John Osorio

Hi @elaith9 I was reading your posts, and a lot of posts. Please if someone already solved this problem please help me. There are very interesting. But I'm having a...

@elaith9 thanks for your answer. @dqthebt24 I was trying with detectnet but I suppose that in fact I didn't use enough epochs. Im going to test again.

We have updated the PR. We follow the equation to apply the destination scaling after the post-operations. As `cudnnConvolutionForward` does not support having its output in `s32` format, we follow...

> There is another class of test cases that fail: no dst scale, but dst is s8: ./build/tests/benchdnn/benchdnn --conv --engine=gpu --skip-impl=ref --dir=FWD_I --dt=s8:s8:s8 --attr-scales=src:common:2 --attr-post-ops=sum mb1ic512iw121oc512ow122kw6pw3nconv1d:21 > > Even dst...

@dzarukin I did the squash as you mentioned. Thank you very much. Let me know if something else needs to be done.

@dzarukin We take into account what was discussed during the meeting. The PR now avoids applying pooling src:dst (u8:u8, s8:s8).

> > @dzarukin We take into account what was discussed during the meeting. The PR now avoids applying pooling src:dst (u8:u8, s8:s8). > > Could you clarify what issue does...

> > @mgouicem We received feedback mentioning that we need to avoid the use of `u8:u8` and `s8:s8`. Maybe @dzarukin have any additional clue about it. If this combination is...

> > > > @mgouicem We received feedback mentioning that we need to avoid the use of `u8:u8` and `s8:s8`. Maybe @dzarukin have any additional clue about it. If this...

To give a little bit of additional context. I found that in this [line](https://github.com/oneapi-src/oneDNN/blob/08fd0eee29cc523ae0f4905c148513c43e7e948d/src/gpu/generic/sycl/layer_normalizations_kernels.hpp#L218). The division is making **v_variance** slightly different from the reference that is computed in benchdnn (there...