warpme_ has quit [Quit: Connection closed for inactivity]
youcai has joined #panfrost
<youcai>
hello, is there currently any way to do hardware accelerated video with panfrost?
<youcai>
like vaapi/vdpau/nvdec or what ..
<hanetzer>
youcai: I don't think mali has that itself, unless you count an opencl program running on it.
<hanetzer>
youcai: there is however what rockchip calls the mpp.
<youcai>
hanetzer: okay, I'm actually testing on exynos5422
<hanetzer>
ah. no specific clues about that, but as far as I know panfrost itself doesn't have a vaapi thingie.
<hanetzer>
like, most desktop gpus have internal extra silicon for video enc/dec
<youcai>
got that, meson give errors when I try to build with only panfrost and --with-vdpau/vaapi
<youcai>
hey but ffmpeg supports opencl.. does that mean I may still take the video decoding load off cpu if I use ffmpeg playback + opencl + panfrost?
fysa has joined #panfrost
_whitelogger has joined #panfrost
Elpaulo has quit [Read error: Connection reset by peer]
Elpaulo has joined #panfrost
anarsoul|c has quit [Quit: Connection closed for inactivity]
<hanetzer>
not sure if the opencl support for panfrost is there yet. it is capable of it in the hardware, however no idea if the driver has support for it yet.
<bbrezillon>
tomeu: was panfrost_gem_shrinker_scan() before that oops happened?
<bbrezillon>
*was panfrost_gem_shrinker_scan() called
kaspter has joined #panfrost
stikonas has joined #panfrost
karolherbst has quit [Ping timeout: 246 seconds]
icecream95 has quit [Ping timeout: 265 seconds]
stikonas has quit [Remote host closed the connection]
stikonas has joined #panfrost
robmur01_ has joined #panfrost
kaspter has quit [Quit: kaspter]
<urjaman>
panfrost OpenCL is very muchly not there yet ... also not sure if there are FOSS openCL video decoders around much (and also the rockchips have the VPU for that)
<urjaman>
i did at one point read from the wiki that some company(TM) (forgot) made an OpenCL VP9 decoder, one of those could be cool on the RK3288 (that has VPU only for H264 and VP8, basically) but alas proprietary
<robmur01_>
I'd be suspicious of a similar list mishap in panfrost_mmu_as_get()
AreaScout_ has joined #panfrost
davidlt has joined #panfrost
<tomeu>
robmur01_: any ideas on why this seems to be easier to reproduce on the t*60 than on the t*20?
raster has joined #panfrost
guillaume_g has quit [Ping timeout: 258 seconds]
<robmur01_>
tomeu: hmm, races around faults/IRQs seem quite sensitive to relative GPU/CPU speeds
<robmur01_>
but for a CPU-side race it might just be down to systems with the bigger GPUs also running bigger more aggressive CPUs (A17/A72)
guillaume_g has joined #panfrost
guillaume_g has quit [Quit: Konversation terminated!]
guillaume_g has joined #panfrost
robmur01_ has quit [Ping timeout: 260 seconds]
<tomeu>
alyssa: seeing lots of:
<tomeu>
Test case 'dEQP-GLES3.functional.shaders.texture_functions.texturelodoffset.sampler2d_fixed_vertex'..
<tomeu>
ERROR: Failed to choose spill node
<tomeu>
which is followed by a couple of faults
<youcai>
Green, hanetzer: thanks, but I think that prevents me from using this mali-midgard any further however. Btw, any desktop environ runs on panfrost now?
<hanetzer>
apparently x11 support is non-existant, so it would have to be wayland-flavoured
<urjaman>
hanetzer: xfce works fine right now in x11
<urjaman>
or like my build is a few days old but anyways
<urjaman>
(not perfect, but fine)
<youcai>
thanks, i memorized some of extra lightweight DE for wayland
<urjaman>
sway seems ok too (that's my non-x11 fallback)
<youcai>
great, should be enough for a tv box
guillaume_g has quit [Quit: Konversation terminated!]
BenG83 has joined #panfrost
anarsoul|c has joined #panfrost
raster has quit [Quit: Gettin' stinky!]
raster has joined #panfrost
yann has quit [Ping timeout: 265 seconds]
davidlt has quit [Ping timeout: 260 seconds]
youcai has quit [Quit: Leaving.]
<alyssa>
TheCycoONE1: Lot of bugfixes have landed in the past few weeks :) so shader magic
<alyssa>
krh: What distro?
<alyssa>
youcai: As others have mentioned, video accel is separate from 3D (and both separate from display) on arm chips
<alyssa>
For Rockchip RK3399, there is a Rockchip VPU which Collabora has been working on upstream support for.
<alyssa>
ezequielg: ^^ has been poking at that, and I think that can cooperate with panfrost
<alyssa>
I'm not sure what the story is on exynos
<alyssa>
icecream95: Odd -- the uniform/work heuristic is to solve a register spilling issue
<alyssa>
It should actually make glmark2 marginally slower
<alyssa>
Though if you're running under X11, it should make the entire machine faster since it's a big boost for glamor, so that in turn lets glmark go faster.
<alyssa>
So I guess it's a win for glmark2-es2 but a loss for glmark2-es2-wayland
<alyssa>
tomeu: Yeah, texturing needs some more work, I can look into it
<alyssa>
Probably have overzealous swizzles, should be an easy fix; I've had a similar bug with load/store registers for OpenCL stuff
<alyssa>
hanetzer: X11 has worked just fine for a while now :)
<hanetzer>
ah :P
<alyssa>
As urjaman said -- fine, not perfect. I am running X+panfrost on my personal Kevin which works nice
<alyssa>
(And panfrost accelerating X / using GPU for shadows and things)
<alyssa>
So turtles all the way down
<alyssa>
Oof. The blob uses some clever tricks to save registers for texture ops (since otherwise a conformant ES3.2 implementation wouldn't be possible)
<alyssa>
In particular, you need to note that texture sizes are bounded at 65k, so it's legal to use 16-bit integer texture coordinates
<alyssa>
So the blob uses half-precision for texelFetch where necessary, which saves you space
<alyssa>
Likewise the offset is legal to be 16-bit integers, to again save space
<alyssa>
This matters since texelFetch requires a full vec4 coordinate
<alyssa>
For a 3D texture you'd have a vec3 offset
<alyssa>
With an LOD attached
<alyssa>
And if it was accessed indirectly that's another integer
<alyssa>
So that would be 9 components -- if they were all full precision, that would fail to compile, since you only get 256-bits of texture registers on Midgard
<alyssa>
So indirect access with texelFetchOffset would fail
<alyssa>
But when you realize that a bunch of those can be half-precision, it fits just fine!
<alyssa>
Though I don't believe indirect sampler access is allowed on ES3.0, so I don't need to worry about that *quite* yet
yann has joined #panfrost
<alyssa>
Meanwhile my notes on textureOffset were wrong so that's just lovely.
davidlt has joined #panfrost
<alyssa>
Er, nvm on the half precision thing -- texelFetch's offsets are allowed to be signed so they can always be inlined, it's regular textureLodOffset you gotta wtach out for
<alyssa>
I'm unable to get the blob to put the textureOffset anywhere but xyz
<alyssa>
It's possible it *has* to be aligned, of course.
<BenG83>
outside of gtk applications everything seems to work fine
<alyssa>
Expected? no
<alyssa>
But it's hard to tell much from that trace, unfortunately.
<alyssa>
What mesa version are you using?
<alyssa>
BenG83: 3D texture tests flake
<alyssa>
Er
<alyssa>
tomeu: ^^
<BenG83>
mesaOpenGL ES profile version string: OpenGL ES 2.0 Mesa 20.0.0-devel (git-2c8742ed85)
<alyssa>
tomeu: dEQP-GLES3.functional.shaders.texture_functions.texture.sampler3d_float_fragment passes when run as part of dEQP-GLES3.functional.shaders.texture_functions.* but fails alone. I'm inclined to blame a resource management issue (setup_slices might be broken for 3D/2D-array -- they worked over the summer but ain't nobody got time to bisect that ;P)
<BenG83>
as I said, only seems to happen with gtk based applications, so I'm not sure if there is something wrong there or how to provide more info
<alyssa>
BenG83: Hm, okay, that's recent..
<alyssa>
Which gtk applications?
<BenG83>
I tried Gimp and some others
<BenG83>
I get a black window, the contents are drawn about 10 minutes later
<alyssa>
BenG83: I wonder if it would be possible to grab an apitrace?
<alyssa>
$ apitrace trace -a egl gimp
<BenG83>
I'll try
<alyssa>
That should spit out gimp.trace, which you can upload somewhere and send to me ... hopefully I'll be able to reproduce from that trace.
<BenG83>
on X the same applications work, but just draw their window content very slow
<BenG83>
and there are less of those crashes in the log
nerdboy has quit [Ping timeout: 265 seconds]
<BenG83>
mmh I ran apitrace trace -o gimp.trace -a egl gimp now two times, but it seems to produce no output file