<_whitenotifier-3>
[scopehal] azonenberg 562d994 - Removed vestigial TDRStepDeEmbed and OFDMDemodulator filters as they were pretty much dead code that didn't work or do anything close to what it needed to
<_whitenotifier-3>
[scopehal] azonenberg 3a1bcc0 - FFTFilter: removed CPU fallback implementation
<_whitenotifier>
[scopehal] azonenberg 831aab1 - FFTFilter: Removed AVX versions of FFT output processing in preparation for going full Vulkan
<_whitenotifier-3>
[scopehal-apps] azonenberg 49f8f2c - Removed references to AVX2 FFT acceleration. Removed unit test for AVX2 accelerated Blackman-Harris window that no longer exists.
<_whitenotifier>
[scopehal-apps] azonenberg 269780b - Filter_FFT test: added CPU implementation moved out from FFTFilter class
<azonenberg>
And with that, all of our filters are now FFTS-free
<d1b2>
<david.rysk> yay finally
<azonenberg>
The codebase as a whole is not
<d1b2>
<david.rysk> do they all have Vulkan implementations where that's helpful?
<azonenberg>
TestWaveformSource still uses FFTS (in the demo oscilloscope, i dont think anything else uses that) so we have to make that use vkFFT
<azonenberg>
and then all of the FFT-ish unit tests use FFTS, but those are planned to transition to FFTW since it's OK if the test binaries which we don't redistribute become GPL
<azonenberg>
so FFTW will be required for devs building unit tests but not end users, and FFTS will be completely removed
<azonenberg>
vkFFT will be the only shipped FFT lib in the application itself
<azonenberg>
so all of our ffts will use it
<azonenberg>
And we do not have vulkan implementations of every potentially acceleratable filter, but we do of everything that uses FFT
<azonenberg>
At least, of the FFT code and immediate postprocessing
<azonenberg>
Right now we do FFT peak detection on the CPU, which means after FFT-ing we have to immediately pull the waveform off the GPU and onto the CPU to do the peak search (and our peak search algorithm sucks)
<azonenberg>
Making a better, GPU accelerated peak detection algorithm is on the wishlist but i have too many things going on to actively work on it
<azonenberg>
but we're at the point that peak detection takes longer than the FFT proper for smaller datasets
sgstair has quit [Read error: Connection reset by peer]