Disable ARCH_OPT_FLAGS in mkl-dnn

[okapies]

At least v0.16, MKL-DNN try to enable -march=native -mtune=native when -DARCH_OPT_FLAGS is not specified (ref):

set(ARCH_OPT_FLAGS "HostOpts" CACHE STRING
    "specifies compiler optimization flags (see below for more information).
    If empty default optimization level would be applied which depends on the
    compiler being used.

    - For Intel(R) C++ Compilers the default option is `-xHOST` which instructs
      the compiler to generate the code for the architecture where building is
      happening. This option would not allow to run the library on older
      architectures.

    - For GNU* Compiler Collection version 5 and newer the default options are
      `-march=native -mtune=native` which behaves similarly to the descriprion
      above.

    - For all other cases there are no special optimizations flags.

    If the library is to be built for generic architecture (e.g. built by a
    Linux distributive maintainer) one may want to specify ARCH_OPT_FLAGS=\"\"
    to not use any host specific instructions")

It has a chance to fail when you build and distribute the mkl-dnn binary. For example, GCC generates newer CPU instructions supported by the detected processors even if as (assembler) in older binutils in your machine doesn't support them (ref):

# make
[  1%] Building CXX object src/CMakeFiles/mkldnn.dir/common/batch_normalization.cpp.o
/tmp/ccs0bO1c.s: Assembler messages:
/tmp/ccs0bO1c.s:211: Error: suffix or operands invalid for `vpsrlq'
/tmp/ccs0bO1c.s:212: Error: suffix or operands invalid for `vpmovsxdq'
/tmp/ccs0bO1c.s:213: Error: no such instruction: `vextracti128 $0x1,%ymm2,%xmm2'
/tmp/ccs0bO1c.s:214: Error: suffix or operands invalid for `vpmuludq'
/tmp/ccs0bO1c.s:215: Error: suffix or operands invalid for `vpsrlq'
/tmp/ccs0bO1c.s:216: Error: suffix or operands invalid for `vpmuludq'
/tmp/ccs0bO1c.s:217: Error: suffix or operands invalid for `vpmovsxdq'
...

It might be a problem especially for building the mkl-dnn binary on virtualized machines because we can't specify which processor is used.

[saudet]

I see, thanks for bringing this up! Let's have it use AVX (without AVX2) by default though, what do you think?

[okapies]

Do you mean such like -march=sandybridge to support Sandy Bridge and later?

https://wiki.gentoo.org/wiki/Safe_CFLAGS

It seems to be good, but I don’t have idea which option is preferable, or minimum requirement for mkl-dnn. We might ask MKL-DNN community which option is good for default.

[saudet]

Something like -msse4.1 -msse4.2 -mavx is probably safer for AMD CPUs, but yeah I don't know if it really matters and might be best to leave it to the defaults for now...

[okapies]

I posted an issue to mkl-dnn repository. oneapi-src/oneDNN#321

[emfomenk]

Hi, MKL-DNN dev here.

Just want to note that Intel MKL-DNN can be compiled with whatever level of optimization is suitable for a user.

For most of the primitives Intel MKL-DNN generates the code on its own at the runtime for the most appropriate instruction set based on the hardware you run the library, so strictly speaking -mnative is not that important. However, there are few cases where we rely on the compiler optimizations. That's why we use -mnative by default.