Native MIC compile error: undefined reference

[Patrick_S_]

Hey all,

i'm currently trying to compile the following native Intel Intrinsics code.

    for ( std::size_t k = 0; k < n; k += 16 ) {

        x_ = _mm512_load_ps( x + k );

        u_ = _mm512_load_ps( u + k );

        xhd_ = _mm512_movehdup_ps( x_ );

        upm_ = _mm512_permute_ps( u_, 177 );

        hdpm_ = _mm512_mul_ps( xhd_, upm_ );

        xld_ = _mm512_moveldup_ps( x_ );

        cmplx_ = _mm512_fmaddsub_ps( xld_, u_, hdpm_ );

        _mm512_store_ps( x + k , cmplx_ );

    }

and thereby I get a compile error:

mic_test.cpp:(.text+0x1097): undefined reference to `_mm512_movehdup_ps'

mic_test.cpp:(.text+0x10b1): undefined reference to `_mm512_permute_ps'

mic_test.cpp:(.text+0x10ce): undefined reference to `_mm512_moveldup_ps'

mic_test.cpp:(.text+0x10e8): undefined reference to `_mm512_fmaddsub_round_ps'

I have included immintrin.h, the memory is 64 byte aligned and I checked

http://software.intel.com/en-us/comment/1726413#comment-1726413

But the Problem is that my program runs in native mode. I don't use pragmas for offloading the code to the MIC.

The strange thing is that a similar code:

    for ( std::size_t k = 0; k < n; k += 16 ) {

        x_ = _mm512_load_ps( x + k );

        z_ = _mm512_load_ps( z + k );

        u_ = _mm512_load_ps( u + k );

        v_ = _mm512_load_ps( v + k );

        zv_ = _mm512_mul_ps( z_, v_ );

        real_ = _mm512_fmsub_ps( x_, u_, zv_);

        zu_ = _mm512_mul_ps( z_, u_ );

        imag_ = _mm512_fmadd_ps( x_, v_, zu_);

        _mm512_store_ps( x + k , real_ );

        _mm512_store_ps( z + k , imag_ );

    }

compiles and runs without problems on the MIC.

Thanks,

Patrick

[Kevin_D_Intel]

The AVX 512 instructions and those particular unresolved intrinsics (discussed in the C++ UG here: http://software.intel.com/en-us/node/485150) are not currently available for Intel Xeon Phi™. They will be available with Knight’s Landing. There is a nice discussion in James’ blog: http://software.intel.com/en-us/blogs/2013/avx-512-instructions.

[Patrick_S_]

Is it documented which AVX512 Intrinsics will work on the current Xeon Phi? or do I have to test each Intrinsics function?

Can I use an inline assembly instruction e.g. vmovshdup instead of _mm512_movehdup_ps? Or did you mean that the whole AVX Instruction ist not available?

[Charles_C_Intel1]

My understanding is that no AVX512 instrincs or instructions work on the current Xeon Phi coprocessor ("Knight's Corner").  Yes, Knights Corner has 512-bit operations, but they are not the same as the AVX512 operations.  Sorry.

[Patrick_S_]

A lot of AVX512 Intrinsics work!

for example this code:

        x_ = _mm512_load_ps( x + k );

        z_ = _mm512_load_ps( z + k );

        u_ = _mm512_load_ps( u + k );

        v_ = _mm512_load_ps( v + k );

        zv_ = _mm512_mul_ps( z_, v_ );

        real_ = _mm512_fmsub_ps( x_, u_, zv_);

        zu_ = _mm512_mul_ps( z_, u_ );

        imag_ = _mm512_fmadd_ps( x_, v_, zu_);

        _mm512_store_ps( x + k , real_ );

        _mm512_store_ps( z + k , imag_ );

[Patrick_S_]

ah ok. Am I right that this link

http://software.intel.com/en-us/node/460862

is the list for all Intrinsics functions that are available for Knights Corner?

[Kevin_D_Intel]

Yes, you are right.

[Patrick_S_]

ok, thanks :)