#asahi-re · 2026-09-14 22:47
bobjrso, current guess: pipe does 1 pixel per clock, macos gets 2. everything else checked and dead#asahi-re · 2026-09-14 22:47
bobjrj`ey: english is not my first language, so it is either broken or llm-polished. broken it is then#asahi-re · 2026-09-14 22:43
bobjrfl0_id_: said it above - english is llm-polished, im not a native speaker. the box, the traces and the numbers are mine. happy to just paste raw dmesg instead if that reads better#asahi-re · 2026-09-14 22:33
bobjrbtw the fw debug gate is just a flag in its data segment, and the data segment is writable from the cpu (the code is not). flipping it makes the fw print the full underrun breakdown with per stage latencies. happy to share details if that is useful to anyone#asahi-re · 2026-09-14 22:32
bobjrthe underrun lands at line 21 / pixel 2664, ~83k pixels into the frame. if the pipe were only 5% short the prefetch buffer would have to be about one line; at half rate it works out to ~11 lines, which is a sane line buffer. so 1 px/clk vs 2 keeps fitting#asahi-re · 2026-09-14 22:32
bobjrchaos_princess: dsc looks out on this pair fwiw. macos at working 4k120 has SupportsDSC=0 on every colour element and DSCRequiredColorElementIDs empty, same in linux. no dsc calls in the macos fw log either. the panel advertises 4k120 uncompressed (displayid 1.2, 1066.5 mhz) and hbr3x4 just barely fits it#asahi-re · 2026-09-14 22:31
bobjralso got the fw to print its own underrun report: "display underrun [frame:1261] [line:21] 0x8, latencies: GP [0, 0] PIODMA [0] SMMU [0] REPLAY [0] PDC [0], DecompFail [0]". so nothing is starving it, it just cannot keep up with the pixel rate. fits the 1 vs 2 pixels per clock idea#asahi-re · 2026-09-14 22:31
bobjrchadmed: yeah, traced it under m1n1 with macos as guest, 4k60 and 4k120 both. iomfb calls match what we send, m3 timing params match, split is 0 on both sides#asahi-re · 2026-09-14 22:27
bobjrlol nope. human, just three nights deep in this. the philips and the numbers are real#asahi-re · 2026-09-14 22:21
bobjrchaos_princess: fits with what I measured - it is not bytes, it is pixels. NV12 at 4K120 reads 1.49 GB/s and still underruns, while 4K60 at 32bpp reads 1.99 GB/s and is fine. Everything up to ~530 Mpix/s works, 1066 does not - as if the pipe does 1 pixel/clock where macOS gets 2. Is that tied to DSC here? There is no dsc code in the dcp driver at all.#asahi-re · 2026-09-14 22:16
bobjrchaos_princess: by default Linux picks colorID 97 = RGB 8bpc, ie 25.6 Gbps of the 25.92 HBR3x4 budget - suspiciously tight. But I forced colorID 130 (YCbCr 4:2:2 10bit, 21.3 Gbps, what macOS picks on this timing) and the band is identical, just green instead of black. So it does not look link-limited.#asahi-re · 2026-09-14 22:14
bobjrchaos_princess: Philips PHL 279M1RV, 4K 120Hz panel, connected over USB-C (DP alt mode, HBR3 x4). 4K60 and 1440p120 are fine; the monitor itself does 4K120 - macOS on the same machine and the same cable drives it cleanly.#asahi-re · 2026-09-14 22:13
bobjrmacOS on the same machine (Ventura under m1n1) drives 4K120 cleanly. Any idea what the pipe needs at ~1 Gpix/s that we are not programming? Happy to dump registers or traces if that helps.#asahi-re · 2026-09-14 22:13
bobjrRuled out by experiment: memory bandwidth (upscaled 1440p source behaves identically), PMGR perf states (same zeros under macOS), DP link rate (same band with YCbCr 4:2:2 10-bit, 21.3 of 25.92 Gbps), colour mode, and every appledrm module parameter.#asahi-re · 2026-09-14 22:12
bobjrMac mini M2 (j473), external display over USB-C on dcpext, 13.5 fw. At 4K120 (timing 94, 1066500 kHz) only the first ~10 lines of each frame are correct, the rest is zeros (black in RGB, green in YCbCr). DCP publishes "bool underrun = 1" on that modeset, and does not at 4K60 or 1440p120.