I think came upon a bug in the mkl_?omatadd routine, especially the complex versions. If both matrices A and B are conjugate transposed, the matrix A is only conjugated. I configured the zomatadd.c example file that intel provided in order to show that this is the case:
As you can see the results are the same, which is incorrect. The second computation gives the wrong answer.
There are actually two more cases I found, where the result is wrong:
- A transposed and B conjugated: This combination conjugate transposes A instead of just transposing it.
- A conjugated and B transposed: This combination conjugate transposes B instead of just transposing it.
The example file is also attached for you to test it.
The issue is closing and we will no longer respond to this thread. If you require additional assistance from Intel, please start a new thread. Any further interaction in this thread will be considered community only.