00:15 karolherbst: RSpliet: why would you need magic?
01:21 HdkR: karolherbst: Because magic is cool duh
01:23 karolherbst: HdkR: sure, but if you hit a break you just pop from the stack until you reached the prebrk
01:24 karolherbst: I don't really see what more is there to do
01:29 HdkR: karolherbst: how are you popping the stack? :)
01:29 karolherbst: with magic of course
01:30 HdkR: lol
01:30 HdkR: Sadly that crap isnt publicly documented at all
01:31 karolherbst: HdkR: well, you have to type the stack entries anyway
01:31 HdkR: and the stack is gone on Volta+
01:31 karolherbst: so you pop until you hit the one you want
01:32 karolherbst: or maybe the hw just some internal pointers to the entries directly and removes everyting above
01:32 karolherbst: HdkR: yeah... volta is weird I've heard
01:33 HdkR: you should document this idea somewhere :>
01:33 karolherbst: there are not even "stack frames" aka no call
01:33 HdkR: Yep
01:33 karolherbst: so we can't haave a builtin library except we spill the regs
01:33 karolherbst: it's annoying
01:33 karolherbst: or we just don't use the regs used by that builtin
01:34 karolherbst: or uhm...
01:34 karolherbst: wait, we had to save anyway
01:34 HdkR: Yea, you have to spill your LR somewhere with calls
01:34 karolherbst: but on volta there is no ret, right?
01:35 karolherbst: so you need to save your jump address
01:35 HdkR: there is a ret, just that it takes a gpr as a address
01:35 HdkR: named branch effectively
01:35 HdkR: more like ARM
01:35 karolherbst: yes
01:37 HdkR: and of course divergence is handled entirely differently
10:13 karolherbst: pmoreau: I had to rework the shader cache patch because master diverged too much, asking if your rb still stands: https://github.com/karolherbst/mesa/commit/d250a27f3d0e3d213d4798354b9d19b1e4763518
10:43 RSpliet: karolherbst: that... doesn't work. You've got one stack pointer per warp, not per thread
10:46 RSpliet: Re undocumented: read https://ieeexplore.ieee.org/document/7011381 . Explains half of NVIDIAs methodology, points to a patent explaining the point of the "bra allwarp", and shows how all of this can be done using explicit predicate registers only instead
10:58 RSpliet: Oh, and explains compiler analysis methods and optimisations for things like thread-invariant branch optimisation
13:11 vedranm: karolherbst: a chemistry tool that uses OpenGL crashes when ran within apitrace, but works normally otherwise, where to start looking?
13:11 karolherbst: vedranm: apitrace errors?
13:12 karolherbst: vedranm: crashes while tracing or while replaying?
13:12 vedranm: while tracing
13:12 karolherbst: sounds like an apitrace bug then
13:12 vedranm: not sure if I saw any errors, just the stack call from the app
13:12 vedranm: *call stack
13:13 vedranm: I'll compile the latest apitrace and retry then
13:13 vedranm: thx
13:15 karolherbst: could be an application bug as well
13:15 karolherbst: lucking out when there is no wrapper
13:53 RSpliet: vedranm: an open source tool? (PyMol by any chance?) Not that I know of issues, but OSS tools are helpful for reproducability ;-)
13:55 vedranm: RSpliet: close
13:55 vedranm: Avogadro
13:56 vedranm: actually PyMOL worked fine
13:56 RSpliet: Crashes only when using apitrace could in theory be the result of an application bug, like performing out-of-bounds writes that normally write to garbage memory, but with apitrace loaded hits a more important buffer.
13:56 vedranm: and on a different PC with different Avogadro and apitrace and GPU, the issue disappears
13:56 vedranm: so I'll dig into what's causing it
13:56 vedranm: RSpliet: sure
13:56 RSpliet: So it might be worth analysing the program with Valgrind...
13:57 RSpliet: Well... it's part python, that's a pain to debug w/ valgrind :-/
14:00 vedranm: yup, good idea
14:00 vedranm: part Python is only for the extensions, I think
14:00 vedranm: like input generators for QM programs etc.
14:00 vedranm: and this crashes on initial display of the molecule
17:37 karolherbst: pmoreau: meh :/ there are some regressions since I last ran piglit. But I guess the changes will be trivial or images related
18:05 vedranm: RSpliet: ugh, can't reproduce it now :-/ when it happens again I'll hopefully observe what is different
21:17 pmoreau: pmoreau: Was the link you sent me the new version of the patch? I can’t figure out what changed. :-D