Any theoretical grounds?? BLAS testers

[keitat]

I am very curious with legacy level 1 BLAS testers, in particular DOT product. According to the reference BLAS tester source code dblat1.f, the difference between DDOT and Oracle (correct number) is tested to a threshold (SFAC) which is hard coded to 9.765625D-4. Is there any theoretical ground? What if a user wants to test with bigger vectors (say N=10000). Since we use BLAS with different (heterogeneous) platforms and runtimes, it will be great to have testers for today's computing systems and common usage.

[langou]

Hi Keita,

0.0009765625 = 1 / 1024 = 2^(-10)

SFAC is used in the test at https://github.com/Reference-LAPACK/lapack/blob/7d90a679e44d0d2add75697923a59b2c15435011/BLAS/TESTING/cblat1.f#L643

Since we use BLAS with different (heterogeneous) platforms and runtimes, it will be great to have testers for today's computing systems and common usage.

Granted. We are working in newer, more scalable tester for BLAS. I do not think we can revised the test suite currently in the LAPACK for the BLAS to work for what you are mentioning.

Cheers, Julien.