If you hack METHOD in atlas_qrmeth.h to always do either PCA copy or no copy, and hardwire nb=4 in ATL_geqrf, then N>10 will get the wrong answer. If tgeqr2 is not called, answer is always correct.
Not sure if it is related, but xsqrtst_pt also fails sanity checks on 32-bit opt8.
G5 and G4 xsqrtst_pt sporadically fail as well, but this is almost certainly unrelated since it only occasionally happens.
Both problems caused by out-of-order writes (ARM and PPC both do this). Problem fixed by turning off PCA for these archs.
Log in to post a comment.
Sign up for the SourceForge newsletter:
You seem to have CSS turned off.
Please don't fill out this field.