inspectrum/liquid-dsp illegal instruction crash on Linux x86-64 build & recent Intel CPUs

[miek]

I was trying to run inspectrum from radioconda-2025.03.14-Linux-x86_64 and it crashes immediately with:

Illegal instruction (core dumped)

Narrowing it down some more, the crash is in libliquid on a vcvttss2usi instruction:

(gdb) run
Starting program: /home/mike/radioconda/bin/inspectrum 

This GDB supports auto-downloading debuginfo from the following URLs:
  <https://debuginfod.ubuntu.com>
Enable debuginfod for this session? (y or [n]) y
Debuginfod has been enabled.
To make this setting permanent, add 'set debuginfod enabled on' to .gdbinit.
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff34246c0 (LWP 1809923)]
[New Thread 0x7ffff2bec6c0 (LWP 1809924)]

Thread 1 "inspectrum" received signal SIGILL, Illegal instruction.
0x00007ffff7967388 in estimate_req_filter_len () from /home/mike/radioconda/bin/../lib/libliquid.so.1
(gdb) bt
#0  0x00007ffff7967388 in estimate_req_filter_len () from /home/mike/radioconda/bin/../lib/libliquid.so.1
#1  0x000055555558848b in SpectrogramPlot::getTunerTaps() ()
#2  0x0000555555588647 in SpectrogramPlot::tunerMoved() ()
#3  0x00007ffff6bdffbf in ?? () from /home/mike/radioconda/bin/../lib/libQt5Core.so.5
#4  0x00005555555888a1 in SpectrogramPlot::setFFTSize(int) ()
#5  0x000055555557dc12 in PlotView::setFFTAndZoom(int, int) ()
#6  0x00007ffff6bdffbf in ?? () from /home/mike/radioconda/bin/../lib/libQt5Core.so.5
#7  0x000055555556fa46 in SpectrogramControls::fftOrZoomChanged(int, int) ()
#8  0x0000555555586601 in SpectrogramControls::setDefaults() ()
#9  0x0000555555574b00 in MainWindow::MainWindow() ()
#10 0x000055555556f0db in main ()
(gdb) disassemble
Dump of assembler code for function estimate_req_filter_len:
   0x00007ffff7967360 <+0>:	sub    $0x8,%rsp
   0x00007ffff7967364 <+4>:	vcomiss 0x5a864(%rip),%xmm0        # 0x7ffff79c1bd0
   0x00007ffff796736c <+12>:	ja     0x7ffff7967390 <estimate_req_filter_len+48>
   0x00007ffff796736e <+14>:	vxorps %xmm2,%xmm2,%xmm2
   0x00007ffff7967372 <+18>:	vcomiss %xmm0,%xmm2
   0x00007ffff7967376 <+22>:	jae    0x7ffff7967390 <estimate_req_filter_len+48>
   0x00007ffff7967378 <+24>:	vcomiss %xmm1,%xmm2
   0x00007ffff796737c <+28>:	jae    0x7ffff79673c0 <estimate_req_filter_len+96>
   0x00007ffff796737e <+30>:	call   *0x92dcc(%rip)        # 0x7ffff79fa150
   0x00007ffff7967384 <+36>:	add    $0x8,%rsp
=> 0x00007ffff7967388 <+40>:	vcvttss2usi %xmm0,%eax
   0x00007ffff796738e <+46>:	ret
   0x00007ffff796738f <+47>:	nop
   0x00007ffff7967390 <+48>:	mov    $0x50,%edx
   0x00007ffff7967395 <+53>:	mov    $0x3,%edi
   0x00007ffff796739a <+58>:	mov    $0x1,%eax
   0x00007ffff796739f <+63>:	vcvtss2sd %xmm0,%xmm0,%xmm0
   0x00007ffff79673a3 <+67>:	lea    0x6156e(%rip),%rcx        # 0x7ffff79c8918
   0x00007ffff79673aa <+74>:	lea    0x619d2(%rip),%rsi        # 0x7ffff79c8d83
   0x00007ffff79673b1 <+81>:	call   *0x94c39(%rip)        # 0x7ffff79fbff0
   0x00007ffff79673b7 <+87>:	xor    %eax,%eax
   0x00007ffff79673b9 <+89>:	add    $0x8,%rsp
   0x00007ffff79673bd <+93>:	ret
   0x00007ffff79673be <+94>:	xchg   %ax,%ax
   0x00007ffff79673c0 <+96>:	mov    $0x53,%edx
   0x00007ffff79673c5 <+101>:	mov    $0x3,%edi
   0x00007ffff79673ca <+106>:	mov    $0x1,%eax
   0x00007ffff79673cf <+111>:	vcvtss2sd %xmm1,%xmm1,%xmm0
   0x00007ffff79673d3 <+115>:	lea    0x61506(%rip),%rcx        # 0x7ffff79c88e0
   0x00007ffff79673da <+122>:	lea    0x619a2(%rip),%rsi        # 0x7ffff79c8d83
   0x00007ffff79673e1 <+129>:	call   *0x94c09(%rip)        # 0x7ffff79fbff0
   0x00007ffff79673e7 <+135>:	xor    %eax,%eax
   0x00007ffff79673e9 <+137>:	jmp    0x7ffff79673b9 <estimate_req_filter_len+89>
End of assembler dump.

This is apparently an AVX-512 instruction, but support for that set has been removed from most recent Intel CPUs (I'm using an i7-13700K - Raptor Lake) so I think the library probably should be built without AVX-512 for good compatibility.

I noticed that conda info includes virtual packages : __archspec=1=skylake which I guess is autodetected from the build runner? I wasn't sure whether it's possible to override that at a global level or if this would have to be patched in the liquid-dsp build, so I opened the issue here rather than on liquid-dsp-feedstock for now.

[ryanvolz]

Thanks for the report, and the diagnosis! It does sound like a rebuild of liquid-dsp is in order. The __archspec=1=skylake virtual package is what conda detects that your machine is compatible with. This is a somewhat new feature, but it means when I rebuild liquid-dsp I should be able to build multiple versions that have compatibility with different CPU instruction sets. The current build, requiring AVX-512, would require __archspec=4, which certainly limits compatibility!

[ryanvolz]

I have updated the liquid-dsp package so that it builds with different levels of instruction sets, with the appropriate one to be chosen by the user's __archspec virtual package. It remains to get the updated package into a release of radioconda, with the installer bundling the version with widest compatibility.

Until then, upgrading the libliquid1 package manually should fix running inspectrum. I think this will do it:

mamba update libliquid1

But you may need

mamba install _x86_64-microarch-level
mamba update libliquid1

[miek]

Yep, that's working now, thanks!