13:48 <ad__> hi all

14:12 <ad__> is there a particular version from where you can suggest to switch to etnaviv from imx-gpu-viv ?

14:38 <cphealy> ad: particular version of what? Kernel? Mesa? Something else?

14:40 <cphealy> If your talking about Mesa, I would suggest using the latest released version (Mesa 20.2) Likewise, for the kernel, go with the latest version you can run on your HW.

14:41 <ad__> cphealy, ok, was referring to kernel, i.e. for a 5.4 (LTS) would be possible to use etnaviv to replace mxc-gpu-viv ?

14:42 <ad__> also, i think both are available in the imx tree, so one should be able to test both, right ?

14:43 <cphealy> ad: I'm using 5.4 with the i.MX6q (GC2000), i.MX6qp (GC3000), and i.MX8M (GC7000lite). That said, I intend on upgrading soon for reasons unrelated to etnaviv.

14:43 <ad__> cphealy, thanks a lot

14:44 <cphealy> What kernel you use likely depends on what your use cases are and what SoC you are working with though.

14:46 <marex> ad__: 5.4 at least, mesa 20.1.y (latest) at least

14:46 <marex> ad__: which SoC are you on ?

14:46 <marex> ad__: and it is always good to switch from the crappy blob

14:46 <ad__> ok i am on imx6dual, was looking about mxc-gpu-viv improvements from 4.9 linux-fslc to 5.4, then i tested a mainline kernel on sabrelite and seen it is using etnaviv, so was trying to understand if etnaviv performance are better/worst from vivante driver

14:46 <marex> ad__: just use mainline on mx6d

14:47 <ad__> hi marex :)

14:52 <ad__> well, woudl be ok a stable 5.4.y (LTS) rebasing imx stuff on, so to have last improvements also for mxc-gpu-viv (so te be able to test both, if it has sense) ?

15:10 <cphealy> ad: The i.MX6dual uses the GC2000. Earlier this year, I benchmarked the GPU using glmark2-es2 on 5.4 with Mesa 19.3.2 and compared against the NXP Vivante blob driver. glmark2-es2 scores showed that etnaviv was a few percent lower than the Vivante blob driver overall, but when looking at the individual tests it appears that most tests are faster with etnaviv except for the texturing tests where etnaviv is considerably slower

15:10 <cphealy> than the Vivante blob. If the texturing issue is resolved with etnaviv, then the overall score for etnaviv would be faster than the Vivante blob driver.

15:13 <ad__> cphealy, great, thanks a lot for these precious infos

15:15 <ad__> and marex, thanks a lot :)

15:19 <cphealy> Question for all on here: This knowledge about etnaviv vs the Vendor blob performance feels like something that should be documented and periodically updated somewhere as it's a common question. Where would be a good place to put this knowledge?

15:24 <cphealy> Perhaps a place for putting a transition guide of something akin to that too as there are things that are different from a coding standpoint when switching between the drivers. For example, video import is different between Vivante and etnaviv in that Vivante uses some vendor proprietary GL extensions while etnaviv uses existing Khronos extensions.

16:51 <marex> ad__: does imx blob even work on 5.4.y ?

16:52 <ad__> marex ?? havent yet tested, but, is it broken in 5.4 ? compraed to 4.9, there are abouot 80 patches ov varipous improvements

16:55 <marex> ad__: I didnt use the blob for years, I only ever use it on some archaic heavily patched vendorkernel where it barely works if I need to see what it does ; but luckily I dumped all the command streams by now, so not even that is needed anymore

16:57 <ad__> well, i believe some out there wants to use imx blob since is supported from nxp

16:58 <ad__> for any issue, nxp says "out _2.0.0 bla bla" is only supported

17:05 <marex> ad__: supported how ? like on the imx community ?

17:05 <marex> I never found any really useful answers there

17:05 <marex> sorry

17:06 <marex> cphealy: probably the visucore wiki, or elinux wiki

17:06 <marex> cphealy: also, some sort of word diff of dEQP-GLES*.perf* tests would help

17:06 <marex> then it would be clear exactly what is slower and what is faster

17:35 <cphealy__> marex: thanks for the recommendation. I'll look at that.

18:20 <sravn> ad__: see also the presentation made by austriancoder at xdc2020: https://xdc2020.x.org/event/9/contributions/618/. It has a nice summary of the performance deltas

18:28 <ad__> sravn, thanks !