<urjaman>
a65e29ccb261f3c127c49e73d0da92124234668f is the first bad commit
<urjaman>
"gallium: simplify throttle implementation" ........... i guess that can have some sort of an impact on pipelining?
<urjaman>
alyssa, bbrezillon ^^ that was the result of my bisect for the leaking fence fd
stikonas has quit [Remote host closed the connection]
<alyssa>
urjaman: Thank you for bisecting
<alyssa>
Oj, hm
<urjaman>
yeah i thought i'd do it since i had both an okay guess at a good commit (one of my previous pulls from reflog) and a not-too-slow reproducer (took less than a minute of youtube watching to kill xfwm4)
<urjaman>
also it was really annoying so i needed to find the good one anyways (and i found nothing obvious in recent history so i thought it'd be interesting)
<alyssa>
Here here
* alyssa
trying to debug
<alyssa>
(Since it's keeping me from getting much done in the way of anything else~)
<alyssa>
^ Apply that to master, let me know if that solves it for you
<urjaman>
building ...
<alyssa>
Ah, right, you're on a c201 :p
<alyssa>
Ahhh, nope, doesn't solve it :|
<alyssa>
Granted it slows it but
<urjaman>
testing ...
<urjaman>
ah hmm ...
<alyssa>
Provided it doesn't regress anything we still want that patch (so pipelining can work w/ corresponding kenrel support) but
<alyssa>
Gah this is over my head
<alyssa>
bbrezillon: ^^
<urjaman>
slows it how much?
<alyssa>
Not sure
<urjaman>
it seems to fix my test case -- or atleast, xfwm4 has 15 fds open and it hasnt changed in the few minutes i've been looking (and playing youtube :P)
<alyssa>
Give it time :-(
<urjaman>
yeah i'll run this mesa, will report if i notice something (and i'll keep an eye on the xfwm4 fds, hasnt changed yet...)
<alyssa>
Baby steps, I guess ... let's modify readmasks to be per-byte
<alyssa>
Ultimately a lot of compiler code will be rewritten to support CL stuff so might as well start here
<alyssa>
Oh gosh
<alyssa>
The
<alyssa>
magnitude of this task
<alyssa>
is only starting to dwell on me
Elpaulo has quit [Ping timeout: 268 seconds]
BenG83 has joined #panfrost
<robmur01>
alyssa: patch seems to work - kmscube now running for more than ~10s and fences apparently not hanging around for long enough to show up in /proc/*/fd at all
<robmur01>
ah, and it's pushed already, cool!
<alyssa>
robmur01: Doesn't totally solve it
<alyssa>
but definitely delays the bad
<alyssa>
and there's no good theoretical reason the patch should solve
<tomeu>
narmstrong: wasn't expecting to see that many flaky tests on t820
<narmstrong>
tomeu: yeah, pretty bad, and I have 0 time to help you guys on mesa to fix the t820
<tomeu>
narmstrong: well, we only need to check it doesn't regresses, not necessarily fix it now
<narmstrong>
tomeu: I'm doing plenty of runs to skip the flaky tests
<tomeu>
narmstrong: if you want, I can take up the MR and figure out what needs to be added to skips and fails
<narmstrong>
tomeu: I'll let you review my flaky tests detection tentative when finished !
<tomeu>
cool!
pH5 has quit [Quit: bye]
yann|work has quit [Ping timeout: 240 seconds]
raster has quit [Quit: Gettin' stinky!]
adjtm has quit [Ping timeout: 240 seconds]
adjtm has joined #panfrost
TheKit has quit [Remote host closed the connection]
davidlt has quit [Ping timeout: 240 seconds]
raster has joined #panfrost
fysa has joined #panfrost
<alyssa>
Ahhh the bugs are piled higher and deeper than I ever imagined !
<anarsoul>
:(
<alyssa>
Trying to fix more 16-bit issues ...
<alyssa>
The dest override field in the disasm has an entropy poblem
<alyssa>
We already know whether it's upper or lower half based on the mask
<alyssa>
Ohhh, I think I get the difference
<alyssa>
The 8-bit mask field is a doubled mask, masking the sources/computation
<alyssa>
The override does the implicit conversion and determines which half to feed into
<alyssa>
It's still sorta wasteful, but it is self-consistent
<alyssa>
Also my handling of vec8 swizzles in the disasm is backwards, look like
<alyssa>
Okay, yeah, fine
<alyssa>
daniels: See what you've done :oooo
<alyssa>
:p
<alyssa>
Okay, in that case, I'm wondering if the best way to proceed is to mask off the source's swizzles explicitly
<alyssa>
--no, wait, that doesn't make sense.
<alyssa>
What?!
<alyssa>
No, it's to parse the mask _as is_
<alyssa>
Rather than stepping down size to parse it
<alyssa>
And then applying the upper? on the override as a shift
<alyssa>
Change the alphabet.
<alyssa>
The one case that's not accounted for is ... what if you wanted to replicate to both upper and lower?
<alyssa>
I don't know if the hw can do it, but I suspect if so, that'd be dest override = 3
robmur01 has quit [Ping timeout: 246 seconds]
raster has quit [Quit: Gettin' stinky!]
* alyssa
can't seem to make the blob do that, maybe it can't
<alyssa>
Oh, well, masks are behaving which is the important part
<alyssa>
daniels: See what you've done! :D
<daniels>
alyssa: oosh
cwabbott has quit [Remote host closed the connection]
tgall_foo has quit [Read error: Connection reset by peer]
cwabbott has joined #panfrost
warpme_ has quit [Quit: Connection closed for inactivity]