Skip to content

LAPACK testing: match eigenvalues by value in xCHKHS test 11 (Reference-LAPACK PR 1491) - #6257

Open
tesch1 wants to merge 1 commit into
OpenMathLib:developfrom
tesch1:chkhs-eig-order
Open

tesch1 wants to merge 1 commit into
OpenMathLib:developfrom
tesch1:chkhs-eig-order

Conversation

@tesch1

@tesch1 tesch1 commented Nov 14, 2026

Copy link
Copy Markdown
Contributor

Since #6158, make lapack-test with TARGET=VORTEXM4 reports one COMPLEX16 error:

 Matrix order=   16, type=18, seed=2120, 768,2414,2241, result  11 is 3.804D+07
 ZHS:    1 out of  2142 tests failed to pass the threshold

This is a false failure in the test, not in the kernels. Test 11 of xCHKHS (from Reference-LAPACK PR 1410) compares the eigenvalues of ZHSEQR with JOB = 'S' and JOB = 'E' by index. For this matrix the two lists hold the same 16 eigenvalues (to about 11 ulp), but eigenvalues 4 and 5 are in opposite order; the printed result is the distance between those two eigenvalues. All ZGEMM calls in the test are correct to rounding; the new kernels only round differently from reference BLAS. The same kind of swap, for well-conditioned eigenvalues (s about 0.5), gives the type 20 failure (result 11 is 1.363D+15) that a VORTEX build can show.

The fix compares each W1(j) with the nearest W3(i), in all four drivers. It is the same change as Reference-LAPACK/lapack#1491, which has the details and the eigenvalue data. This PR can also wait for that one to be merged upstream.

Testing on Apple M4 (TARGET=VORTEXM4, current develop):

  • make lapack-test: 0 errors in all precisions (1 COMPLEX16 error before).
  • nep.in for s, d, c and z: all pass.

Test 11 (from Reference-LAPACK PR 1410) compares the eigenvalues of
xHSEQR with JOB = 'S' and JOB = 'E' by index. The two paths can return
the eigenvalues in a different order, so a swap of two eigenvalues
fails the test although the two lists agree. Compare each W1(j) with
the nearest W3(i) instead.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant