AMD AOCC 2.2 Helping Squeeze Extra Performance Out Of AMD EPYC 7002 "Rome" CPUs

In a few cases, the AOCC Zen 2 optimizations weren't paying off.

One software package where being built under AOCC 2.2 was repeatedly paying off was Intel's oneDNN (formerly known as MKL-DNN and DNNL). With this deep neural network library, the AOCC-generated binaries tended to be slightly faster than Clang 10. which in general for oneDNN were much faster than the GCC 10 binary.

When building Intel's SVT-VP9 video encoder with AOCC 2.2 was another case where the AMD EPYC 7742 2P server was enjoying better performance over Clang or GCC.