ChanServ changed the topic of #panfrost to: Panfrost - FLOSS Mali Midgard & Bifrost - Logs https://oftc.irclog.whitequark.org/panfrost - <macc24> i have been here before it was popular
vstehle1 has quit [Ping timeout: 480 seconds]
vstehle1 has joined #panfrost
rasterman has joined #panfrost
icecream95 has joined #panfrost
<icecream95>
robmur01: Those patches don't seem to have helped..
<icecream95>
I added another file to https://gitlab.freedesktop.org/-/snippets/2680 which has the kernel logs from a second Oops, but two of them happened at exactly the same time so they are interleaved
icecream95 has quit [Read error: Connection reset by peer]
icecream95 has joined #panfrost
icecream95 is now known as Guest4093
Guest4093 has quit []
icecream95 has joined #panfrost
alpernebbi has joined #panfrost
strongtz[m] has joined #panfrost
psydroid[m] has joined #panfrost
stebler[m] has joined #panfrost
icecream95 has quit [Quit: leaving]
icecream95 has joined #panfrost
camus1 has quit [Remote host closed the connection]
camus has joined #panfrost
toggleton[m] has joined #panfrost
<robmur01>
maybe there's still a race somewhere
<robmur01>
although I do note that your build ID is still "5.14.0-rc4" where it would typically be "5.14.0-rc4+" if extra patches were applied
<icecream95>
robmur01: I used a source tarball, so it's not a git repo which I guess the build scripts would need to know whether to append that "+"
<robmur01>
ah, that's cool - I know there are a number of ways that can be misleading (like somehow ending up with some old commit hash in CONFIG_LOCALVERSION as a colleague did once - that can *really* confuse debugging...)
<icecream95>
It it's a race, it's not a particularly hard one to hit, because I got two Oopses at exactly the same time again..
<icecream95>
..and again
icecream95 has quit [Quit: Reconnecting]
icecream95 has joined #panfrost
<icecream95>
I can reproduce it with for i in {1..10}; do timeout 5 glmark2-es2 -b refract; done, which doesn't even cause any GPU faults
<robmur01>
well, at least it doesn't look like a race between teardown and the fault handler any more - if anything it smells like some kind of double-free
<icecream95>
[Other ways I considered included BITSET_RANGE(bits_frac, bits_int + bits_frac - 1)]
<alyssa>
icecream95: sure, will change that
<alyssa>
also note to self: comment at the top of pan_clear.c doesn't match the code I ended up writing, will amned
<icecream95>
Having tests means that it's easy to swap out things like this without worrying about breaking anything
<alyssa>
Indeed
<alyssa>
aside - whether "color" or "colour" is used is a coin flip, i am sorry
<icecream95>
I vote for correct spelling wherever possible; we can change all the other drivers if people complain about incosistency
<alyssa>
I also vote for correct spelling, except the API headers are all mispelled..
<icecream95>
I'm sure people won't mind if we change those as well:
<icecream95>
#ifdef GL_VERSION_4_7
<icecream95>
glClear(GL_COLOUR_BUFFER_BIT);
<icecream95>
#else
<icecream95>
glClear(GL_COLOR_BUFFER_BIT);
<icecream95>
#endif
<alyssa>
Hihihihi
<alyssa>
eh.. back to work..
<alyssa>
icecream95: Made the change to the factor calculation :+1:
* icecream95
looks at the :+1: and wonders if there was an off-by-one somewhere
<alyssa>
it should be a thumbs up emoji
<alyssa>
but i'm typing from a plain TTY on linux on m1
<alyssa>
since the hack who announced they'd deliver linux graphics drivers for the m1 still hasn't 7 months later... oh wait.
<icecream95>
Anyway I used to keep complaining that I couldn't see them whenever you used emoji
bluebugs has quit [Quit: Leaving]
indy has quit []
indy has joined #panfrost
<icecream95>
Is this idea too much of a hack?
<icecream95>
For each UBO quadword, store a mask of read words, but if a read crosses a quadword boundary, put the mask in both the upper bits of the quadword and the next quadword
<icecream95>
(So a four component read at offset three would give the masks [0b1111000, 0b111])
<icecream95>
bi_pick_ubo will use the entire mask for the check against PAN_MAX_PUSH, but only the lower four bits for the calls to pan_add_pushed_ubo
<icecream95>
That means I only need to test for the "pushed" bit once in bi_opt_push_ubo, because cross-quadword reads are only pushed if both words are pushed
chewitt has joined #panfrost
* icecream95
assumes the answer is "yes" but implements it anyway