could not use blas when building with cmake

[hubutui]

according to https://github.com/apache/incubator-mxnet/blob/master/make/config.mk#L117, USE_BLAS could be blas. However, https://github.com/apache/incubator-mxnet/blob/master/cmake/ChooseBlas.cmake seems don't deal with blas.

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended labels: Build

[hubutui]

Kindly remind, cmake provides FindBLAS module for finding variety blas package, see https://cmake.org/cmake/help/v3.15/module/FindBLAS.html

[leleamol]

@mxnet-label-bot add [Build, Blas]

[larroy]

#14871

[larroy]

@edisongustavo

[hubutui]

emm, using the FindBLAS module would be a great idea. And blas (blas, cblas, openblas, or atlas) is provided in Linux distribution, which doesn't need to be compiled from source again.

[larroy]

@hubutui could you help review the above PR with these suggestions? Thanks.

[hubutui]

@larroy Thanks bro. I read the pr days ago. I would like to help, but I have little experience with cmake. There several thoughts comes to my mind:

1. for mkl, mklml, and mkldnn, use the current solution, since cmake could not deal with it using FindBLAS.
2. for other blas (atlas, openblas, etc), use the FindBLAS module from cmake.

[hubutui]

actually, I'm a little confused about the math library selection. It seems that we could choose one of:

• atlas
• openblas
• mkl
  where for mkl, we could choose the full mkl, mkl subset (aks mklml), or mkl-dnn.
  However, https://github.com/apache/incubator-mxnet/blob/master/docs/tutorials/mkldnn/MKLDNN_README.md#1 says that we could use openblas instead of mkl with mkl-dnn.
  Am I right that USE_BLAS is independent with USE_MKLDNN? And there is no USE_BLAS=mkldnn?

[TaoLv]

@hubutui I think you mean the BLAS library selection. MKL-DNN is a performance library for deep learning computations rather than a BLAS library, although it provides some GEMM implementations. So MKL-DNN doesn't participate the BLAS library selection.

Am I right that USE_BLAS is independent with USE_MKLDNN? And there is no USE_BLAS=mkldnn?

Yes, you're right.

[hubutui]

@TaoLv Then https://mxnet.incubator.apache.org/versions/master/install/build_from_source.html#math-library-selection lists mkl full, mklml and mkl-dnn in the same level, which really make users confused.
So we could choose:

• Apple Accelerate
• ATLAS
• MKL (MKL, MKLML)
• OpenBLAS
  as the BLAS library selection. And we could enable MKL-DNN submodule to further speed up CPU computing?

[hubutui]

emm, finally, I build mxnet with openblas & mkldnn, it works great for me.