03:05 jenatali: Again after auth... glehmann just suppress that warning?
03:06 jenatali: mareko: I'm on leave at the moment, I'll be back in early December. If this is urgent I can try to look before that to unblock you
04:36 mareko: jenatali: it's not
04:36 mareko: enjoy the leave
04:37 jenatali: Thanks, remind me in a few weeks and I'll be happy to look
07:21 mlankhorst: daniels: patchwork still 502/504?
10:47 mlankhorst: /9
11:14 biju: OOps in drm_bridge_connector_hdmi_cec_init+0x8/0x40 [drm_display_helper] (P)
11:14 biju: Any one noticed this issue with today's next?
11:15 biju: 22.653432] drm_bridge_connector_hdmi_cec_init+0x8/0x40 [drm_display_helper] (P)
11:15 biju: [ 22.660942] drm_bridge_connector_init+0x6a0/0x6cc [drm_display_helper]
11:15 biju: [ 22.667584] rzg2l_du_encoder_init+0xa0/0x230 [rzg2l_du_drm]
15:42 alyssa: dschuermann: looking into the smoothstep fails now
15:43 alyssa: it's interesting because that test gets 100% constant folded
15:43 alyssa: so I guess we have a slightly buggy constant folding somewhere
15:43 alyssa: cant repro on hk tho
15:43 alyssa: (can try again at intel once i get to the office)
15:45 alyssa: possibly affected if you opt algebraic before you constant fold
15:46 alyssa: Okay, reproduced on honeykrisp
15:46 karolherbst: this reminds me of this CL CTS bug where a constant folded kernel was used to determine how to GPU rounds 🙃
15:47 alyssa: dschuermann: https://rosenzweig.io/bad.txt https://rosenzweig.io/good.txt
15:47 alyssa: effect of opt algebraic after/before const folding
15:51 alyssa: float division by zero
15:52 alyssa: this is an edge case for smoothstep
15:52 alyssa: Possibly the test is invoking UB actually?
15:52 alyssa: `Results are undefined if edge0 ≥ edge1`
15:54 alyssa: spirv:
15:54 alyssa: %46 = OpFSub %float %float_1_1756356en38 %float_1_17549435en38
15:54 alyssa: %47 = OpFSub %float %float_1_1756356en38 %float_1_17549435en38
15:54 alyssa: %48 = OpExtInst %float %1 SmoothStep %46 %47 %float_0_5
15:54 alyssa: so it's interpolating with edge0=edge1 which is UB
15:54 alyssa: CTS bug. wahoo!
15:54 dschuermann: oh great :]
15:54 alyssa: you can take it from here, I gotta run
15:55 alyssa: hope this helps tho :p
15:55 dschuermann: haha, I kill you :P
15:55 dschuermann: thanks
15:55 alyssa: lazy ass workaround is to nir_opt_constant_fold in vk_spirv_to_nir somewhere
15:55 alyssa: but someone should fix CTS anyway
15:55 dschuermann: or put it into known fail list ...
15:56 alyssa: if youre not gonna fix CTS, the workaround in mesa is 1loc for all drivers
15:56 alyssa: since everybody except radv (AHEM!) uses vk_spirv_to_nir
15:57 dschuermann: lol, I didn't even know that
15:57 alyssa: radv should use it
16:00 dschuermann: sounds about right
16:00 alyssa: please
16:00 alyssa: ?
16:00 alyssa: :)
16:00 dschuermann: I'll look into it
16:00 alyssa: thx
16:00 dschuermann: so, after copy-prop a constant-fold?
16:01 alyssa: yea i think
16:11 frankbinns: bbrezillon: I read your cover letter for panfrost/panthor cached mappings and now I'm thinking https://elixir.bootlin.com/linux/v6.17.5/source/drivers/gpu/drm/imagination/pvr_gem.c#L372 is completely redundant...
16:44 bbrezillon: frankbinns: it is
16:45 bbrezillon: frankbinns: but maybe it's here to cover for the missing cache flush I was mentionning
16:52 frankbinns: bbrezillon: ah right, I'll go dig (or get someone else to :D)
19:39 alyssa: dschuermann: opened CTS ticket
20:29 dschuermann: alyssa: <3 thanks
22:43 alyssa: :]