<d1b2>
<Mughees> Ok. And what if I want to discard output stream (i.e. output stream doesn't get displayed if input is digital)?
<azonenberg>
@mughees: just set the data to nullptr
<azonenberg>
or do you mean eliminate the stream entirely?
<azonenberg>
Dynamic stream behavior like this is relatively new territory
<azonenberg>
So we havent worked out all of the corners of what isn/isn't possible or should be done
<d1b2>
<Mughees> Yeah Entirely
<azonenberg>
ditto for things like feedback paths in the filter graph. I get that there's benefits for control systems but everything was architected around a linearized DAG
<azonenberg>
i'd like to think about how to make it work well but it's something we need to think about how to do well
<d1b2>
<Mughees> Ok for now for my application, i want amplitude of pulses plotted which doesn’t make sense for digital signal inputs.
<d1b2>
<Mughees> I could simply pass the input as output for this case as for digital signals a plot of pulses amplitude is technically equal to the signal itself
<d1b2>
<Mughees> So im kind of good
<d1b2>
<Mughees> On a separate note, what work related to scalar type u have in mind
<azonenberg>
Oh, the stuff i suggested weeks ago? I already implemented that because I needed it for something I was doing lol
<azonenberg>
I guess there is still one filter that would be useful to have, an array indexing filter
<azonenberg>
returns the Nth element of an analog waveform, if it exists, as a scalar
<azonenberg>
(given N as a scalar)
<azonenberg>
@mughees how are you at PLL design btw?
<azonenberg>
Our CDR PLL filter needs to be replaced with a better PLL that has specific bandwidth characteristics etc
<d1b2>
<Mughees> Never done anything with it. Except setup plls in a microcontroller for embedded software programming 😊
<azonenberg>
Well then you probably aren't the guy we should throw at that filter then :)
<d1b2>
<Mughees> Yeah .. I’ll discuss with Louis and M for whats next
<azonenberg>
Gimme 5 mins to go through the list of tickets and i'll suggest some things
<azonenberg>
There are a few that i'd love to do that are blocked on other things
<azonenberg>
somebody submitted a PR for a decode that outputted the data as digital samples but i think an analog waveform makes more sense. or maybe even both
<azonenberg>
(also said PR is old and has bitrotted probably... but feel free to use it as a starting point)
<d1b2>
<Mughees> Any idea why I may be getting this error after I keep a filter displayed for some time (20-30 secs): terminate called after throwing an instance of 'vk::DeviceLostError' what(): vk::Device::waitForFences: ErrorDeviceLost Aborted (core dumped)
<azonenberg>
Device lost error usually indicates the GPU stopped responding for some reason
<azonenberg>
and the driver reset it
<azonenberg>
I've seen it here and there but havent root caused it yet
<d1b2>
<Mughees> ok. It becomes stable if sample depth is kept low
<d1b2>
<Mughees> for example 10ks intead of 100ks
<azonenberg>
yeah that suggests you might be bumping into a timeout somewhere in the GPU driver where a shader takes too long or something
<azonenberg>
but 100k should not be timing out
<azonenberg>
100M maybe on a low end integrated GPU
<azonenberg>
There are some bugs i need to fix wrt intel integrated GPUs and how some of the shader coordinate math is done
<azonenberg>
what card are you using?
Bird|otherbox has quit [Ping timeout: 265 seconds]
<d1b2>
<Mughees> i have a nvidia one too...but somehow ubuntu picked this one up
<d1b2>
<david.rysk> are you running on the intel or the nvidia? if nvidia, are you running the nvidia driver or the nouveau driver?
<azonenberg>
Can you test and let me know if you are seeing problems specifically at memory depths over 65536 or 2097152 samples?
<d1b2>
<Mughees> i did not bother to install nvidia
<azonenberg>
(2^16 and 32*2^16 respectively)
<d1b2>
<david.rysk> probably safe to say that nouveau will have bugs especially on newer cards
<d1b2>
<Mughees> ok
<azonenberg>
@david.rysk its almost certainly an optimus setup using the intel by default unless you request use of the NV
<azonenberg>
for a specific app
<azonenberg>
@mughees so... the intel vulkan driver has a max X axis thread count of 2^16
<azonenberg>
Bad Things are likely to happen if you do anything on one that tries to launch more than 2^16 thread blocks
<azonenberg>
typically we run 32 threads per block so 32 * 2^16 is where i would expect to have problems
<azonenberg>
we have some shaders that assume you can run arbitrarily many thread blocks in one axis; these need to be refactored to run on a 2D grid instead of 1D
<azonenberg>
That said, i would expect this to be an immediate failure
<azonenberg>
the first time the shader runs vulkan would complain you asked for an invalid configuration
<azonenberg>
it should not cause a hang or eventual device lost error, which leads me to believe this is not the problem you are hitting
Bird|otherbox has joined #scopehal
<d1b2>
<Mughees> 65536 fails, 65540 fails, 65520 fails, 30000 works
<d1b2>
<Mughees> I am a GPU virgin guys,,, English please 😄
<d1b2>
<Mughees> Failure becomes very quick if I drag some waveform into a eparate window
<azonenberg>
This seems like it's a different bug then
<azonenberg>
Maybe poke Louis and have him take a look?
<d1b2>
<Mughees> yeah will bug him....makes taking screenshots for documentation frustrating....
<d1b2>
<Mughees> My logic is that sticking with intel integrated should be more stable
<azonenberg>
Nvidia is our most well tested platform
<d1b2>
<david.rysk> over nouveau yes, over nvidia no
<azonenberg>
the nvidia binary driver that is
<azonenberg>
it's what i use and i think most of the other core devs
<azonenberg>
more specifically 64-bit x86 CPUs, linux, and nvidia GPU with binary driver
<d1b2>
<Mughees> what if poor people need to use glscopeclinet 🙂
<azonenberg>
I'm not saying i dont want to support other hardware
<azonenberg>
Just that this is what I have, so bugs that hit that hardware get noticed and fixed fast
<d1b2>
<david.rysk> those bugs need to be tracked and fixed 🙂
<azonenberg>
I agree they need to be fixed
<azonenberg>
The hardest part is getting a dev who knows GPU stuff to be able to reproduce it
<azonenberg>
e.g. my work laptop has an intel integrated GPU and an nvidia card, and seems to run ngscopeclient fairly reliably on eithert
<azonenberg>
(as long as you dont hit the max waveform size for the intel card, wrt that shader limit i mentioned above)
<azonenberg>
but that's in the megapoint range not tens of kpoints
<d1b2>
<Mughees> 50Ks also working good
<azonenberg>
The number being close to 64K does suggest something funky is happening with shader block sizes, but i cant debug it without seeing it happen in front of me
<d1b2>
<david.rysk> these kinds of bugs can be very dependent on gpu model and driver version
<azonenberg>
Exactly
<azonenberg>
So the hardest bit about fixing it is probably getting that exact hardware setup in front of a dev
<azonenberg>
(a dev with GPU experience rather than someone writing filters who doesnt know vulkan)
<d1b2>
<Mughees> yeah GPU is definitely complex stuff