00:46 RSpliet: imirkin: in RA, is there a way of fairly accurately estimating register usage between liveness analysis and the actual assignment?
00:47 RSpliet: I mean, in theory there is, although allocation of doubles could cause an underestimate, but in practice?
00:58 RSpliet: pmoreau: GL_ARB_shading_language_420pack ... they must've been very high when proposing that :-D
02:03 imirkin: RSpliet: doubles are double-wide registers, RA is aware of that
02:03 imirkin: RSpliet: at every point in time you know how many values are live, so ... "yes", i guess?
02:25 RSpliet: imirkin: RA might be aware of that, but I guess liveness is still only an approximation. If your widest liveness is 4 registers, but you've come to a point where reg 0 and 2 are assigned and you need a double, you're screwed and need to use {3,4} (or {4,5} if there's alignment constraints)
02:29 RSpliet: I've gotten kind of interested in the problem of banked registers. some paper mentioned Kepler appears to have four banks (or I guess 4*32 considering warp size), gets best performance if you make sure that in a single instruction your regs are spread over the banks as well as possible
02:30 imirkin: yeah, with groups that becomes an issue - not just doubles. groups come up a lot. e.g. texturing.
02:33 RSpliet: For heuristics on spreading registers over banks it seems useful to have such a conservative estimate though. Let's you determine whether picking a register other than the first available risks increasing the total reg usage
02:33 RSpliet: *lets
02:46 imirkin: tbh i haven't given it too much thought
02:47 imirkin: we keep the current set of regs in a BitSet
02:47 imirkin: and they're allocated out of there
10:46 RSpliet: yep, I found that logic yesterday... the "easy read" part of _ra :-)
11:14 karolherbst: mupuf_: ping on 80901, the issue was the stupid vbios table, right?
11:14 karolherbst: or was it somethign else?
11:31 karolherbst: ohh wait, different issue
11:32 mupuf_: karolherbst: no worries, I will answer the guy
11:32 mupuf_: and no, it looks like osmething went wrong
11:33 mupuf_: it needs bisecting, for sure
11:36 karolherbst: well, reverting one commit helps
11:36 karolherbst: https://lists.freedesktop.org/archives/nouveau/2014-March/016589.html
11:36 karolherbst: I will add his vbios
11:38 karolherbst: uhhhh
11:38 karolherbst: this vbios is odd
11:39 karolherbst: nvbios | wc: 1092973 3290985 30706395
11:39 karolherbst: :O
11:39 karolherbst: HWEA is like endless
11:40 karolherbst: no boost table, no cstep table
11:40 karolherbst: well thet ables are tehre, just empty
11:41 karolherbst: oh well, it is an old fermi card, so no surprise really
11:42 karolherbst: "id = 0x22, data = 0x501e -- fan_min = 30, fan_max = 80"
12:01 mupuf_: karolherbst: I need to see what could lead us not to set a PWM fan for him
12:02 mupuf_: no changes were made in nouveau for the fan between 4.6 and 4.7
12:03 karolherbst: yeah no clue
13:02 karolherbst: mupuf_: is there any disadvantage in using the values you calculate? or are there values with big relative errors?
14:08 karolherbst: meh, we should make a list with games having crappy vsync implementations and write to the devs to fix it
14:09 Yoshimo: i seriously doubt that they will fix it
14:11 RSpliet: why fix when they can just claim we should be providing them with 60fps, rocking the cards at full speed ;-)
14:12 karolherbst: RSpliet: I was talking about the broken ones
14:12 karolherbst: I have games tearing at 60 fps
14:12 RSpliet: on the official driver too?
14:12 karolherbst: no quite sure, but I would have to use bumblebee, things are different there compared to prime
14:13 karolherbst: but I think I had some, yes
14:13 Yoshimo: why would it be different than the situation on windows karol? Games have problems and bad code, nvidia and amd release new drivers, everyones happy
14:14 RSpliet: sometimes game dev fixes things, sometimes GPU vendor fixes/works around things
14:15 RSpliet: I recall seeing VDPAU tear on NVE7 a few years ago using the official driver. Shit happens.
14:16 RSpliet: where was I getting at? oh yes... if this tearing problem exists on NVIDIA, I presume that the game dev has decided that this is either acceptable or a result of PRIME/bumblebee bugs they haven't seen before
14:16 RSpliet: you can nudge them, but don't get your hopes up too high. Nouveau just isn't important enough in its current state