travis: added ppc64le and s390x test.

[junaruga]

This PR is to add ppc64le and s390x test to Travis CI. There are 2 commits. 1st commit is to rebase the commits used in #604 .

Travis is ok on my forked repository here.

It seems Debian minimap2 deb package already has both ppc64le and 390x builds using simde in their Makefile. I am not sure if we should set -fopenmp-simd -DSIMDE_ENABLE_OPENMP flag in Makefile.simde like Debian's one is doing.

[mbrcic]

Flags -fopenmp-simd -DSIMDE_ENABLE_OPENMP under SIMDe do indeed bring performance gains for ppc64le and s390x, assuming there is OpenMP support. These flags help when SIMDe does not know the translation to platform-native intrinsics so instead adds vectorization hints. In that case, OpenMP SIMD gives the best performance.

For ARM NEON, there should be no gains, because SSE4.1 and SSE2 are (almost?) completely transformed to NEON intrinsics.

[mr-c]

@mbrcic Luckily -fopenmp-simd/-qopemp-simd only require compiler support (gcc/icc respectively) and not the entire OpenMP API :-)

[junaruga]

@mbrcic thanks for the detailed explanation. I found an explanation of -fopenmp-simd -DSIMDE_ENABLE_OPENMP in simde README.

For best performance, in addition to -O3 (or whatever your compiler's equivalent is), you should enable OpenMP 4 SIMD support by defining SIMDE_ENABLE_OPENMP before including any SIMDe headers, and enabling OpenMP support in your compiler.

I also found the explanation of -fopenmp-simd at gcc mailing list. That is why -fopenmp-simd is better than -fopenmp. I remember that in gcc 4.8 or 4.9 ppc64le version -fopenmp-simd is not supported while -fopenmp is supported.

[junaruga]

@lh3 Could you check this PR? Here is the Travis result: https://travis-ci.org/github/lh3/minimap2/builds/688098723 . Thanks.

[lh3]

Closing as we are moving away from Travis. Thanks anyway.