Dear BLAS developers,
there seems to be a bug in the gemv routines for the special
case where n=0 and beta /not= 1, i.e.,
call dgemv(TRANS,M, 0, ALPHA, A, LDA, X, INCX, 0.5, Y, INCY)
leaves Y unchanged but according to the manpage n=0 is allowed
and this operation should scale Y.
This bug is present in all BLAS implementations I have tested, including
NETLIB-reference, AMD-ACML, Intel-MKL, GOTO, GOTO2, ATLAS,
SUN-performance-lib.
The gbmv-routines seem also to be affected by this bug.
Level 3 routines (at least gemm-routines) handle this situation
correctly.
Best regards,
Dennis Diehl
|