4.6.1. Error Handling in AOCL-BLAS - 5.2 English - 57404

The original Netlib BLAS defined an error handling function XERBLA, which is called within BLAS2 and BLAS3 routines if an incorrect input argument is detected. Only incorrect matrix, vector sizes, and options for specifying transpose matrix, upper or lower in a symmetric matrix, and so on can be detected. BLAS does not detect extreme values (such as Inf or NaNs) within the supplied matrices and vectors, it is the user’s responsibility to check for these if required.

The functionality of Netlib’s XERBLA is to print a message to standard output and stop execution of the process. Stopping is extremely unhelpful in many applications and usage scenarios. Thus, AOCL-BLAS, in common with other similar libraries, has traditionally disabled the stop statement. In AOCL 4.2, the functionality of AOCL-BLAS has been enhanced to give users more choice over both stopping the application on error and printing a message on error. In AOCL 5.0 this functionality was added to the similar cblas_xerbla error handling function. The choices are specified by setting each of the environment variables BLIS_STOP_ON_ERROR and BLIS_PRINT_ON_ERROR to 0 or 1 to respectively disable or enable the functionality. The default values for each are:

When the stop on error is disabled, no error code is passed back to the user application through the BLAS interface arguments, unlike the INFO argument used in LAPACK routines. Therefore, AOCL-BLAS has also added an extra function to return the value of INFO from the previous call to a BLAS routine made by the same thread. The function can be called as follows:

**In C/C++:**

#include <blis.h>
...
gint_t info_value = bli_info_get_info_value();

**In Fortran:**

integer :: info_value
integer, external :: bli_info_get_info_value
...
info_value = bli_info_get_info_value()

If the returned value is not zero, the value indicates the argument
in the preceding BLAS call that was incorrect.