I understand that MKLML was experimental but like you pointed out it has been around for a while, so I am presuming some bugs would have been fixed. Users have also seen about 30% performance boost with MKL, now with the latest release we are dropping support and forcing to use another experimental which we know(from all the discussion during the release) that its not stable. Also OpenBlas=MKL does not bring all the features for MKLML and MKLML will have a smaller sized library per this response by a Intel member https://github.com/intel/mkl-dnn/issues/102.
IMHO this is not a good practice, I also wish that a discussion happen on dev list before we drop support, so more people provide feedback. On Wed, May 16, 2018 at 11:54 AM, Anirudh <[email protected]> wrote: > Hi Naveen, > > There was an earlier flag for using MKLML which was a subset of MKL. I > think this flag was removed when MKLDNN was added. > I think the only replacement to the old MKLML is the MKLDNN feature that > was added. ( I am assuming this also installs the MKLML) > So if users don't want to use MKLDNN feature they will have to install the > full MKL library and set USE_BLAS=mkl if I am not wrong. > > I would also like to understand more from others about what is the best way > to transition for users who were using MKLML flag earlier but don't want to > use > MKLDNN flag now ? (Both flags were experimental, but MKLML flag had been > around for some time.) > > Anirudh > > > On Wed, May 16, 2018 at 11:38 AM, Naveen Swamy <[email protected]> wrote: > > > How do I build the latest release(1.2.0rc3) with MKL enabled, it looks > like > > its now defaulting to the experimental MKLDNN. I do not want to MKLDNN as > > it is experimental. > > > > -Naveen > > >
