#asahi-gpu · 2022-10-03 18:22
marcanif we need to drive them it'll be over i2c via the DCP i2c interface, most likely, if not a higher level abstraction, but that doesn't *necessarily* need to go through the i2c subsystem at this point (not the first DRM driver to not do that for bridges)#asahi-gpu · 2022-10-03 17:46
marcanor something like that#asahi-gpu · 2022-10-03 17:46
marcanI think the DP-HDMI bridges should be just child nodes of dcp or wherever they make sense, and just have a compatible (in case we need to start driving them) and say what the right connector type is on the other end#asahi-gpu · 2022-10-03 17:44
marcanwe're going to have to cook our own#asahi-gpu · 2022-10-03 17:44
marcanI don't expect existing DT bindings to be very useful for this, since they're mostly intended for real hardware, not a giant firmware abstraction#asahi-gpu · 2022-10-03 14:40
marcanI think that happens over the aux channel#asahi-gpu · 2022-10-03 14:40
marcanthe link rate would be determined dynamically by the DP negotiation stuff, so that makes sense#asahi-gpu · 2022-10-03 10:00
marcanbut you still have separate EPICs for the underlying dptx right?#asahi-gpu · 2022-10-03 10:00
marcanah, so that's like some shared one that is special for split mode#asahi-gpu · 2022-10-03 09:58
marcanhm, I expected more EPIC instances to show up#asahi-gpu · 2022-10-03 09:57
marcanand IOMFB needs to know about the tiling config to set up two outputs#asahi-gpu · 2022-10-03 09:56
marcanmy understanding from strings and such is that IOMFB is shared and the EPICs get split#asahi-gpu · 2022-10-03 09:56
marcanI didn't see any obvious place for that other than some unused bits in the EP messages#asahi-gpu · 2022-10-03 09:56
marcanthen there has to be some field somewhere for the index though#asahi-gpu · 2022-10-03 09:56
marcanif IOMFB really splits in half then that would be interesting#asahi-gpu · 2022-10-03 09:55
marcanif there is only one IOMFB then that needs a tiling config from EDID to make sense of both outputs#asahi-gpu · 2022-10-03 09:55
marcanwell, I mean, "just work" in what sense?#asahi-gpu · 2022-10-03 09:54
marcanI do expect split EPICs for both sides (so you could poke i2c separately etc) but I don't expect the main IOMFB to be split#asahi-gpu · 2022-10-03 09:53
marcansven: isn't the AV interface one of the EPICs? I doubt the main IOMFB is split#asahi-gpu · 2022-10-03 08:25
marcan16h idle time on my M1 Pro :p#asahi-gpu · 2022-10-03 08:23
marcanI think only some models do#asahi-gpu · 2022-10-03 08:23
marcanBut it's actually hard to find out exactly what intel CPUs have FIVRs these days#asahi-gpu · 2022-10-03 08:22
marcanIt's the same concept, Apple just does it on the other side of the motherboard, which isn't that much further away#asahi-gpu · 2022-10-03 08:22
marcanI don't think AMD does it?#asahi-gpu · 2022-10-03 08:22
marcanIntel has been doing on-package converters on and off lately#asahi-gpu · 2022-10-03 08:18
marcan:)#asahi-gpu · 2022-10-03 08:18
marcanM1 Ultra has 200 or so#asahi-gpu · 2022-10-03 08:18
marcanyou know, like those motherboards that advertise 8 or 10 power phases on x86#asahi-gpu · 2022-10-03 08:18
marcanpower phases as in independent switching converters#asahi-gpu · 2022-10-03 08:18
marcan(probably some big table somewhere initialized by iBoot)#asahi-gpu · 2022-10-03 08:18
marcanI think PMGR directly sequences SPMI commands to those when CPU clocks change, no firmware involved#asahi-gpu · 2022-10-03 08:17
marcanand each has like 20+ power phases#asahi-gpu · 2022-10-03 08:17
marcanthere's 10 of those chips under the M1 Ultra, it's pretty ridiculous#asahi-gpu · 2022-10-03 08:15
marcanmassively parallel CLVRs right under the SoC#asahi-gpu · 2022-10-03 08:15
marcanbut Apple have really fancy PMUs for that, especially on the pro machines#asahi-gpu · 2022-10-03 08:15
marcanupclocking needs to wait for the voltage to go up#asahi-gpu · 2022-10-03 08:15
marcandownclocking is 5µs#asahi-gpu · 2022-10-03 08:15
marcanit's the voltage that kills you#asahi-gpu · 2022-10-03 08:15
marcanso you do the math for how many CPU clock transitions per second we can do :)#asahi-gpu · 2022-10-03 08:15
marcanI have the CPU clock latencies somewhere, but IIRC it was like 50µs for slowest to fastest clock, including voltage ramp-up#asahi-gpu · 2022-10-03 08:14
marcanyou don't have to actually retrain for each power transition, only periodically for temperature compensation#asahi-gpu · 2022-10-03 08:14
marcanyeah, but once it's trained you keep the same training#asahi-gpu · 2022-10-03 08:13
marcanit's those 1000+ high-speed IOs that eat all the power#asahi-gpu · 2022-10-03 08:13
marcanit's mostly about stopping clocks and the IO interface, and yeah self refresh while that's down#asahi-gpu · 2022-10-03 08:11
marcanthis mobile stuff is super optimized for fast power transitions#asahi-gpu · 2022-10-03 08:11
marcanwe know the RAM can change perf modes pretty fast too#asahi-gpu · 2022-10-03 08:11
marcanthat's not surprising for LPDDR I think#asahi-gpu · 2022-10-03 08:10
marcanwhich is a lot better than other SoCs that have RAM powered on all the time during scanout#asahi-gpu · 2022-10-03 08:10
marcanso they can burst out big chunks out of RAM, at full speed, then power it down#asahi-gpu · 2022-10-03 08:10
marcanjust not the whole frame