crash on Linux x86_64 (cmake build, git 8f7e986184afb)

[ayounes-nviso]

Hello,

With latest develop (8f7e986) with cmake build I can reproduce a crash on various ubuntu platform (16.04, 17.10 and 18.04); there is no crash on an earlier version (02ef20a), here is the backtrace:

Thread 1 "nv3dfi_video_cl" received signal SIGILL, Illegal instruction.
0x00007fffe4b11383 in sgemm_nn (args=0x7fffffffc4c0, range_m=0x0, range_n=0x0,
sa=0x7fffdcf1d000, sb=0x7fffdd03d000, dummy=0)
at src/driver/level3/level3.c:254
254 if ( alpha[0] == ZERO
(gdb) bt
#0 0x00007fffe4b11383 in sgemm_nn (args=0x7fffffffc4c0, range_m=0x0, range_n=0x0, sa=0x7fffdcf1d000, sb=0x7fffdd03d000, dummy=0)
at src/driver/level3/level3.c:254
#1 0x00007fffe4b1112c in cblas_sgemm (order=CblasRowMajor, TransA=CblasNoTrans, TransB=CblasNoTrans, m=10, n=4200, k=27, alpha=1, a=0x555555d22340, lda=27, b=0x555556236110, ldb=4200, beta=0, c=0x5555561e1b90, ldc=4200)
at src/interface/gemm.c:422

[ayounes-nviso]

In Release mode I have a different backtrace.
additional info: I am compiling and using a static library libopenblas.a

Program received signal SIGILL, Illegal instruction.
0x00007fffe4d27fa2 in alloc_mmap ()
(gdb) bt
#0 0x00007fffe4d27fa2 in alloc_mmap ()
#1 0x00007fffe4d283ab in blas_memory_alloc ()
#2 0x00007fffe4cec06f in gotoblas_init ()
#3 0x00007ffff7de5733 in call_init (env=0x7fffffffdbb8, argv=0x7fffffffdb88, argc=5, l=) at dl-init.c:72
#4 0x00007ffff7de5733 in _dl_init (main_map=0x7ffff7ffe170, argc=5, argv=0x7fffffffdb88, env=0x7fffffffdbb8) at dl-init.c:119
#5 0x00007ffff7dd60ca in _dl_start_user () at /lib64/ld-linux-x86-64.so.2

[martin-frbg]

What is your hardware ? Luckily "good" and "bad" versions appear to be only six days apart so this should be easy to track down, but the only changes directly affecting SGEMM were for SkylakeX only.

[ayounes-nviso]

Currently on old Dell laptop, Sandy Bridge I think:
Intel(R) Core(TM) i7-2720QM CPU @ 2.20GHz

[martin-frbg]

Weird. Both compiling and running happens on this system ? (Slight hope that you might have built the library on Haswell without including DYNAMIC_ARCH support for other/older processors)

[brada4]

https://valid.x86.fr/ynbz2s
extmodel is not handled, but should be Sandybridge from the picture?

[ayounes-nviso]

Yes, compiling and running happens on this same system and DYNAMIC_ARCH is set to ON.

But the same crash happened on another system (Xeon or something) so I don't think this crash is specific to Sandybridge.

My guess is that it is related to the recent cmake changes for x64 ?

[ayounes-nviso]

The crash also happened yesterday on those platforms:
Docker running on i7 9xx (Nehalem Class Core i7), gcc-6
Ubuntu 17.10 running on i7-6500U CPU (Skylake?), gcc-7

[martin-frbg]

Err, it is now dawning on me that the change to add -march-skylakex-avx512 "where required" may actually be adding it unconditionally rather than for that specific target...wonder why this did not break the CI builds though. Could you try removing the two lines from system_check.cmake that I added in #1798 ?

[ayounes-nviso]

If I remove those 2 lines I get back the compilation failure (see #1797)

[martin-frbg]

Duh. But these lines actually belong in system.cmake, like

if (DEFINED TARGET AND ${TARGET} STREQUAL "SKYLAKEX" AND NOT NO_AVX512)
  set (CCOMMON_OPT "${CCOMMON_OPT} -march=skylake-avx512")
  set (FCOMMON_OPT "${FCOMMON_OPT} -march=skylake-avx512")
endif()

somewhere around line 42, before the if (DEFINED TARGET)
At least I think so until proven otherwise ...