#
✕ nick:pac85
#asahi-re · 2026-02-21 22:19
pac85Shiz: Ty!
#asahi-re · 2026-02-21 20:04
pac85https://www.irccloud.com/pastebin/aucTMxMd/
#asahi-re · 2026-02-21 20:04
pac85fl0_id: great! For more context the two moves at the top contain an hardcoded address to a buffer containg two pointers which in turn point to the two kernel args I'm passing. For a shader that takes no args and does nothing this shader looks like this:
#asahi-re · 2026-02-21 19:49
pac85(that was after adding the lengths of those instrs)
#asahi-re · 2026-02-21 19:48
pac85https://www.irccloud.com/pastebin/3qqvokjj/
#asahi-re · 2026-02-21 19:48
pac85fl0_id: yeah I was aware, that's the disassembler I was referring to. Here are the dumps if anyone wants to look
#asahi-re · 2026-02-21 19:28
pac85We thought perhaps we can post dumps of such shaders so that people looking into the isa and that are more familiar with the encoding can perhaps shine some light on this.
#asahi-re · 2026-02-21 19:28
pac85It seems like the CDM points to a shader that somehow jumps to the real shader. That uses a bunch of instructions that the current disassembler cannot parse.
#asahi-re · 2026-02-21 19:28
pac85We have enough to dump the CDM and shaders. This is where we made a finding that we felt was worth sharing:
#asahi-re · 2026-02-21 19:28
pac85So we figured out the new macos uapi, and started reversing the new cmdstream.
#asahi-re · 2026-02-21 19:28
pac85Me and somebody else decided to have some fun reverse engineering the M3 gpu in the weekends.
#asahi-re · 2026-02-21 19:28
pac85Hi there.