Repository navigation
import numpy crashes at xgetbv instruction in npy_cpu_supports on a VM #15342
Description
Activity
While trying to learn more about this problem, I found a bug report at another project where there was a crash at xgetbv on a machine that had avx support disabled. The fix there was to check that the CPUID.1:ECX.OSXSAVE bit is set first. I wasn't sure if something like that might be applicable for numpy too.
I got the same error on a Google Cloud VM.
The /proc/cpuinfo I got is:processor : 3 vendor_id : GenuineIntel cpu family : 6 model : 44 model name : Westmere E56xx/L56xx/X56xx (Nehalem-C) stepping : 1 microcode : 0x1 cpu MHz : 2199.816 cache size : 4096 KB physical id : 3 siblings : 1 core id : 0 cpu cores : 1 apicid : 3 initial apicid : 3 fpu : yes fpu_exception : yes cpuid level : 11 wp : yes flags : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss syscall nx rdtscp lm constant_tsc rep_good nopl nonstop_tsc pni pclmulq dq vmx ssse3 fma cx16 pcid sse4_1 sse4_2 movbe popcnt aes avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single tpr_shadow flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle av x2 smep bmi2 erms invpcid rtm rdseed adx smap arat bugs : cpu_meltdown spectre_v1 spectre_v2 spec_store_bypass l1tf mds bogomips : 4399.63 clflush size : 64 cache_alignment : 64 address sizes : 40 bits physical, 48 bits virtual power management:So it appears that this thing does have avx and avx2 support in general, but not the
xgetbvinstruction.In my case the crash points to the specific version and line:
https://github.com/numpy/numpy/blob/v1.17.3/numpy/core/src/umath/cpuid.c#L30It seems that the
xgetbvinstruction is called without first checking whether it is available.According to "Intel(R) Advanced Vector Extensions Programming Reference", Ref. # 319433-011
Note: It is unwise for an application to rely exclusively on CPUID.1:ECX.AVX[bit 28]
or at all on CPUID.1:ECX.XSAVE[bit 26]: These indicate hardware support but not
operating system support. If YMM state management is not enabled by an operating
systems, AVX instructions will #UD regardless of CPUID.1:ECX.AVX[bit 28].@mattip, @seiko2plus does the new/proposed cpu detection code solve this issue? And @qfan is there a simple fix you can suggest to our current cpu detection functionality?
@mattip, @seiko2plus does the new/proposed cpu detection code solve this issue? And @qfan is there a simple fix you can suggest to our current cpu detection functionality?
I do not know any x86 assembly. But the "Intel(R) Advanced Vector Extensions Programming Reference" PDF (you can google it, but intel does not provide a direct link) has a 17-line pseudo-code for checking this, near the "note" I quoted in previous comment.
@ryandesign, yes that's right,
OSXSAVEbit must be checked first before executingxgetbv,
@seberg, @qfan, gh-13421 should fix this issue, please could you try it? also you can dump the new module attribute__cpu_features__that come with it. it should contain the runtime detection result.Reacted by Sebastian Berg@ryandesign, yes that's right,
OSXSAVEbit must be checked first before executingxgetbv,
@seberg, @qfan, gh-13421 should fix this issue, please could you try it? also you can dump the new module attribute__cpu_features__that come with it. it should contain the runtime detection result.I applied your pull request to numpy/master and it does work on that Westmere CPU. Great Job!
>>> import numpy >>> numpy.__cpu_features__ {'MMX': True, 'SSE': True, 'SSE2': True, 'SSE3': True, 'SSSE3': True, 'SSE41': True, 'POPCNT': True, 'SSE42': True, 'AVX': False, 'F16C': True, 'XOP': False, 'FMA4': False, 'FMA3': False, 'AVX2': False, 'AVX512F': False, 'AVX512CD': False, 'AVX512ER': False, 'AVX512PF': False, 'AVX5124FMAPS': False, 'AVX5124VNNIW': False, 'AVX512VPOPCNTDQ': False, 'AVX512VL': False, 'AVX512BW': False, 'AVX512DQ': False, 'AVX512VNNI': False, 'AVX512IFMA': False, 'AVX512VBMI': False, 'AVX512VBMI2': False, 'AVX512BITALG': False, 'AVX512_KNL': False, 'AVX512_KNM': False, 'AVX512_SKX': False, 'AVX512_CLX': False, 'AVX512_CNL': False, 'AVX512_ICL': False, 'VSX': False, 'VSX2': False, 'VSX3': False, 'NEON': False, 'NEON_FP16': False, 'NEON_VFPV4': False, 'ASIMD': False, 'FPHP': False, 'ASIMDHP': False, 'ASIMDDP': False, 'ASIMDFHM': False} >>>I'm going to lose access to this VM tomorrow, though. So I probably won't be able to test this again if you make any change later.
Reacted by Sayed Awadgh-13421 should fix this issue, please could you try it?
It does on my macOS Sierra VM, thank you.
also you can dump the new module attribute
__cpu_features__that come with it.$ python3.7 Python 3.7.6 (default, Jan 17 2020, 18:32:15) [Clang 9.0.0 (clang-900.0.39.2)] on darwin Type "help", "copyright", "credits" or "license" for more information. >>> import numpy >>> numpy.__cpu_features__ {'MMX': True, 'SSE': True, 'SSE2': True, 'SSE3': True, 'SSSE3': True, 'SSE41': True, 'POPCNT': True, 'SSE42': True, 'AVX': False, 'F16C': False, 'XOP': False, 'FMA4': False, 'FMA3': False, 'AVX2': False, 'AVX512F': False, 'AVX512CD': False, 'AVX512ER': False, 'AVX512PF': False, 'AVX5124FMAPS': False, 'AVX5124VNNIW': False, 'AVX512VPOPCNTDQ': False, 'AVX512VL': False, 'AVX512BW': False, 'AVX512DQ': False, 'AVX512VNNI': False, 'AVX512IFMA': False, 'AVX512VBMI': False, 'AVX512VBMI2': False, 'AVX512BITALG': False, 'AVX512_KNL': False, 'AVX512_KNM': False, 'AVX512_SKX': False, 'AVX512_CLX': False, 'AVX512_CNL': False, 'AVX512_ICL': False, 'VSX': False, 'VSX2': False, 'VSX3': False, 'NEON': False, 'NEON_FP16': False, 'NEON_VFPV4': False, 'ASIMD': False, 'FPHP': False, 'ASIMDHP': False, 'ASIMDDP': False, 'ASIMDFHM': False} >>>Reacted by Sayed Awad1.19.0 fixes this. Thanks.
Hi, I run the MacPorts automated build system and we have been experiencing an issue for awhile where
import numpycauses Python to crash. I don't know if this happens for any users running macOS directly on their Macs but it does happen on one of our build machines—the one running macOS Sierra 10.12. The other build machines (which run earlier and later macOS versions) don't experience the problem. All the build machines are virtual machines running under VMware ESXi 6 on 2009-model Xserves.I believe this is similar to or the same as #13059 but in case it's not I didn't want to mix up my information in that issue. In that issue, @eric-wieser speculated that
HAVE_XGETBVis being set incorrectly. As my lldb trace below shows, we are crashing at thexgetbvinstruction.@seberg asked in #13059 if the problem was the VM claiming to support AVX when it does not. I have used the sample program posted by @ulido in #10330 to check whether the machine claims to have avx support (using
__builtin_cpu_supports("avx")). In my case here in this macOS 10.12 VM, the sample program says "AVX not supported!"Reproducing code example:
Error message:
I am sorry, I am not familiar with gdb, but here's what happens under lldb:
Numpy/Python version information:
Python 3.7.6
Numpy 1.18.1
macOS 10.12.6 Build version 16G2136
Xcode 9.2 Build version 9C40b
Apple LLVM version 9.0.0 (clang-900.0.39.2)