ikarso has quit [Quit: Connection closed for inactivity]
brook has quit [Remote host closed the connection]
brook has joined #beagle
brook has quit [Remote host closed the connection]
brook has joined #beagle
brook has quit [Remote host closed the connection]
buzzmarshall has joined #beagle
lucascastro has quit [Remote host closed the connection]
brook has joined #beagle
vagrantc has joined #beagle
florian has quit [Quit: Ex-Chat]
DansDroids has joined #beagle
<DansDroids>
Hi i was wondering if anyone can help me, I am trying to find out how to add a fbtft driver for the ili9488 to my pocketbeagle, running 'sudo modprobe fbtft_device name=list; dmesg | tail -30' shows no entry for the ili9488 but I found a compatible driver at https://github.com/birdtechstep/fbtft... I have cloned the repo but am unsure how to add
<DansDroids>
the drivers to my kernel..
<zmatt>
DansDroids: here's some notes on building a customized version of the standard beaglebone kernel: https://pastebin.com/eLhrp1Hg
<zmatt>
(note: for the 4.19-ti kernel series you'll want to use branch ti-linux-4.19.y of github.com/RobertCNelson/ti-linux-kernel-dev )
<DansDroids>
thank you, i take it there are no commands and i have to rebuild the kernel from scratch?
<zmatt>
not sure what you mean by that... though, this looks like it's designed for older kernels which don't already have fbtft
<DansDroids>
ah sorry, i was hoping to find a cli command i can use to add the new folder of fbtft drivers to my existing list as the 4.19.94-ti kernel i am running only has a few drivers bundled..
<zmatt>
I mean, it's possible to build individual drivers out of tree, but I'm not sure how that works if you've just got a single file
<DansDroids>
hahaha me neither... thank you for the help, i will look over that page you sent
<zmatt>
looks like nothing has changed much though, so it may be sufficient to just copy fb_ili9488.c over to drivers/staging/fbtft/ in the kernel tree and add it to its Kconfig and Makefile
<DansDroids>
ok, that sounds alot easier! tbh kernel compiling really confuses me!! im good at python and c but for some reason i always glaze over when it comes to kernels lol
<zmatt>
right, but the repo I linked to has build scripts that will take care of all that and produces an installable debian package
<DansDroids>
ahh ok, just script based?
<zmatt>
?
<DansDroids>
i dont know, ive look over that github page for the ti series a few times and it still confuses me.. im going to move onto another project.. thank you for the help
<rob_w>
i need to apply some filters (Fir, pass, rms etc ) onto some data blocks in a fast way, can i offload such stuff via opengl into the gpu , like a shader or such onto a img ?
<zmatt>
rob_w: what hardware are you talking about?
<rob_w>
beaglebone powervr
<rob_w>
or am437x with sgx
<zmatt>
I mean, since they don't support opencl or whatever yeah your only option would be to somehow try to use (or abuse) shaders to do the computation you want
<zmatt>
keep in mind that latency will probably be substantial, so whether this is "fast" will depend on your definition of fast
<rob_w>
dont they run opengles n such
<zmatt>
I just used neon-optimized code on the cortex-a8 for FIR filtering
<zmatt>
opengl ES and ES2 yeah
<rob_w>
how is neon optimisation actually realised ?
<zmatt>
I just used C/C++ code with neon intrinsics.. I could probably have further optimized it using asm but the compiler output seemed okay enough and performed sufficiently well for our application
<zmatt>
(note: I used 32-bit signed integer arithmetic, not float)
<zmatt>
(with 64-bit accumulation)
<rob_w>
sounds nice
<zmatt>
yeah, the stereo FIR core ends up compiling down to https://pastebin.com/mKrvget9 which handles two FIR coefficients per loop iteration (for both channels at the same time)
<zmatt>
it's not super performant (we get 2.4% cpu load for 128-coeff stereo FIR) but it was good enough
<zmatt>
@ 48 kHz
<rob_w>
if you use plain c code for that example, how do they compare in speed ?
<rob_w>
but i guess i ve look and learn more .. more .. more
<zmatt>
lemme check
<rob_w>
like ballpark ,, are we talking percents or factors
<zmatt>
without neon intrinsics (i.e. relying on auto-vectorization by the compiler) I get about 40%-50% slowdown and it also doubles the size of the code output
<zmatt>
of course this is still optimized code
<rob_w>
u mean the instruciton binary size
<zmatt>
yeah the size of the .o containing the fir filtering code
<zmatt>
of course this is just what my code happens to yield, your mileage may vary
<rob_w>
well those numbers do look great
<zmatt>
well, even with intrinsics this is still almost 4 cpu cycles/coeff/(stereo)sample even though neon can theoretically dispatch a vmlal every cycle which would be 1 cycle/coeff/(stereo)sample
<zmatt>
I think loading the data is the limiting factor
otisolsen70 has quit [Quit: Leaving]
Guest28 has joined #beagle
<Guest28>
Hi
<Guest28>
Hi
Guest28 has quit [Quit: Client closed]
<zmatt>
I guess he just really wanted to say greetings