#asahi-gpu · 2021-03-23 21:11
chrisfDarkShadow44: when the buffer is a parameter to the entrypoint function you know it's uniform across all lanes#asahi-gpu · 2021-03-23 21:11
chrisfDarkShadow44: when the buffer is a parameter to the entrypoint function you know it's uniform across all lanes#asahi-gpu · 2021-03-23 21:09
chrisfDarkShadow44: some of the details of that pseudocode are wrong [what happens with a hole in the mask, etc], i had pretty incomplete information at the time#asahi-gpu · 2021-03-23 21:09
chrisfDarkShadow44: some of the details of that pseudocode are wrong [what happens with a hole in the mask, etc], i had pretty incomplete information at the time#asahi-gpu · 2021-03-23 21:05
chrisfdougall has since figured out a lot more of the encoding but we're short a description of exactly what happens#asahi-gpu · 2021-03-23 21:05
chrisfdougall has since figured out a lot more of the encoding but we're short a description of exactly what happens#asahi-gpu · 2021-03-23 21:03
chrisfDarkShadow44: there's a better description of the "simple" loads and stores in the older docs at https://hthh.github.io/m1-gpu-re/#asahi-gpu · 2021-03-23 21:03
chrisfDarkShadow44: there's a better description of the "simple" loads and stores in the older docs at https://hthh.github.io/m1-gpu-re/#asahi-gpu · 2021-03-23 21:02
chrisfDarkShadow44: should be the base address of the buffer#asahi-gpu · 2021-03-23 21:02
chrisfDarkShadow44: should be the base address of the buffer#asahi-gpu · 2021-03-18 15:19
chrisfmarcan: on socs designed for android the display controllers are often enormous -- tons of planes, etc#asahi-gpu · 2021-03-18 15:18
chrisfmarcan: do we expect the apple display controller to be small like desktop, or a mobile-style behemoth?#asahi-gpu · 2021-03-17 22:47
chrisftease out the unknown things one by one#asahi-gpu · 2021-03-17 22:46
chrisfpossibly look at the (much simpler) command stream for compute shader only#asahi-gpu · 2021-03-17 22:44
chrisfmorelightning[m]: if you're interested in command stream side, compare what you're getting with what `demo` does#asahi-gpu · 2021-03-17 20:42
chrisfthat seems fairly benign as things going wrong go ;)#asahi-gpu · 2021-03-17 20:40
chrisfstick lldb on it an see how the wrapper went wrong?#asahi-gpu · 2021-03-17 20:33
chrisfbloom: ^#asahi-gpu · 2021-03-17 20:33
chrisfsuggests we dont understand something#asahi-gpu · 2021-03-17 19:45
chrisfhaving to interact with an IOSurface adds a lot of complexity#asahi-gpu · 2021-03-17 19:45
chrisfthe drawing example i looked at was offscreen, yes#asahi-gpu · 2021-03-17 19:32
chrisfmorelightning[m]: figure out why :)#asahi-gpu · 2021-03-17 19:24
chrisfhttps://developer.apple.com/documentation/metal/basic_tasks_and_concepts/performing_calculations_on_a_gpu?preferredLanguage=occ#asahi-gpu · 2021-03-17 19:13
chrisfmorelightning[m]: and with at least one trivial drawing sample#asahi-gpu · 2021-03-17 19:13
chrisfmorelightning[m]: it works with the MetalComputeBasic sample#asahi-gpu · 2021-03-17 19:13
chrisfmorelightning[m]: simpler example, or dig into why it doesnt work#asahi-gpu · 2021-03-17 19:00
chrisf:+1:#asahi-gpu · 2021-03-17 18:57
chrisfDarkShadow4444: *no* taking the driver binaries apart, etc. only observe their behavior from the outside.#asahi-gpu · 2021-03-17 18:56
chrisfDarkShadow4444: note the topic#asahi-gpu · 2021-03-17 18:41
chrisfthat sounds correct -- and the Dt field is cache control#asahi-gpu · 2021-03-17 18:41
chrisfim only half paying attention#asahi-gpu · 2021-03-17 18:39
chrisfbloom, dougall jump in any time eh :)#asahi-gpu · 2021-03-17 18:38
chrisfbut it's given up its secrets fairly readily so far#asahi-gpu · 2021-03-17 18:38
chrisfi mean, you have to run macos ;)#asahi-gpu · 2021-03-17 18:38
chrisfit's really not too unpleasant#asahi-gpu · 2021-03-17 18:37
chrisfeverything interesting is done in userspace, and you can get underneath it#asahi-gpu · 2021-03-17 18:36
chrisfyes, and there's only a few -- of the "allocate some memory", "execute this command buffer in memory", variety#asahi-gpu · 2021-03-17 18:35
chrisfthe wrapper intercepts the IOKit calls to the kernel driver#asahi-gpu · 2021-03-17 18:29
chrisfmorelightning[m]: all inside mac os#asahi-gpu · 2021-03-17 18:26
chrisfyou run them :)#asahi-gpu · 2021-03-17 18:25
chrisfmorelightning[m]: everything so far is done by intercepting between the userspace driver (which does the "metal" stuff) and the kext#asahi-gpu · 2021-03-17 18:16
chrisfpretty much. dougall has some other tricks, but ^ works.#asahi-gpu · 2021-03-17 18:13
chrisfwe also need to reverse the bulk of the command stream structures. `demo` in the gpu repo has a lot of mysterious stuff in it still#asahi-gpu · 2021-03-17 18:11
chrisfhave you got an M1 to poke at?#asahi-gpu · 2021-03-17 18:11
chrisfauthor is here in the channel :)#asahi-gpu · 2021-03-17 18:10
chrisfDarkShadow4444: one big area of useful things would be to help dougall flesh out the missing pieces#asahi-gpu · 2021-03-17 18:08
chrisfDarkShadow4444: https://dougallj.github.io/applegpu/docs.html#asahi-gpu · 2021-03-17 18:05
chrisfdougall: any chance you could fill out the device_load etc docs a bit more? there's clearly a bunch of stuff rattling around in your head only, about typed loads etc#asahi-gpu · 2021-02-16 00:51
chrisfdougall: neat!#asahi-gpu · 2021-02-15 02:07
chrisfso can we do big strided & masked stores to uniform space just like device space stores?