Problem with blasfeo_drowpe? about blasfeo HOT 4 CLOSED

Hi Lander,

thank you for your bug report.
Could you specify some more information about your build (e.g. TARGET, LA, MF), such that I can identify the issue?

I just tried out with Haswell target, high-performance LA and panel major matrix format, and the output of the permutation you propose is correctly 0 2 1.
Notice that kmax should be set to 3 in this case (i.e. 3 permutations are applied), if you set it to 2 you would get the output of the first two permutations, i.e. 1 2 0.

from blasfeo.

Hello Gianluca,

Thank you for the feedback. To avoid confusion: the output I obtain is 0 2 1 as well, so the same as you. I however expect to get 1 2 0 as the correct result. If 0 2 1 is the expected output, could you clarify what the function is exactly supposed to do?

Thank you,
Lander

from blasfeo.

Hi Lander,

the routine uses the encoding of a permutation vector ipiv that is found e.g. in the standard LAPACK routine dgetrf for LU factorization (with the only difference that in LAPACK the indexes start from 1, while in BLASFEO I preferred to use the C standard array indexing of starting from 0, so it is exactly the standard LAPACK encoding minus one).

ipiv = [1, 2, 0] means (if applied to a vector):

• swap element 0 of the original vector with the element at ipiv[0]=1
• afterwards, swap element 1 of the output vector from the previous step with the element at ipiv[1]=2
• afterwards, swap element 2 of the output vector from the previous step with the element at ipiv[2]=0

from blasfeo.

Dear Gianluca,

Thanks a lot for the explanation. I misunderstood the purpose of the function.

Kind regards,
Lander

from blasfeo.