can compile with MKLDNN + MKLML but no MKL

[MaeThird]

Because MKL is really large and needs register in Intel' web. It's tedious in the production environment.
I tried to compile the mxnet with MKLDNN and MKLML without MKL but failed in the linking processing. It looks like the 'mshadow' module needs the MKL
I think if the entire MKL is convenient there will be no MKLML anymore. if the proj is not designed for production it's just a toy for ever.

[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

[pengzhao-intel]

@MaeThird thanks for trying MXNet with MKLDNN backend.
Yes, it should work without MKL build even in mshadow.
Could you try the build with 'USE_BLAS=open`?

Please paste the error message when you build without MKL.

[pengzhao-intel]

cc @yinghu5 @NeoZhangJianyu

[MaeThird]

Hi @pengzhao-intel
MKLDNN + MKLML + CBLAS( Atlas or OpenBlas) works as expected.
MKLDNN + MKLML + USE_BLAS_MKL=1 (Actually there was no MKL,I just replaced it with MKLML) with the error output:

 53%] Built target mxnet_static
[ 55%] Built target dmlc
[ 84%] Built target mkldnn
[ 84%] Built target libomp-needed-headers
[ 92%] Built target omp
[ 93%] Built target mxnet
[ 94%] Built target gtest
[ 94%] Built target gtest_main
[ 94%] Linking CXX executable mxnet_unit_tests
/usr/bin/ld: ../libmxnet.a(elemwise_unary_op_basic.cc.o): in function `void mxnet::op::UnaryOp::MKL_Compute<mxnet::op::mshadow_op::erf, mxnet::op::mkl_func::erf>(nnvm::NodeAttrs const&, mxnet::OpContext const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&, std::vector<mxnet::OpReqType, std::allocator<mxnet::OpReqType> > const&, std::vector<mxnet::TBlob, std::allocator<mxnet::TBlob> > const&)':
elemwise_unary_op_basic.cc:(.text._ZN5mxnet2op7UnaryOp11MKL_ComputeINS0_10mshadow_op3erfENS0_8mkl_func3erfEEEvRKN4nnvm9NodeAttrsERKNS_9OpContextERKSt6vectorINS_5TBlobESaISF_EERKSE_INS_9OpReqTypeESaISK_EESJ_[_ZN5mxnet2op7UnaryOp11MKL_ComputeINS0_10mshadow_op3erfENS0_8mkl_func3erfEEEvRKN4nnvm9NodeAttrsERKNS_9OpContextERKSt6vectorINS_5TBlobESaISF_EERKSE_INS_9OpReqTypeESaISK_EESJ_]+0x93): undefined reference to `vsErf'
/usr/bin/ld: elemwise_unary_op_basic.cc:(.text._ZN5mxnet2op7UnaryOp11MKL_ComputeINS0_10mshadow_op3erfENS0_8mkl_func3erfEEEvRKN4nnvm9NodeAttrsERKNS_9OpContextERKSt6vectorINS_5TBlobESaISF_EERKSE_INS_9OpReqTypeESaISK_EESJ_[_ZN5mxnet2op7UnaryOp11MKL_ComputeINS0_10mshadow_op3erfENS0_8mkl_func3erfEEEvRKN4nnvm9NodeAttrsERKNS_9OpContextERKSt6vectorINS_5TBlobESaISF_EERKSE_INS_9OpReqTypeESaISK_EESJ_]+0xf9): undefined reference to `vdErf'
collect2: error: ld returned 1 exit status
make[2]: *** [tests/CMakeFiles/mxnet_unit_tests.dir/build.make:354: tests/mxnet_unit_tests] Error 1
make[1]: *** [CMakeFiles/Makefile2:1825: tests/CMakeFiles/mxnet_unit_tests.dir/all] Error 2

It turns out MKLML cannot replace the MKL as the blas using.
Another small issue is that [https://github.com/intel/mkl-dnn/blob/41bee20d7eb4a67feeeeb8d597b3598994eb1959/src/common/utils.cpp#L53] strncpy(value, buffer, value_length);will case

In file included from /usr/include/string.h:494,
                 from /home/mae/mxnet/incubator-mxnet/3rdparty/mkldnn/src/common/utils.cpp:17:
In function ‘char* strncpy(char*, const char*, size_t)’,
    inlined from ‘int mkldnn::impl::mkldnn_getenv(char*, const char*, int)’ at /home/mae/mxnet/incubator-mxnet/3rdparty/mkldnn/src/common/utils.cpp:53:24:
/usr/include/bits/string_fortified.h:106:34: error: ‘char* __builtin_strncpy(char*, const char*, long unsigned int)’ output truncated before terminating nul copying as many bytes from a string as its length [-Werror=stringop-truncation]
  106 |   return __builtin___strncpy_chk (__dest, __src, __len, __bos (__dest));
      |          ~~~~~~~~~~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
/home/mae/mxnet/incubator-mxnet/3rdparty/mkldnn/src/common/utils.cpp: In function ‘int mkldnn::impl::mkldnn_getenv(char*, const char*, int)’:
/home/mae/mxnet/incubator-mxnet/3rdparty/mkldnn/src/common/utils.cpp:49:34: note: length computed here
   49 |             value_length = strlen(buffer);

with GCC-9.1 . It's handy to fix it.

[pengzhao-intel]

MKLDNN + MKLML + USE_BLAS_MKL=1 (Actually there was no MKL,I just replaced it with MKLML) with the error output:

This is not the right usage because MKLML is just a subset of MKL and only works for MKL-DNN library.
It doesn't target to work as a full BLAS library.

You can get MKL by yum or apt to avoid web register.
https://software.intel.com/en-us/articles/installing-intel-free-libs-and-python-yum-repo
https://software.intel.com/en-us/articles/installing-intel-free-libs-and-python-apt-repo

As you see the "vsErf" is not included in MKLML but it's a part of MKL.

cc @TaoLv

[MaeThird]

@pengzhao-intel Thanks for your help.

[TaoLv]

Thanks for pinging, @pengzhao-intel .

@MaeThird ,

Regarding the first error, yes, the vsErf function is not part of MKLML. So you need full MKL to take advantage of this function.
Regarding the compiling error with GCC 9+, it's fixed in MKL-DNN 0.20 release and just upstreamed to mxnet half an hour ago (#15422). Would you mind trying the master branch to see if the error is still there?