BUG: Build of numpy 1.23.0 fails on debian-stretch

[mkoeppe]

Describe the issue:

On debian-stretch (gcc-6_6.3.0-18+deb9u1_amd64), building numpy 1.23.0 from source fails, see details in https://trac.sagemath.org/ticket/34110#comment:4
This is a regression from numpy 1.22.4.

Reproduce the code example:

pip install .

Error message:

No response

NumPy/Python version information:

1.23.0

[mattip]

The error seems to be a slight differences in SIMD compilation. @seiko2plus is there a known problem with gcc 6?:

In file included from numpy/core/src/common/simd/avx512/avx512.h:75:0,
                   from numpy/core/src/common/simd/simd.h:50,
                   from /sage/local/var/lib/sage/venv-python3.10.3/var/tmp/sage/build/numpy-1.23.0/src/numpy/core/src/npysort/x86-qsort.dispatch.cpp:17,
                   from build/src.linux-x86_64-3.10/numpy/core/src/npysort/x86-qsort.dispatch.avx512_skx.cpp:21:
  numpy/core/src/common/simd/avx512/arithmetic.h: In function 'npy_uint32 npyv_reducemax_u32(npyv_u32)':
  numpy/core/src/common/simd/avx512/arithmetic.h:387:89: error: invalid conversion from 'int' to '_MM_PERM_ENUM' [-fpermissive]
           npyv_u32 a3 = _mm512_max_epu32(a2, _mm512_shuffle_epi32(a2, (1<<6 | 0<<4 | 3<<2 | 2)));
                                                                       ~~~~~~~~~~~~~~~~~~~~^~~~
...
 In file included from build/src.linux-x86_64-3.10/numpy/core/src/npysort/x86-qsort.dispatch.avx512_skx.cpp:21:0:
  /sage/local/var/lib/sage/venv-python3.10.3/var/tmp/sage/build/numpy-1.23.0/src/numpy/core/src/npysort/x86-qsort.dispatch.cpp: In function 'npy_int partition_vec(type_t*, npy_intp, npy_intp, zmm_t, zmm_t, zmm_t*, zmm_t*)':
  /sage/local/var/lib/sage/venv-python3.10.3/var/tmp/sage/build/numpy-1.23.0/src/numpy/core/src/npysort/x86-qsort.dispatch.cpp:654:64: error: there are no arguments to '_knot_mask16' that depend on a template parameter, so a declaration of '_knot_mask16' must be available [-fpermissive]
       vtype::mask_compressstoreu(arr + left, _knot_mask16(gt_mask), curr_vec);
      

[rgommers]

It looks like we don't have a CI job with the lowest-supported GCC version; we should probably fix that.

[mkoeppe]

Related to that, https://numpy.org/doc/stable/user/building.html#prerequisites could probably be made clearer. There's still talk about GCC 4.x. https://docs.scipy.org/doc/scipy/dev/toolchain.html#c-compilers is much better

[seiko2plus]

@mattip,

is there a known problem with gcc 6?

Regards to the current build error, here are some:

• limited support for AVX512 mask operations

  • synonyms intrinsics _k[and, or, xor, not]_mask16 don't exist, we usually use _mm512_[and, or, xor, not] since its widely supported even on old versions of MSVC.
  • AVX512BW mask operations aren't supported at all, usually we fall down to integer registers
  • AVX512DQ same as BW, as workaround we cast around __mmask16.
  • most of these cases are handled in

    numpy/numpy/core/src/common/simd/avx512/operators.h

    Line 146 in 75ae7be

    /***************************

• emulated reduce intrinsics aren't supported

[mkoeppe]

Thanks all!