Renato Barros Arantes

Hi @KeijiBranshi, Yes, when I say TensorFlow 2.16 I mean tf-nightly. Thanks!

Hi @kawakami-k, Sorry for the delay in responding to you. It is similar: ``` ./test_shuffle --gtest_filter="ShuffleSimpleBF16/shuffle_test_t.TestsShuffle/4" Note: Google Test filter = ShuffleSimpleBF16/shuffle_test_t.TestsShuffle/4:-*_GPU* [==========] Running 1 test from 1 test suite....

Hi @mgouicem The root cause is that OneDNN conducts all fp16 post-op operations in fp32 and then downgrades to fp16 at the end, leading to a rounding error. To bypass...

Hi @mgouicem, > This is the expected behavior in oneDNN, and this is exactly what benchdnn checks. I just confirmed that these two tests pass on x64 implementation, which has...

Hi @mgouicem, > Could you clarify how this remove the f32->f16 downconversion? I would expect that f32->f16 conversion to be here independently of the number of post-ops since convolution accumulates...

Hi @shu1chen, Your answer is not related to my question that is about Pytorch, and not an example that you said I mentioned, but I did not. Maybe you are...

Now that the API is enabled, I expected the example `examples/ukernels/cpu_brgemm.cpp` to be functional. Still, when running it after enabling `-DDNNL_EXPERIMENTAL_UKERNEL=ON`, it returns the error `Kernel is not supported on...