Continuation from the following thread: https://lists.apache.org/thread.html/bcb1bd5046ff51049a0556098e756578f6fa6564831d77fddb56432f@%3Cdev.mxnet.apache.org%3E
I am also +1 for making it on master and testing until 1.5.0. We can decide later on (before 1.5.0) to enable mkldnn as default for the nightly build (pip install --pre build) to try to get more feedback if needed. - What the story is like when there's no AVX instructions present on CPUs. Do we get an illegal instruction error, or does it fallback gracefully? According to this issue ( https://github.com/apache/incubator-mxnet/issues/11911), AVX2 is the minimum requirement for pre-build binaries. - Are there any outstanding issues when MKLDNN is enabled? -There is one issues with quantization int8 of mkldnn (will create issue about it when team gives me reproducible code snippet). Additionally, we are waiting to merge the PR to build mkldnn statically with mac/linux when building from source after MKL is added to the CI. - MKLDNN is a submodule dependency, are we pulling the latest commit or releases? If not we should move to releases before we make it a default I agree. We should tag mxnet only to releases from now on. Currently it is tagged to 0.17.1 Please let me know if there any other outstanding issues, else we are going to make mkldnn / cmake default in the Make/CMakefile. Alex
