#
✕ nick:marcan
#asahi-gpu · 2022-10-03 08:10
marcanlike at least 1MB or something
#asahi-gpu · 2022-10-03 08:10
marcanno, I think it's a big fat buffer
#asahi-gpu · 2022-10-03 08:10
marcansorry, bad math, 33MB :)
#asahi-gpu · 2022-10-03 08:09
marcan24MB of SRAM for 4K, for each DCPEXT? don't think so
#asahi-gpu · 2022-10-03 08:09
marcanthat too
#asahi-gpu · 2022-10-03 08:09
marcanI don't think it has enough cache for whole frames, but yeah
#asahi-gpu · 2022-10-03 08:08
marcanmy guess is that they burst large chunks of framebuffer out of RAM at a time, so the RAM can be put to sleep most of the time
#asahi-gpu · 2022-10-03 08:08
marcana lot of DISPEXT seems to be cache
#asahi-gpu · 2022-10-03 08:08
marcanapple just put them into the low-end laptops too
#asahi-gpu · 2022-10-03 08:08
marcanand yes, that's the thing, M1/M2 are ipad socs
#asahi-gpu · 2022-10-03 08:07
marcanone DISPEXT is bigger than two PCPU cores
#asahi-gpu · 2022-10-03 08:07
marcanhttps://twitter.com/marcan42/status/1461390053389729794
#asahi-gpu · 2022-10-03 08:07
marcanI am not surprised they don't throw the silicon at it on the low end SoCs
#asahi-gpu · 2022-10-03 08:07
marcaner, DISPEXT I mean
#asahi-gpu · 2022-10-03 08:07
marcanthe DCP is actually *huge*
#asahi-gpu · 2022-10-03 08:06
marcanM2 is the same
#asahi-gpu · 2022-10-03 08:06
marcanbut who knows what Apple HW engineers are thinking...
#asahi-gpu · 2022-10-03 08:06
marcanand I guess most of the time only one of those is used
#asahi-gpu · 2022-10-03 08:06
marcanon the silicon you can see how each DCP has two mirrored halves
#asahi-gpu · 2022-10-03 08:06
marcanit's a waste of hardware though, I agree
#asahi-gpu · 2022-10-03 08:05
marcan(2 on M1 Pro, 4 on M1 Max, 8 on M1 Ultra)
#asahi-gpu · 2022-10-03 08:05
marcanthe pro machines have a separate DCPEXT, each with two pipes, for every supported monitor output
#asahi-gpu · 2022-10-03 08:05
marcanapple clearly designed these things with two pipes for tiled stuff only, even if you can abuse it for independent displays
#asahi-gpu · 2022-10-03 08:05
marcanwell, it's only sort-of fake
#asahi-gpu · 2022-10-03 08:03
marcanthe DCP powers that on when the output comes up
#asahi-gpu · 2022-10-03 08:03
marcanso there's *two* layers of firmware
#asahi-gpu · 2022-10-03 08:03
marcanthere's a cortex-m3 for realtime stuff
#asahi-gpu · 2022-10-03 08:03
marcanalso, it's inception
#asahi-gpu · 2022-10-03 08:02
marcansame firmware interface for any given macOS version for all socs
#asahi-gpu · 2022-10-03 08:02
marcanplus, it abstracts out platform differences for us
#asahi-gpu · 2022-10-03 08:02
marcan(as insane as it is, yes)
#asahi-gpu · 2022-10-03 08:02
marcanand I certainly am not looking forward to reverse engineering clocking, training, RT bandwidth calculations, and all that nonsense
#asahi-gpu · 2022-10-03 08:02
marcanthere's a *lot* of stuff in there
#asahi-gpu · 2022-10-03 08:02
marcanwell... the megabytes of firmware beg to differ
#asahi-gpu · 2022-10-03 08:01
marcan(which is possible)
#asahi-gpu · 2022-10-03 08:01
marcanbut yes, if we wanted to bypass the firmware, the only practical option is to just never power on the coprocessor and do everything ourselves directly against hardware
#asahi-gpu · 2022-10-03 08:01
marcanI'm not sure we want to go there
#asahi-gpu · 2022-10-03 08:01
marcanplus then that means we get to implement wonders like DP training
#asahi-gpu · 2022-10-03 08:01
marcanyes, but it's easy to crash the firmware doing that
#asahi-gpu · 2022-10-03 07:59
marcanyeah.
#asahi-gpu · 2022-10-03 07:59
marcanI can see this being practical for the special case of two identical monitors with identical EDIDs
#asahi-gpu · 2022-10-03 07:59
marcanimagine the combinatronics...
#asahi-gpu · 2022-10-03 07:59
marcanso then what, we read the EDIDs manually via the EPIC i2c interfaces, and conjure up some "merged" EDID with a tiled descriptor, then inject that?
#asahi-gpu · 2022-10-03 07:58
marcanfin
#asahi-gpu · 2022-10-03 07:58
marcansees no tiled descriptor, refuses to use #1
#asahi-gpu · 2022-10-03 07:58
marcanit probably just picks #0 and uses that
#asahi-gpu · 2022-10-03 07:58
marcanbut then that's the thing, you plug in two displays with two EDIDs, route them to separate pipes of the DCP, then what?
#asahi-gpu · 2022-10-03 07:58
marcanRight
#asahi-gpu · 2022-10-03 07:58
marcanI suspect this is all going to be, if possible at all, a very, very limited hack (that probably shouldn't even be enabled by default)
#asahi-gpu · 2022-10-03 07:57
marcanLike, how does the top-level decide what "mode" to expose downstream? Does it rely on tiled display descriptors in the EDID? Do we need to start conjuring up weird EDIDs that merge the real EDIDs somehow and use the virtual EDID feature (I *think* that is a thing)?

plus ›