Bug 730468 – dev-libs/ncnn-20200413 - In file included from .../work/ncnn-20200413/src/layer/x86/convolution_x86.cpp:22: /usr/lib/gcc/x86_64-pc-linux-gnu/9.3.0/include/fmaintrin.h:63:1: error: inlining failed in call to always_inline ‘__m256 _mm256_fmadd_ps(__m256 ...

• Status: RESOLVED FIXED
• Product: Gentoo Linux

[Description Sergey 'L29Ah' Alirzaev 2020-07-02 17:12:50 UTC]

FAILED: src/CMakeFiles/ncnn.dir/layer/x86/convolution_x86.cpp.o 
/usr/lib/ccache/bin/x86_64-pc-linux-gnu-g++  -I/var/tmp/portage/dev-libs/ncnn-20200413/work/ncnn-20200413/src -Isrc -I/var/tmp/portage/dev-libs/ncnn-20200413/work/ncnn-20200413/src/layer  -O2 -pipe -march=ivybridge -mtune=native -ftree-vectorize -malign-data=cacheline -mtls-dialect=gnu2 -fPIC   -Wall -Wextra -Wno-unused-function -fvisibility=hidden -fvisibility-inlines-hidden -fopenmp -MD -MT src/CMakeFiles/ncnn.dir/layer/x86/convolution_x86.cpp.o -MF src/CMakeFiles/ncnn.dir/layer/x86/convolution_x86.cpp.o.d -o src/CMakeFiles/ncnn.dir/layer/x86/convolution_x86.cpp.o -c /var/tmp/portage/dev-libs/ncnn-20200413/work/ncnn-20200413/src/layer/x86/convolution_x86.cpp
In file included from /usr/lib/gcc/x86_64-pc-linux-gnu/9.3.0/include/immintrin.h:107,
                 from /var/tmp/portage/dev-libs/ncnn-20200413/work/ncnn-20200413/src/layer/x86/convolution_x86.cpp:22:
/usr/lib/gcc/x86_64-pc-linux-gnu/9.3.0/include/fmaintrin.h: In function ‘_ZN4ncnnL24conv3x3s1_winograd23_sseERKNS_3MatERS0_S2_S2_RKNS_6OptionE._omp_fn.1’:
/usr/lib/gcc/x86_64-pc-linux-gnu/9.3.0/include/fmaintrin.h:63:1: error: inlining failed in call to always_inline ‘__m256 _mm256_fmadd_ps(__m256, __m256, __m256)’: target specific option mismatch
   63 | _mm256_fmadd_ps (__m256 __A, __m256 __B, __m256 __C)
      | ^~~~~~~~~~~~~~~
In file included from /var/tmp/portage/dev-libs/ncnn-20200413/work/ncnn-20200413/src/layer/x86/convolution_x86.cpp:32:
/var/tmp/portage/dev-libs/ncnn-20200413/work/ncnn-20200413/src/layer/x86/convolution_3x3.h:506:45: note: called from here
  506 |                     _sum3n = _mm256_fmadd_ps(_r0n, _k3n, _sum3n);
      |                              ~~~~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~


and a lot of similar errors

[Comment 1 Sergey 'L29Ah' Alirzaev 2020-07-02 17:14:50 UTC]

Created attachment 647418 [details]
build log

[Comment 2 Sergey 'L29Ah' Alirzaev 2020-07-02 17:15:21 UTC]

Created attachment 647420 [details]
emerge --info

[Comment 3 Piotr Karbowski (RETIRED) 2020-07-08 19:30:12 UTC]

It seems that march=ivybridge triggers it, works here without march and with march=zenvn2. Can you confirm that dropping march fixes your issue? If so, then we need to bring it up upstream. There's one new ncnn version that I plan to package this weekend so maybe this one will have better results.

[Comment 4 Sergey 'L29Ah' Alirzaev 2020-07-12 11:59:42 UTC]

Yeah, dropping march=ivybridge makes it build.

[Comment 5 Ionen Wolkens 2021-05-30 04:51:30 UTC]

This seems still present in 20210525 version.

-march=ivybridge -mno-avx will also work for this.

Current case here seem to be caused by lacking FMA but any cases where have avx but not avx2 will lead to similar issues. I don't think this code gets tested on avx + no-avx2 much.

If wanted to workaround this from the ebuild until upstream gets to revising these, the "dirty" way would either be:
A) find src -type f -exec sed -i s/__AVX__/__AVX2__/g {} + || die
or
B) use cpu_flags_x86_avx2 || append-flags -mno-avx

For 'B' I don't particularly like the idea of introducing a USE for a workaround, 'A' might be worth some consideration (not that I like it much either).

[Comment 6 Larry the Git Cow 2021-05-31 19:30:41 UTC]

The bug has been closed via the following commit(s):

https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=182a324f44e60684c6c72750ead69fb7149b3012

commit 182a324f44e60684c6c72750ead69fb7149b3012
Author:     Ionen Wolkens <[e-mail]>
AuthorDate: 2021-05-29 07:38:30 +0000
Commit:     Piotr Karbowski <[e-mail]>
CommitDate: 2021-05-31 19:30:38 +0000

    dev-libs/ncnn: add 20210525
    
    ebuild updates:
    - respect $(get_libdir)
    - build shared library over static (this new version also exports
      additional symbols needed by waifu2x-ncnn-vulkan for shared linking)
    - install more docs
    - add IUSE for tools and vulkan
    - add ZLIB license for *_mathfun.h
    - add temporary workaround for bug 730468
    - scrubbed previous patch and added upstream issue link (still needed)
    
    Closes: https://bugs.gentoo.org/730468
    Signed-off-by: Ionen Wolkens <[e-mail]>
    Signed-off-by: Piotr Karbowski <[e-mail]>

 dev-libs/ncnn/Manifest                             |  1 +
 dev-libs/ncnn/files/ncnn-fix-glslang-include.patch | 10 +--
 dev-libs/ncnn/metadata.xml                         |  4 ++
 dev-libs/ncnn/ncnn-20210525.ebuild                 | 76 ++++++++++++++++++++++
 4 files changed, 84 insertions(+), 7 deletions(-)

[Comment 7 Larry the Git Cow 2022-04-20 02:40:09 UTC]

The bug has been referenced in the following commit(s):

https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=92a9da73390b1f2d90011e6b6a6fa8cf25512cec

commit 92a9da73390b1f2d90011e6b6a6fa8cf25512cec
Author:     Ionen Wolkens <[e-mail]>
AuthorDate: 2022-04-20 01:06:53 +0000
Commit:     Ionen Wolkens <[e-mail]>
CommitDate: 2022-04-20 02:39:39 +0000

    dev-libs/ncnn: add 20220419
    
    Also remove now unnecessary avx workaround wrt bug #730468,
    entire usage was refactored upstream and the sed now cause
    issues instead.
    
    Bug: https://bugs.gentoo.org/730468
    Signed-off-by: Ionen Wolkens <[e-mail]>

 dev-libs/ncnn/Manifest             |  1 +
 dev-libs/ncnn/ncnn-20220419.ebuild | 73 ++++++++++++++++++++++++++++++++++++++
 2 files changed, 74 insertions(+)