Self-test errors on macOS about nsimd

[gquintin]

Hi Erik,

Is it possible to run the following tests on your MAC and tell me the results?

1. If you still have the binaries of the tests, run them manually on your terminal:

   ./tests.cxx_adv.notl.u8.cpp98 && echo ok || echo fail

2. Run the same commands several times to see if the results are reproducible?

   for i in `seq 1 10`; do
       mkdir build-$i
       cd build-$i
       ../nstools/bin/nsconfig .. -Dsimd=avx2 -Dmpfr='-I/opt/local/include -L/opt/local/lib -lmpfr'
       ../nstools/bin/nstest -j$(nproc) 1>../log-n-${i}.txt 2>&1
       ../nstools/bin/nstest -j1 1>../log-1-${i}.txt 2>&1
   done

3. Do the same but with GCC. From what I know of MACOS, GCC is by default a symlink to clang. So make sure that GCC is actually GCC:

   mkdir build-gcc
   cd build-gcc
   ../nstools/bin/nsconfig .. -Dsimd=avx2 -Dmpfr='-I/opt/local/include -L/opt/local/lib -lmpfr' -comp=gcc
   ../nstools/bin/nstest -j$(nproc) 1>../log-n-${i}.txt 2>&1
   ../nstools/bin/nstest -j1 1>../log-1-${i}.txt 2>&1

I am asking you this because I saw weird behaviors on ony of my servers when compiling with clang. I compiled NSIMD for AARCH64 and ran the tests. I got some fails. But then when I SSHed into the server and ran the tests manually it worked fine and was successful.

[gquintin]

Re Erik,

I have ran the tests with clang several times and got no error. Still investigating.

[eschnett]

I realize that the ninja tests command before the nstest command already fails. I assume it is necessary to complete so that all tests are available?

The error reported is

$ ninja -j1 tests
[1/1738] clang++ -std=c++11 -pedantic ....modules.random.philox_2x32_10.cpp11.d
FAILED: tests.modules.random.philox_2x32_10.cpp11
clang++ -std=c++11 -pedantic /Users/eschnett/src/nsimd/build/../tests/modules/random/philox_2x32_10.cpp -Wall -Wextra -Wdouble-promotion -Wconversion -Wsign-conversion -fPIC -O2 -I/Users/eschnett/src/nsimd/build/../include -DNDEBUG -I/opt/local/include -L/opt/local/lib -Wl,-rpath,/opt/local/lib -lmpfr -DAVX2 -mavx2 -DFMA -mfma -DFP16 -mf16c -L. '-Wl,-rpath,$ORIGIN' -lnsimd_avx2 -o tests.modules.random.philox_2x32_10.cpp11 -MMD -MF tests.modules.random.philox_2x32_10.cpp11.d
In file included from /Users/eschnett/src/nsimd/build/../tests/modules/random/philox_2x32_10.cpp:29:
/Users/eschnett/src/nsimd/build/../tests/modules/random/reference.hpp:239:17: error: no matching function for call to 'mulhilo64'
  uint64_t lo = mulhilo64(0xD2B74407B1CE6E93ULL, ctr.v[0], &hi);
                ^~~~~~~~~
/Users/eschnett/src/nsimd/build/../tests/modules/random/reference.hpp:116:18: note: candidate function not viable: no known conversion from 'uint64_t *' (aka 'unsigned long long *') to 'u64 *' (aka 'unsigned long *') for 3rd argument
NSIMD_INLINE u64 mulhilo64(u64 a, u64 b, u64 *hip) {
                 ^
1 error generated.
ninja: build stopped: subcommand failed.

This is with

$ clang --version
clang version 11.0.0
Target: x86_64-apple-darwin20.1.0
Thread model: posix
InstalledDir: /opt/local/libexec/llvm-11/bin

[gquintin]

Which clang is it? (homebrew, macports, Apple)
Because Linux clang does work. I first have to get my hands on a MAC to fix this.

[eschnett]

This is MacPorts Clang, most recent version.

MacPorts GCC 10 (configured as suggested above) has the same problem. There are also two additional errors:

/opt/local/include/gcc10/c++/x86_64-apple-darwin20/bits/error_constants.h:135:24: error: 'EOWNERDEAD' was not declared in this scope
  135 |       owner_dead =     EOWNERDEAD,
      |                        ^~~~~~~~~~
/opt/local/include/gcc10/c++/x86_64-apple-darwin20/bits/error_constants.h:151:34: error: 'ENOTRECOVERABLE' was not declared in this scope
  151 |       state_not_recoverable =    ENOTRECOVERABLE,
      |                                  ^~~~~~~~~~~~~~~

[gquintin]

I will be able to fix the Clang error but the GCC one... I am not sure...

[gquintin]

Installing Xcode and macports on an old Mac. It is slow....

[eschnett]

It turns out the GCC errors are known to GCC https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93151 and to MacPorts, and will likely be fixed there at some point.

[eschnett]

I just came across a web page that said that uint64_t is unsigned long long on Linux, but is unsigned long on macOS. Maybe there is a mismatch between u64 and uint64_t?

This change seems to help:

$ git diff
diff --git a/tests/modules/random/reference.hpp b/tests/modules/random/reference.hpp
index cb47ea3..bd28bff 100644
--- a/tests/modules/random/reference.hpp
+++ b/tests/modules/random/reference.hpp
@@ -235,8 +235,8 @@ NSIMD_INLINE tab64x1_t _philox2x64bumpkey(tab64x1_t key) {
 }

 NSIMD_INLINE tab64x2_t _philox2x64round(tab64x2_t ctr, tab64x1_t key) {
-  uint64_t hi;
-  uint64_t lo = mulhilo64(0xD2B74407B1CE6E93ULL, ctr.v[0], &hi);
+  u64 hi;
+  u64 lo = mulhilo64(0xD2B74407B1CE6E93ULL, ctr.v[0], &hi);
   tab64x2_t out = {{hi ^ key.v[0] ^ ctr.v[1], lo}};
   return out;
 }

[gquintin]

I suspected as much.

[eschnett]

Yes, with these changes I can build all tests.

[eschnett]

The errors appear random. Here are results from three runs:

-- SUMMARY: 48 fails out of 8214 tests
-- FAILED: ./tests.allocator.cpp98
-- FAILED: ./tests.c_base.fms.u64.c99
-- FAILED: ./tests.c_base.fnma.u8.c99
-- FAILED: ./tests.c_base.gt.f16.c89
-- FAILED: ./tests.c_base.load2a_ravel.u8.c99
-- FAILED: ./tests.c_base.load3a.i32.c89
-- FAILED: ./tests.c_base.load3u_ravel.f64.c99
-- FAILED: ./tests.c_base.load4u_ravel.i16.c89
-- FAILED: ./tests.c_base.nbtrue.f32.c89
-- FAILED: ./tests.c_base.neg.f64.c99
-- FAILED: ./tests.c_base.reinterpret.f64_to_f64.c89
-- FAILED: ./tests.c_base.reinterpret.i8_to_u8.c89
-- FAILED: ./tests.c_base.reinterpretl.f16_to_i16.c89
-- FAILED: ./tests.c_base.ziplo.i16.c99
-- FAILED: ./tests.cxx_adv.load4u.f64.cpp11
-- FAILED: ./tests.cxx_adv.min.u64.cpp11
-- FAILED: ./tests.cxx_adv.ne.u64.cpp98
-- FAILED: ./tests.cxx_adv.orb.i64.cpp11
-- FAILED: ./tests.cxx_adv.orl.f32.cpp98
-- FAILED: ./tests.cxx_adv.reinterpretl.i8_to_u8.cpp98
-- FAILED: ./tests.cxx_adv.shr.i32.cpp11
-- FAILED: ./tests.cxx_adv.unziplo.u16.cpp98
-- FAILED: ./tests.cxx_base.eq.f32.cpp98
-- FAILED: ./tests.cxx_base.iota.i32.cpp98
-- FAILED: ./tests.cxx_base.masko_loadu1.u32.cpp98
-- FAILED: ./tests.cxx_base.maskz_loadu1.i64.cpp98
-- FAILED: ./tests.cxx_base.notb.f32.cpp11
-- FAILED: ./tests.cxx_base.shl.u64.cpp11
-- FAILED: ./tests.cxx_base.unziplo.i64.cpp11
-- FAILED: ./tests.cxx_base.zip.i32.cpp98
-- FAILED: ./tests.cxx_base.ziphi.i64.cpp98
-- FAILED: ./tests.modules.fixed_point.andl.fp_8_2.cpp11
-- FAILED: ./tests.modules.fixed_point.andl.fp_8_6.cpp11
-- FAILED: ./tests.modules.fixed_point.max.fp_8_6.cpp98
-- FAILED: ./tests.modules.fixed_point.min.fp_8_1.cpp98
-- FAILED: ./tests.modules.fixed_point.mul.fp_4_3.cpp11
-- FAILED: ./tests.modules.fixed_point.orb.fp_16_7.cpp11
-- FAILED: ./tests.modules.fixed_point.rec.fp_4_4.cpp98
-- FAILED: ./tests.modules.spmd.abs.i64.cpp11
-- FAILED: ./tests.modules.spmd.gt.u64.cpp98
-- FAILED: ./tests.modules.spmd.notl.u64.cpp98
-- FAILED: ./tests.modules.spmd.orl.u16.cpp98
-- FAILED: ./tests.modules.spmd.reinterpret.f16_u16.cpp98
-- FAILED: ./tests.modules.tet1d.div.u16.cpp11
-- FAILED: ./tests.modules.tet1d.le.f64.cpp98
-- FAILED: ./tests.modules.tet1d.lt.i32.cpp11
-- FAILED: ./tests.modules.tet1d.mul.f64.cpp98
-- FAILED: ./tests.modules.tet1d.round_to_even.u16.cpp98



-- SUMMARY: 18 fails out of 8214 tests
-- FAILED: ./tests.c_base.abs.u64.c99
-- FAILED: ./tests.c_base.fnma.f32.c89
-- FAILED: ./tests.c_base.ge.f16.c89
-- FAILED: ./tests.c_base.to_mask.u64_to_u64.c99
-- FAILED: ./tests.c_base.upcvt.u16_to_f32.c99
-- FAILED: ./tests.cxx_adv.le.f16.cpp11
-- FAILED: ./tests.cxx_adv.upcvt.f32_to_i64.cpp98
-- FAILED: ./tests.cxx_adv.upcvt.f32_to_u64.cpp98
-- FAILED: ./tests.cxx_base.cvt.u32_to_u32.cpp98
-- FAILED: ./tests.cxx_base.eq.i8.cpp11
-- FAILED: ./tests.cxx_base.gather.i64.cpp98
-- FAILED: ./tests.cxx_base.iota.i64.cpp11
-- FAILED: ./tests.cxx_base.notl.u64.cpp11
-- FAILED: ./tests.modules.fixed_point.sub.fp_4_4.cpp11
-- FAILED: ./tests.modules.spmd.notl.f32.cpp11
-- FAILED: ./tests.modules.spmd.to_mask.i16.cpp98
-- FAILED: ./tests.modules.tet1d.max.u32.cpp98
-- FAILED: ./tests.modules.tet1d.mul.f64.cpp98



-- SUMMARY: 18 fails out of 8214 tests
-- FAILED: ./tests.c_base.iota.f64.c99
-- FAILED: ./tests.c_base.load2u_ravel.i8.c99
-- FAILED: ./tests.c_base.round_to_even.i64.c99
-- FAILED: ./tests.c_base.sub.u16.c99
-- FAILED: ./tests.c_base.trunc.i64.c89
-- FAILED: ./tests.cxx_adv.cvt.u64_to_i64.cpp98
-- FAILED: ./tests.cxx_base.masko_loadu1.f16.cpp11
-- FAILED: ./tests.cxx_base.reinterpretl.u8_to_i8.cpp11
-- FAILED: ./tests.cxx_base.shl.i64.cpp98
-- FAILED: ./tests.cxx_base.unziplo.f64.cpp98
-- FAILED: ./tests.modules.fixed_point.add.fp_4_2.cpp98
-- FAILED: ./tests.modules.fixed_point.fma.fp_16_2.cpp98
-- FAILED: ./tests.modules.spmd.cvt.u64_f64.cpp11
-- FAILED: ./tests.modules.spmd.fms.u8.cpp98
-- FAILED: ./tests.modules.spmd.lt.u8.cpp98
-- FAILED: ./tests.modules.spmd.mul.u32.cpp98
-- FAILED: ./tests.modules.tet1d.ceil.i16.cpp98
-- FAILED: ./tests.modules.tet1d.min.i16.cpp11

Luckily, the errors all go away when running in serial, i.e. without the '-j' option.

[gquintin]

Hi Erik, currently compiling on our MAC to debug all this.

[gquintin]

Hi Erik, I fixed and pushed the "uint64_t" bug. I found more and will replace them all. I keep this issue open as I have like you the behavior on MAC.

[gquintin]

@eschnett, I am trying on a MAC M1 and got the same errors as you with homebrew:

In file included from /opt/homebrew/Cellar/gcc/10.2.0_3/include/c++/10.2.1/system_error:39,
                 from /opt/homebrew/Cellar/gcc/10.2.0_3/include/c++/10.2.1/bits/ios_base.h:46,
                 from /opt/homebrew/Cellar/gcc/10.2.0_3/include/c++/10.2.1/ios:42,
                 from /opt/homebrew/Cellar/gcc/10.2.0_3/include/c++/10.2.1/ostream:38,
                 from /Users/gquintin/nsimd/build/../include/nsimd/cxx_adv_api.hpp:29,
                 from /Users/gquintin/nsimd/build/../tests/cxx_adv/addv.f16.cpp:33:
/opt/homebrew/Cellar/gcc/10.2.0_3/include/c++/10.2.1/aarch64-apple-darwin20/bits/error_constants.h:135:24: error: 'EOWNERDEAD' was not declared in this scope
  135 |       owner_dead =     EOWNERDEAD,
      |                        ^~~~~~~~~~
/opt/homebrew/Cellar/gcc/10.2.0_3/include/c++/10.2.1/aarch64-apple-darwin20/bits/error_constants.h:151:34: error: 'ENOTRECOVERABLE' was not declared in this scope
  151 |       state_not_recoverable =    ENOTRECOVERABLE,
      |                                  ^~~~~~~~~~~~~~~

How did you manage to compile anyway?

[gquintin]

I am falling back to Xcode for now.

[eschnett]

I installed gcc via Spack, and this works for me.

[gquintin]

@eschnett I have fixed the issue. I just learnt at my expense that multithreading + fork do not play well together... Back to basics with simple C programming. You can test whether it works on your mac with the following:

git clone -b v2.2 https://github.com/agenium-scale/nsimd.git
export NSTOOLS_CHECKOUT_LAST_COMMIT=1
bash build-tests for sse2/sse42/avx/avx2 with gcc
cd build-avx2-gcc
../nstools/bin/nstest -j40

I have done the POSIX part, I am now debugging the Win32 part.

[gquintin]

All fixes have been done to nstest, I have tested it so it should work on Linux, Windows and MacOS. I am closing this issue.