#
#asahi-gpu · 2023-03-29 04:27
i509vcbIf I had to guess the PPP (Primitive Processing Pipeline) does primitive assembly?
#asahi-gpu · 2023-03-29 04:13
i509vcb> The three pipelines the kernel sees they TA, 3D, and CP, and they map to VDM, PDM, CDM
#asahi-gpu · 2023-03-29 04:13
i509vcbSearching through the logs I did find this:
#asahi-gpu · 2023-03-29 04:13
i509vcbWhat specifically makes a pixel shader a pixel shader? A lot of sources I found seem to conflate fragment and pixel shaders
#asahi-gpu · 2023-03-29 04:08
linaFrom what alyssa told me, AGX/PowerVR does not support *fragment* shaders, it only supports pixel shaders, that's why it's called PDM ^^
#asahi-gpu · 2023-03-29 03:42
i509vcbPixel Data Master (PDM) is the hardware which dispatches fragment shaders? (I'd think fragment since that sounds most similar to pixel)
#asahi-gpu · 2023-03-28 19:30
jannauthe gfx cluster on pro/max/ultra have to communicate somehow
#asahi-gpu · 2023-03-28 19:25
alyssalol
#asahi-gpu · 2023-03-28 19:23
cr1901gfx cards do Point To Point Protocol now?
#asahi-gpu · 2023-03-28 19:04
alyssaISP image synthesis processor
#asahi-gpu · 2023-03-28 19:04
i509vcbI have the logs don't worry there
#asahi-gpu · 2023-03-28 19:04
alyssaPPP i forgot but grep the logs
#asahi-gpu · 2023-03-28 19:04
alyssaCDM/PDM/VDM are compute/pixel/vertex data master, same as powervr
#asahi-gpu · 2023-03-28 19:03
alyssai509vcb: sure, i'd love a glossary in https://docs.mesa3d.org/drivers/asahi.html
#asahi-gpu · 2023-03-28 18:59
i509vcband possibly a super high level description of what the hardware associated with the abbreviation does
#asahi-gpu · 2023-03-28 18:58
i509vcbI'm wondering if I should write a patch to note all these abbreviations
#asahi-gpu · 2023-03-28 15:28
alyssai509vcb: Compute Data Master, i.e the hardware that dispatches compute kernels
#asahi-gpu · 2023-03-28 12:27
linaI did not! ^^
#asahi-gpu · 2023-03-28 11:44
jannaulina: did you see https://www.youtube.com/watch?v=dgd88zE3O38 ?
#asahi-gpu · 2023-03-28 10:52
_jannau_there are multiple fb sized buffers but all only attached to the gpu. I would have expected that at least the currently displayed fb is attached to dcp
#asahi-gpu · 2023-03-28 10:43
linaI did see framebuffer-sized things...
#asahi-gpu · 2023-03-28 10:32
_jannau_the actual framebuffers (by size) aren't attached to the display-subsystem though
#asahi-gpu · 2023-03-28 10:23
linajannau: we map to DCP anything that "could be" a framebuffer (to make kmsro work), so it can have stuff that has nothing to do with framebuffers... though we might be too eager right now. I'll take a look tomorrow ^^
#asahi-gpu · 2023-03-28 09:13
jannaubest bet for that is dcpep_cb_allocate_buffer/dcpep_cb_release_mem_desc but strange that I never saw it before 6.2.
#asahi-gpu · 2023-03-28 09:11
jannauit has 3 objects of 60817408 bytes, which is a tad excessive for framebuffers on j314. most common size is 2523136 bytes
#asahi-gpu · 2023-03-28 09:06
jannaulina: looks like dcp might something according to /sys/kernel/debug/dma_buf/bufinfo (currently "Total 73 objects, 497156096 bytes", was fewer objects yesterday) but I have no idea what. Judging by the size it's not framebuffers
#asahi-gpu · 2023-03-28 04:46
i509vcbIt seems to be compute related from searching around the gallium driver?
#asahi-gpu · 2023-03-28 04:44
i509vcbWhat does CDM mean in agx-land?
#asahi-gpu · 2023-03-28 04:08
alyssalina: irregularly scheduled reminder that i appreciate you ^^
#asahi-gpu · 2023-03-27 16:58
i509vcbSmithay plans to use IN_FENCE_FD eventually so that we don't need to keep using glFlush and blocking on Vulkan, but we aren't done with that yet
#asahi-gpu · 2023-03-27 14:23
linaAlso I tried to reproduce the issue and I ended up crashing the GPU firmware instead... that's not good...
#asahi-gpu · 2023-03-27 14:23
linaAh yeah, I doubt kwin uses that
#asahi-gpu · 2023-03-27 14:23
jannauthe kms plane has an "IN_FENCE_FD" property for explictic fences from userspace: https://www.kernel.org/doc/html/v4.15/gpu/drm-kms.html#explicit-fencing-properties
#asahi-gpu · 2023-03-27 14:03
linaI haven't looked at the KMS side API but if it supports both an explicit fence as well as implicit buffer fences, I would expect only implicit ones since that's how the usual WSI stuff works so far as far as I know
#asahi-gpu · 2023-03-27 14:02
linaI meant like no fence attached at all
#asahi-gpu · 2023-03-26 15:46
jannauthere was roughly 2 1/2 days of use before the GEM_BIND failures so it could be an issue that manifest only over time/use
#asahi-gpu · 2023-03-26 15:44
jannauI haven't seen any fb with explicit fence, most fbs from kwin get implicit fences
#asahi-gpu · 2023-03-26 15:16
linaGEM_BIND should never fail unless an invariant was violated (debug build should catch that) or something went very weirdly wrong with VA tracking or memory exhaustion or something...
#asahi-gpu · 2023-03-26 15:15
linaBut I suspect there is a bad codepath somewhere around that
#asahi-gpu · 2023-03-26 15:15
linajannau: It is expected that the FB has no explicit fence if and only if the render was issued before export, actually completed, and then the buffer was exported when it was known to be idle.
#asahi-gpu · 2023-03-26 14:51
jannausecond kwin_wayland segfault in apple_dri after failed DRM_IOCTL_ASAHI_GEM_BIND, unfortuantely without mesa debug symbols, mesa already recompiled
#asahi-gpu · 2023-03-26 13:05
jannauis it expected that the fb has no explicit fence?
#asahi-gpu · 2023-03-26 13:04
jannaulina: the pink boxes could be a fencing error. I see drm_gem_plane_helper_prepare_fb end up sometimes without fence
#asahi-gpu · 2023-03-25 22:23
handlerugthat's nice to hear
#asahi-gpu · 2023-03-25 22:23
handlerugoh I must have remembered my experience with 1.x
#asahi-gpu · 2023-03-25 22:21
flibitijibiboah yeah - i will say, those updated banks were actually done with the linux-native studio, the appimage. seemed to work great!
#asahi-gpu · 2023-03-25 22:03
handlerugconcerned about this as well
#asahi-gpu · 2023-03-25 22:03
handlerugflibitijibibo: got it! I think it'd be cleaner to have a newer FMOD version supported by the game itself anyway because correctly patching it requires FMOD Studio (which is either Windows/macOS or Wine) and obviously the player to sign up on the FMOD website (as I think there might be some issues with distributing their libraries and tooling), and I believe the modding community seems to be
#asahi-gpu · 2023-03-25 16:38
winterAh, okay. I figured since this is specifically a Mesa issue (not a GPU issue), it'd go there. Guess I should try to reproduce on ALARM Asahi, too.
#asahi-gpu · 2023-03-25 16:32
jannauI think the preferred place is still https://github.com/AsahiLinux/linux/issues/72

plus ›