Fix Travis CI

[cyrilbouvier]

The Goal of the PR is to fix Travis CI. Currently, two out of the four jobs fail: https://travis-ci.org/github/linbox-team/fflas-ffpack/builds/741688422

First, commit 1569db4c51 should fix the homebrew problem with the OS X job (a similar changes was made in Givaro a few months ago).

The other problem is with the job using gcc 4.9 (see issue #313). To fix this issue, I see two possibilities:

1. drop support of gcc 4.9, which is old (last release in 2016) and not supported anymore (https://gcc.gnu.org/gcc-4.9/)
2. do not use any intrinsics not recognize by gcc 4.9. For example, for the first missing intrinsic _mm512_castps512_ps256, I was able to remove it by rewriting the blendv method in simd512_float.inl:

    static INLINE CONST vect_t blendv(const vect_t a, const vect_t b, const vect_t mask) {
   -
   -        __m256 lowa  = _mm512_castps512_ps256(a);
   -        __m256 higha = _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(a),1));
   -
   -        __m256 lowb  = _mm512_castps512_ps256(b);
   -        __m256 highb = _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(b),1));
   -
   -        __m256 lowmask  = _mm512_castps512_ps256(mask);
   -        __m256 highmask = _mm256_castpd_ps(_mm512_extractf64x4_pd(_mm512_castps_pd(mask),1));
   -
   -        __m256 reslow = _mm256_blendv_ps(lowa, lowb, lowmask);
   -        __m256 reshigh = _mm256_blendv_ps(higha, highb, highmask);
   -
   -        __m512 res = _mm512_castps256_ps512(reslow);
   -        res = _mm512_insertf32x8(res, reshigh, 1);
   -
   -        return res;
   +        return _mm512_mask_blend_ps (
   +                    _mm512_movepi32_mask (_mm512_castps_si512 (mask)), a, b);
    }

   The advantage of this new code is that it is faster (only two instructions, as the cast is here for compilation only). The main drawback is that _mm512_movepi32_mask requires AVX512DQ and not just AVX512L. But , according to issue #312 , other parts of the simd code already silently require AVX512DQ.

More work is needed to remove the other mising intrinsics if we go with option 2.

[cyrilbouvier]

Commit 1569db4 did fix the homebrew problem with the OS X job but revealed a failure for test-fger

The problem does not seem to be specific to OS X, as I was able to reproduce it. I had one failed run of test-fger on more than 100 runs, with the following error message:

Checking Modular_implem<int64_t, uint64_t, uint64_t> modulo 165033971... FAIL
a   :132720555
m   :44, n   : 1
incx :1, incy : 1, ldC : 1
Error C[0,0]=18463374 D[0,0]=24729322
Error C[1,0]=54479037 D[1,0]=144408618
Error C[2,0]=67881147 D[2,0]=99501009
Error C[3,0]=84658322 D[3,0]=141056375
Error C[4,0]=35823758 D[4,0]=44270387
Error C[5,0]=35369714 D[5,0]=15002300
Error C[6,0]=114908161 D[6,0]=42428348
Error C[7,0]=53084340 D[7,0]=15942759
Error C[8,0]=94996110 D[8,0]=104641692
Error C[9,0]=39583949 D[9,0]=143044872
Error C[10,0]=140921121 D[10,0]=136917392
Error C[11,0]=43433855 D[11,0]=74462683
Error C[12,0]=143759473 D[12,0]=89899069
Error C[13,0]=96683823 D[13,0]=75362525
Error C[14,0]=28275651 D[14,0]=78298271
Error C[15,0]=149033254 D[15,0]=138700481
Error C[16,0]=126252883 D[16,0]=66669534
Error C[17,0]=28226674 D[17,0]=51199424
Error C[18,0]=81191980 D[18,0]=144124893
Error C[19,0]=69444028 D[19,0]=61800979
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
FAILED 

Edit: I was able to find a seed that trigger the error:

./test-fger -s 355057717

Edit 2: It seems this issue is already under investigation: see issue #307 and PR #289

[cyrilbouvier]

A more exhaustive list of intrinsics not recognized by gcc 4.9:

• _mm512_castps512_ps256: see original post
• _mm512_mullo_epi64: used in mullo for Simd512\. Another problem is that it requires AVX512DQ but is only protected by the AVX512F macro (see #312)
  • _mm512_castpd256_pd512: used in blendv and hadd for Simd512\
• _mm512_castpd512_pd256: used in blendv and hadd for Simd512\
• _mm512_castps256_ps512: used in blendv and hadd for Simd512\
• _mm512_castps_pd: used in hadd for Simd512\
• _mm512_castsi512_pd: used in eq, lesser, lesser_eq, greater, greater_eq for Simd512\
• _mm512_castsi512_ps: used in eq, lesser, lesser_eq, greater, greater_eq for Simd512\
• _mm512_cmpgt_epu64_mask: used in greater for Simd512\
• _mm512_cvtepi64_pd: used in mod for Simd512\. Another problem is that it requires AVX512DQ but is only protected by the AVX512F macro (see #312)
• _mm512_cvtpd_epi64: used in mod for Simd512\. Another problem is that it requires AVX512DQ but is only protected by the AVX512F macro (see #312)
• _mm512_insertf32x8: used in hadd for for Simd512\. Another problem is that it requires AVX512DQ but is only protected by the AVX512F macro (see #312)

[ClementPernet]

OK to merge this PR as is in order to fix the hombrew problem. Regarding the support for 4.9 as minimal gcc version, it seems that the numerous avx512 undefined intrinsics is a strong enough argument to move to gcc-5 as minimal supported gcc: we should not waste our time writing workarounds. Still the one you propose deserves its own PR, if it improves the current implementation. The fger bug is already mentionned in #307 and need our attention!

[ClementPernet]

LGTM . Merging