AVX2 disabled on Windows as workaround for gcc bug

[sbabic44]

On Keysight 3000T, software throws shader error and quits. PC is Windows, Intel 4600 HD, OpenGL 4.3. latest driver.. Errors in att..

Error 450.txt

Tried changing GLSL version: 450 to 430

Error 430.txt

[azonenberg]

Looks like the 4600 HD doesn't support 64-bit integer arithmetic (GL_ARB_gpu_shader_int64). Do you have a machine with a discrete graphics card you can try?

Intel integrated cards have been lagging in support for a while, but even fairly old NVidia/AMD discrete GPUs should work.

[sbabic44]

Yes, I'm afraid I will need to upgrade.. Will try on laptop it has Nvidia.. And that was freaky fast answer, respect !! Anyways, I'll try and report back.. Reagrds,

[sbabic44]

Yes, that was that, i managed to start it, Now I get other error: GetConsoleScreenBufferInfo() failed (6); When I try to start hystogram or FFT or any graphical analysis. Statistics and text measurements work fine.

[azonenberg]

This is on Windows, I assume? Can you post a screenshot of the error showing both the glscopeclient window and the console?

[bvernoux]

I confirm on my side glscopeclient "crash" with FFT too with Windows build it seems related to alignment to be confirmed... https://github.com/anthonix/ffts is used in glscopeclient and is known to have some https://github.com/anthonix/ffts/issues in fact and seems dead (not maintained) since few years, latest commit was on Jun 17, 2017. So far I was searching an alternative but it is not easy to find something good fast/multi-platform with open license like MIT or BSD license... For information the https://github.com/linkotec/ffts fork which is more recent does not build correctly with MSYS2 mingw64 ...

[sbabic44]

Yes it crashes with Segmentation fault:

sinisa@t440p MINGW64 ~ $ glscopeclient scoop:agilent:lan:3104t.:5025 GetConsoleScreenBufferInfo() failed (6); Warning: glscopeclient works best with the OMP_WAIT_POLICY environment variable set to PASSIVE Segmentation fault

sinisa@t440p MINGW64 ~

Main screen closes, nothing to see there.

I tried FFT, histograms, and few more and it crashes.

[bvernoux]

This issue shall be probably renamed as it is an issue with FFTS for Windows with MSYS2 mingw64 it seems it does not affect GNU/Linux to be confirmed I have fixed that issue it was an issue to compute the next highest power of 2 see https://github.com/azonenberg/scopehal/pulls So now all crash related to ffts are fixed it was impacting following scopehal classes:

• JitterSpectrumFilter
• FFTFilter
• DeEmbedFilter
• TestWaveformSource (used in demo mode)

[bvernoux]

I confirm this issue is fixed now with latest master since https://github.com/azonenberg/scopehal-apps/commit/27078c095a68189306fa3baed975ac4610fcd380

[azonenberg]

@sbabic44 can you please test and confirm this is fixed?

Also, I added a backward compatibility version of the shaders which does not require GL_ARB_gpu_shader_int64, please test on your 4600 HD and see if it works there too?

[sbabic44]

Andrew, thanks for the effort. It actually started with this on command line:

`sinisa@WKS MINGW64 ~/ffts/build/scopehal-apps/build $ glscopeclient --debug myscope:demo:null:null GetConsoleScreenBufferInfo() failed (6); Warning: glscopeclient works best with the OMP_WAIT_POLICY environment variable set to PASSIVE Detecting CPU features... * AVX2

Warning: Warning: Can't parse preference value 13,370000 for preference hidden_s etting, ignoringWarning: Warning: Can't parse preference value 42,090000 for preference test_rea l, ignoring ` If I try to create FFT on one of channels, or create histogram it still drops out with no warning....

But shaders work with 4600HD on win10 now just fine..

Best regards

[azonenberg]

So looks like we still have a problem.

Additionally, it seems like the preference parsing is getting messed up by locales that uses commas as a decimal separator. Let me file a separate ticket for that.

[azonenberg]

@sbabic44 Can you try with latest code and see how it works? Try with and without the --noopencl command line argument.

[someone--else]

On rev 3e9523c (current master) / Windows segfault happens immediately after doing RF->FFT on any demo scope waveform (see call stack below)

Console output:

Detecting CPU features... * AVX2 OpenCL support: not present at compile time. GPU acceleration disabled. Context: OpenGL 4.2 compatibility profile GL_VENDOR = NVIDIA Corporation GL_RENDERER = Quadro T1000/PCIe/SSE2 GL_VERSION = 4.2.0 NVIDIA 452.66 GL_SHADING_LANGUAGE_VERSION = 4.20 NVIDIA via Cg compiler Initial GL error code = 0 GL_ARB_gpu_shader_int64 = supported

Call stack: (segfault happens on this line: __m256 vcos = _mm256_cos_ps(vscale); at i == 0, data/out seems aligned properly)

libscopeprotocols.dll!FFTFilter::CosineSumWindowAVX2(const float * data, size_t len, float * out, float alpha0) (c:\apps\msys64\home\root\gls\scopehal-apps\lib\scopeprotocols\FFTFilter.cpp:714) libscopeprotocols.dll!FFTFilter::HammingWindow(const float * data, size_t len, float * out) (c:\apps\msys64\home\root\gls\scopehal-apps\lib\scopeprotocols\FFTFilter.cpp:822) libscopeprotocols.dll!FFTFilter::ApplyWindow(const float * data, size_t len, float * out, FFTFilter::WindowFunction func) (c:\apps\msys64\home\root\gls\scopehal-apps\lib\scopeprotocols\FFTFilter.cpp:672) libscopeprotocols.dll!FFTFilter::DoRefresh(FFTFilter * const this, AnalogWaveform * din, std::vector<EmptyConstructorWrapper, AlignedAllocator<EmptyConstructorWrapper, 64> > & data, double fs_per_sample, size_t npoints, size_t nouts, bool log_output) (c:\apps\msys64\home\root\gls\scopehal-apps\lib\scopeprotocols\FFTFilter.cpp:439) libscopeprotocols.dll!FFTFilter::Refresh(FFTFilter * const this) (c:\apps\msys64\home\root\gls\scopehal-apps\lib\scopeprotocols\FFTFilter.cpp:315) libscopehal.dll!Filter::RefreshIfDirty(Filter * const this) (c:\apps\msys64\home\root\gls\scopehal-apps\lib\scopehal\Filter.cpp:191) OscilloscopeWindow::_ZN18OscilloscopeWindow17RefreshAllFiltersEv._omp_fn.0(void)() (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\OscilloscopeWindow.cpp:3081) libgomp-1.dll![Unknown/Just-In-Time compiled code] (Unknown Source:0) OscilloscopeWindow::RefreshAllFilters(OscilloscopeWindow * const this) (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\OscilloscopeWindow.cpp:3079) OscilloscopeWindow::OnAllWaveformsUpdated(OscilloscopeWindow * const this, bool reconfiguring) (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\OscilloscopeWindow.cpp:2913) OscilloscopeWindow::PollScopes(OscilloscopeWindow * const this) (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\OscilloscopeWindow.cpp:2852) OscilloscopeWindow::OnTimer(OscilloscopeWindow * const this) (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\OscilloscopeWindow.cpp:493) sigc::bound_mem_functor1<bool, OscilloscopeWindow, int>::operator()(const sigc::bound_mem_functor1<bool, OscilloscopeWindow, int> * const this, sigc::type_trait_take_t _A_a1) (c:\apps\msys64\mingw64\include\sigc++-2.0\sigc++\functors\mem_fun.h:2066) sigc::adaptor_functor<sigc::bound_mem_functor1<bool, OscilloscopeWindow, int> >::operator()<int&>(const sigc::adaptor_functor<sigc::bound_mem_functor1<bool, OscilloscopeWindow, int> > * const this, int & _A_arg1) (c:\apps\msys64\mingw64\include\sigc++-2.0\sigc++\adaptors\adaptor_trait.h:89) sigc::bind_functor<-1, sigc::bound_mem_functor1<bool, OscilloscopeWindow, int>, int, sigc::nil, sigc::nil, sigc::nil, sigc::nil, sigc::nil, sigc::nil>::operator()(sigc::bind_functor<-1, sigc::bound_mem_functor1<bool, OscilloscopeWindow, int>, int, sigc::nil, sigc::nil, sigc::nil, sigc::nil, sigc::nil, sigc::nil> * const this) (c:\apps\msys64\mingw64\include\sigc++-2.0\sigc++\adaptors\bind.h:1124) sigc::internal::slot_call0<sigc::bind_functor<-1, sigc::bound_mem_functor1<bool, OscilloscopeWindow, int>, int, sigc::nil, sigc::nil, sigc::nil, sigc::nil, sigc::nil, sigc::nil>, bool>::call_it(sigc::internal::slot_rep * rep) (c:\apps\msys64\mingw64\include\sigc++-2.0\sigc++\functors\slot.h:136) libglibmm-2.4-1.dll![Unknown/Just-In-Time compiled code] (Unknown Source:0) ScopeApp::run(ScopeApp * const this, std::vector<std::__cxx11::basic_string<char, std::char_traits, std::allocator >, std::allocator<std::__cxx11::basic_string<char, std::char_traits, std::allocator > > > filesToLoad, bool reconnect, bool nodata, bool retrigger, bool nodigital, bool nospectrum) (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\ScopeApp.cpp:120) main(int argc, char ** argv) (c:\apps\msys64\home\root\gls\scopehal-apps\src\glscopeclient\main.cpp:289)

[someone--else]

Turning off AVX2 via g_hasAvx2 solves the problem

Might be related to this: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412

[azonenberg]

Rather than turning off AVX2 on Windows, what happens if you add -mavx to the compile command line? AVX1 should be pretty well supported and it might change the ABI enough to make stuff align correctly.

[someone--else]

Same segfault with -mavx

I think using clang instead of gcc might help, but not sure how to set CMake up for it to test

[azonenberg]

Just checking in on this: after https://github.com/glscopeclient/scopehal/pull/484, is this issue still relevant or can we close it?

[someone--else]

Linked GCC issue seems to be still open, so turning AVX off on Windows as a temp workaround is still relevant and shouldn't be reverted. Perhaps we can have a separate issue dedicated to reverting it when GCC is fixed?

There is also a way to fix this without fixing GCC by moving all relevant stack variables to classes or heap and aligning them manually, but I don't know if it's worth it since code quality will suffer as a result

[azonenberg]

Hmm. Let me rename this issue then to reflect the true situation.

[hmartinez82]

Do Clang builds also suffer from this? If not https://github.com/ngscopeclient/scopehal/pull/966