Skip to content

stest_rfp: STFSM test is very sensitive to small numeric errors in BLAS STRSM  #679

Open
@martin-frbg

Description

@martin-frbg

As originally reported in OpenMathLib/OpenBLAS#3648 , the test for STFSM in sdrvrf3.f compares against the result of STRSM with a fairly small default threshold. While this happens to work with the reference implementation of the BLAS, spurious failure may occur with a user-supplied BLAS (as observed with OpenBLAS) although the deviation is within the expectations for a single precision calculation making use of FMA or similar hardware features.
While this situation is more or less described in the FAQ as a "minor testing failure" (http://www.netlib.org/lapack/faq.html#_how_do_i_interpret_lapack_testing_failures) it may be causing unnecessary concern for distribution packagers etc. Suggested options would be to compare against DTRSM, or trivially to increase the threshold. The same (non?)issue very likely exists with ctest_rfp/CTFSM

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions