Intel MKL FATAL ERROR: Cannot load libmkl_avx.so or libmkl_def.so

[emailweixu]

When paddle is compiled with MKL (by setting INTEL_MKL_ROOT when running cmake), _swig_paddle.so is linked against some libmkl_*.so. When MKL functions are actually used by the program, another MKL library libmkl_avx.so is dynamically loaded by MKL and some symbols in libmkl_avx.so need to be resolved from libmkl_intel_sequential.so. However, the symbols of libmkl_intel_sequential.so are not visible to libmkl_avx.so because when python imports an .so as an module the so lib is dlopened without RTLD_GLOBAL flag by default.

There are several ways to solve this issue:

1. Using static library. Which can be achieved using the following two methods:
   i. Change cmake/cblas.cmake to search for static libraries libmkl_*.a explicitly at several find_library() calls.
   ii. Move .so files under INTEL_MKL_ROOT/lib/intel64 to a different location

2. Using RTLD_GLOBAL to do import. This can be achieved by adding a line at paddle/py_paddle/__init__.py:
   sys.setdlopenflags(sys.getdlopenflags() | ctypes.RTLD_GLOBAL)

[lcy-seso]

I encountered the same problem and this issue explains the reason. Thank you.

[luotao1]

I encountered the same problem before.

However, as Paddle support MKLML & MKLDNN currently, we could use MKLML to instead INTEL_MKL_ROOT. That is, setting WITH_MKLML=ON and WITH_MKLDNN=ON, all the MKL related libraries will be auto downloaded, compiled and installed.

Thus, can we remove INTEL_MKL_ROOT? @emailweixu @tensor-tang