 5). It differs from implementation in SLARTG and DLARTG which both are set CS=0 and SN=1 .

Just curious, is there any paper which substantiate LAPACK's choice in this case? Or the reason is the phrase in the source code: "without doing any floating point operations (saves work in SBDSQR when there are zeros on the diagonal)"? But what about continuity in this case? More common question is, how much LAWNs are mandatory for LAPACK development.