在 2016年11月16日星期三 UTC+8上午12:49:13，Ángel de Vicente写道：
>
> Hi,
>
> I wanted to try this out. My test Fortran module is the following, which
> I compile with: ifort -mkl -o test_mod.so test_mod.f90 -shared -fpic
> ,----
> | module testmodule
> |
> | implicit none
> |
> | double precision, external :: ddot
> | double precision, dimension(3) :: d,e
> |
> | integer :: i
> |
> | contains
> |
> | function dot_f()
> | double precision :: dot_f
> |
> | do i = 1,3
> | d(i) = 1.0d0*i
> | e(i) = 3.5d0*i
> | end do
> |
> | dot_f = ddot(3,d,1,e,1)
> | end function dot_f
> |
> | END module testmodule
> `----
>
> and my test.jl
> ,----
> | ccall((:testmodule_mp_dot_f_, "./test_mod"),Float64, ())
> `----
>
> If I try to run test.jl I get, as per the OP:
> ,----
> | julia> include("test.jl")
>
>
> | Intel MKL FATAL ERROR: Cannot load libmkl_avx.so or libmkl_def.so.
> `----
>
>
> "Steven G. Johnson" <[email protected] <javascript:>> writes:
> > I got
> > Intel MKL FATAL ERROR: Cannot load libmkl_avx2.so or libmkl_def.so.
> >
> > Possibly you need to add the directory containing these files
> > (/opt/intel/composer_xe_2015.0.090/mkl/lib or similar?) to your
> LD_LIBRARY_PATH
> > environment variable, so that the runtime linker knows where to find
> them.
>
> In this case this is not enough.
>
> If you try to open the library directly you get:
> ,----
> | julia>
> Libdl.dlopen("/opt/intel/2016.0.1/compilers_and_libraries_2016.1.150/linux/mkl/lib/intel64/libmkl_avx2.so")
>
>
> | ERROR: could not load library
> |
> "/opt/intel/2016.0.1/compilers_and_libraries_2016.1.150/linux/mkl/lib/intel64/libmkl_avx2.so"
>
>
> |
> /opt/intel/2016.0.1/compilers_and_libraries_2016.1.150/linux/mkl/lib/intel64/libmkl_avx2.so:
>
>
> | undefined symbol: mkl_dft_fft_fix_twiddle_table_32f
>
> | in dlopen(::String, ::UInt32) at ./libdl.jl:90 (repeats 2 times)
> `----
>
> and nm confirms that those symbols are undefined:
> ,----
> | [angelv@duna intel64]$ nm libmkl_avx2.so | grep fft_fix
> | U mkl_dft_fft_fix_twiddle_table_32f
> | U mkl_dft_fft_fix_twiddle_table_64f
> `----
>
> and they are acutally defined in libmkl_core.so
> ,----
> | [angelv@duna intel64]$ nm libmkl_core.so | grep fft_fix
> | 00000000018e3020 D mkl_dft_fft_fix_twiddle_table_32f
> | 00000000018e2800 D mkl_dft_fft_fix_twiddle_table_64f
> | [angelv@duna intel64]$
> `----
>
>
> So, a workaround is to open libmkl_core.so first with the flag
> RTLD_GLOBAL and then run the test.jl code:
>
> ,----
> | julia>
> |
> Libdl.dlopen("/opt/intel/2016.0.1/compilers_and_libraries_2016.1.150/linux/mkl/lib/intel64/libmkl_core.so",Libdl.RTLD_GLOBAL)
>
>
> | Ptr{Void} @0x0000000003af6fa0
>
>
> |
>
>
> | julia> include("test.jl")
>
>
> | 49.0
> `----
>
> But, to be honest, I don't fully understand if this will be enough for
> all codes using MKL or perhaps other dependencies are there which forces
> you to open more libraries manually with RTLD_GLOBAL. But at least it
> points in the right direction.
>
> Cheers,
> --
> Ángel de Vicente
> http://www.iac.es/galeria/angelv/
>
You have solved the problem, that's nice. But maybe I should reconsider
whether to use mkl in such a uncomfortable way.
