Build issue on OSX

Question

Build issue on OSX

Closed this issue 6 years ago · 9 comments

Hi, on OSX I'm getting a build error, either with --enable-precompilation or without (giving the same error on "make check").

./configure --prefix=$HOME/sandbox --with-blas-libs="-lopenblas" --enable-precompilation
make

First error amongst many:

In file included from ../../../fflas-ffpack/fflas/fflas_simd/simd512.inl:76:
../../../fflas-ffpack/fflas/fflas_simd/simd512_float.inl:554:10: error: argument to '__builtin_ia32_rndscaleps_mask' must be a constant integer
return _mm512_roundscale_ps(a, i);

Not sure what I'm missing
Help would be most appreciated.
Thx. Mark

fflas-build-errors.txt

Answer 1 · 2018-09-26T15:51:27.000Z

This seems to have been an issue with the configuration incorrectly determining that I had AVX-512 extensions. Everything compiles when --disable-avx512f and --disable-avx512dq options are used on the autoconfig.

Answer 2 · 2018-09-26T19:09:30.000Z

Hi Mark, sorry for my delayed reply.
This mis-detection of avx-512 is still a bug, so let me reopen the issue so we can try an fix it.
Could you give us more feedback on your setting:
CPU (not only i5-i7 but also the 4 digits following it), which compiler (clang --version or gcc -v), etc.
Thanks for your help.

Answer 3 · 2018-09-26T19:15:41.000Z

Hi Clément, Here’s some more info: % clang --version Apple LLVM version 10.0.0 (clang-1000.10.44.2) Target: x86_64-apple-darwin18.0.0 Thread model: posix InstalledDir: /Library/Developer/CommandLineTools/usr/bin % sysctl -n machdep.cpu.brand_string Intel(R) Xeon(R) W-2140B CPU @ 3.20GHz Best, Mark

…

__ Dr. Mark Giesbrecht, Professor David R. Cheriton School of Computer Science, University of Waterloo, Canada Phone:+1.519.888.4080<tel:+15198884080> Email:mwg@uwaterloo.ca<mailto:mwg@uwaterloo.ca> URL:https://cs.uwaterloo.ca/~mwg On Sep 26, 2018, at 3:11 PM, Clément Pernet <notifications@github.com<mailto:notifications@github.com>> wrote: Hi Mark, sorry for my delayed reply. This mis-detection of avx-512 is still a bug, so let me reopen the issue so we can try an fix it. Could you give us more feedback on your setting: CPU (not only i5-i7 but also the 4 digits following it), which compiler (clang --version or gcc -v), etc. Thanks for your help. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub<#179 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJWon3hDV5EE8LF1zLhqweYHe2m74ZyGks5ue9FigaJpZM4W04yM>.

Answer 4 · 2018-09-26T19:17:46.000Z

Now that I look further, this is weird: that processor should have AVX-512 and AVX-512DQ! Best, Mark

…

__ Dr. Mark Giesbrecht, Professor David R. Cheriton School of Computer Science, University of Waterloo, Canada Phone:+1.519.888.4080<tel:+15198884080> Email:mwg@uwaterloo.ca<mailto:mwg@uwaterloo.ca> URL:https://cs.uwaterloo.ca/~mwg On Sep 26, 2018, at 3:15 PM, Mark Giesbrecht <mwg@uwaterloo.ca<mailto:mwg@uwaterloo.ca>> wrote: Hi Clément, Here’s some more info: % clang --version Apple LLVM version 10.0.0 (clang-1000.10.44.2) Target: x86_64-apple-darwin18.0.0 Thread model: posix InstalledDir: /Library/Developer/CommandLineTools/usr/bin % sysctl -n machdep.cpu.brand_string Intel(R) Xeon(R) W-2140B CPU @ 3.20GHz Best, Mark

__ Dr. Mark Giesbrecht, Professor David R. Cheriton School of Computer Science, University of Waterloo, Canada Phone:+1.519.888.4080<tel:+15198884080> Email:mwg@uwaterloo.ca<mailto:mwg@uwaterloo.ca> URL:https://cs.uwaterloo.ca/~mwg On Sep 26, 2018, at 3:11 PM, Clément Pernet <notifications@github.com<mailto:notifications@github.com>> wrote: Hi Mark, sorry for my delayed reply. This mis-detection of avx-512 is still a bug, so let me reopen the issue so we can try an fix it. Could you give us more feedback on your setting: CPU (not only i5-i7 but also the 4 digits following it), which compiler (clang --version or gcc -v), etc. Thanks for your help. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub<#179 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJWon3hDV5EE8LF1zLhqweYHe2m74ZyGks5ue9FigaJpZM4W04yM>.

Answer 5 · 2018-09-26T19:40:32.000Z

Ok then, that's what I was starting to guess: you're the happy first one to compile our avx512 code with llvm. The intrinsics are implemented slightly differently from gcc and this may be the reason why we did not see them before. The first instances seem to be easy to fix.
I've pushed a fix to master. Let me know it fixes it.

Answer 6 · 2018-09-26T19:50:45.000Z

Thanks! New processors are always fun. Your fix gets me a little further. Next collection of errors looks like: In file included from ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.h:97: ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.inl:363:44: error: too few arguments to function call, expected 8, have 4 R0 = simd::set(r0[0], r1[0], r2[0], r3[0]); // could be ... ~~~~~~~~~ ^ ../../../fflas-ffpack/fflas/fflas_simd/simd512_int64.inl:103:2: note: 'set' declared here static INLINE CONST vect_t set(const scalar_t x0, const scalar_... ^ In file included from fflas_L1_inst.C:34: In file included from ../../../fflas-ffpack/fflas/fflas.h:109: In file included from ../../../fflas-ffpack/fflas/fflas_fgemm.inl:443: In file included from ../../../fflas-ffpack/fflas/fflas_fgemm/fgemm_classical.inl:43: In file included from ../../../fflas-ffpack/fflas/fflas_igemm/igemm.h:45: In file included from ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.h:97: ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.inl:385:21: error: no member named 'get' in 'Simd512_impl<true, true, true, 8>' r0[0] = simd::get(R0, 0); ~~~~~~^ ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.inl:386:21: error: no member named 'get' in 'Simd512_impl<true, true, true, 8>' r1[0] = simd::get(R0, 1); ~~~~~~^ ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.inl:387:21: error: no member named 'get' in 'Simd512_impl<true, true, true, 8>' r2[0] = simd::get(R0, 2); ~~~~~~^ ../../../fflas-ffpack/fflas/fflas_igemm/igemm_kernels.inl:388:21: error: no member named 'get' in 'Simd512_impl<true, true, true, 8>' r3[0] = simd::get(R0, 3); ~~~~~~^ In file included from fflas_L1_inst.C:34: Full output from config and make are attached. Now if I could only get Linbox to build as well…! Cheers, Mark

…

__ Dr. Mark Giesbrecht, Professor David R. Cheriton School of Computer Science, University of Waterloo, Canada Phone:+1.519.888.4080<tel:+15198884080> Email:mwg@uwaterloo.ca<mailto:mwg@uwaterloo.ca> URL:https://cs.uwaterloo.ca/~mwg On Sep 26, 2018, at 3:40 PM, Clément Pernet <notifications@github.com<mailto:notifications@github.com>> wrote: Ok then, that's what I was starting to guess: you're the happy first one to compile our avx512 code with llvm. The intrinsics are implemented slightly differently from gcc and this may be the reason why we did not see them before. The first instances seem to be easy to fix. I've pushed a fix to master. Let me know it fixes it. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub<#179 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AJWonxkSg0rDqLu5FAhEf_dVI9qMOlpPks5ue9gygaJpZM4W04yM>.

Answer 7 · 2018-09-26T20:00:35.000Z

Well these ones are less trivial. I just wonder how it could work with gcc! I'll need to look into it with fresh eyes tomorow.
Meanwhile, you can safely use fflas-ffpack with the --disable-avx512f --disable-avx512dq that you mentionned. No big performance drop should be noticed.

Answer 8 · 2018-09-27T16:14:41.000Z

No problem and thanks. I’m happy to try things out as you get time for the fixes. Best, Mark

…

__ Dr. Mark Giesbrecht, Professor David R. Cheriton School of Computer Science, University of Waterloo, Canada Phone:+1.519.888.4080 <tel:+15198884080> Email:mwg@uwaterloo.ca <mailto:mwg@uwaterloo.ca> URL:https://cs.uwaterloo.ca/~mwg <https://cs.uwaterloo.ca/~mwg>

On Sep 26, 2018, at 4:00 PM, Clément Pernet ***@***.***> wrote: Well these ones are less trivial. I just wonder how it could work with gcc! I'll need to look into it with fresh eyes tomorow. Meanwhile, you can safely use fflas-ffpack with the --disable-avx512f --disable-avx512dq that you mentionned. No big performance drop should be noticed. — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#179 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AJWonzUJiiOrZ7h7QR36BePeyPfwDxvqks5ue9zlgaJpZM4W04yM>.

Answer 9 · 2018-10-11T12:48:35.000Z

Mark,
This should be fixed now, since PR #183 was merged. I could reproduce the bug check that it had been fixed on a skylake scalable machine, but feel free to let me know if you can confirm it on your side.
I'm closing the issue. If the problem remains, feel free to reopen.