Unexpected Transpose Offset Unit-Diagonal

[Bryan Flynt - NOAA Affiliate]

I discovered an unexpected behavior when I take the Transpose of an Offset Unit-Diagonal.  The Non-Unit variation performs as expected but when switched to Unit Diagonal the same operation leaves out the Transpose operation on the diagonal.  I narrowed it down to a simple "bli_daddm" example which I've attached (C++) code but I get the same behavior on all the diagonal operations (bli_?addd, bli_?axpyd, etc.).

Is this the expected behavior ???

If so can you explain how the "diaga" argument is applied?  Thanks.

Non-Unit Behavior (Expected)

 | 2  2  0  0 |^T     | 3  3  3  3 |    | 5  5  5  5 |
 | 2  2  2  0 |    +  | 3  3  3  3 | =  | 5  5  5  5 |
 | 2  2  2  2 |       | 3  3  3  3 |    | 3  5  5  5 |
 | 2  2  2  2 |       | 3  3  3  3 |    | 3  3  5  5 |

Unit Behavior (Not Expected)

 | 2  1  0  0 |^T     | 3  3  3  3 |    | 5  6  5  5 |
 | 2  2  1  0 |    +  | 3  3  3  3 |  = | 3  5  6  5 |
 | 2  2  2  1 |       | 3  3  3  3 |    | 3  3  5  6 |
 | 2  2  2  2 |       | 3  3  3  3 |    | 3  3  3  5 | 

--

Bryan T. Flynt
Global Systems Division

NOAA Earth System Research Laboratory

325 Broadway R/GSD6                                   

Boulder, CO 80303
(303) 497-3875       

      

[Devin Matthews]

Dear Bryan,

I submitted this on Github as https://github.com/flame/blis/issues/468. Follow-up will happen there.

Devin Matthews

--
You received this message because you are subscribed to the Google Groups "blis-discuss" group.
To unsubscribe from this group and stop receiving emails from it, send an email to blis-discuss...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/blis-discuss/CAGu9PXp55p91XLBvEu7BtTRaJUgdJeuw7wu%2BMO_U2c_U5-XOrg%40mail.gmail.com.