07:26 <robmur01> icecream95: given what https://images.anandtech.com/doci/16694/GPU_24.png says, I'd imagine that native commands are ones that aren't emulated

07:44 <icecream95> robmur01: Okay. I guess the iterators would interrupt the MCU upon seeing an unsupported opcode, which can then do whatever it needs to?

07:44 <icecream95> That would probably be easy to verify if I actually had hardware

07:44 <icecream95> Assuming that the debug printfs can be read from software...

07:47 <icecream95> It would be possible to mess with the firmware binary to change how printf works if needed, but that might be too close to binary reverse-engineering

08:00 <robmur01> I don't have access to the specs so this really is just me inferring rather than any wink-wink-nudge-nudge, but it seems logical - not unlike a microcoded CPU, indeed

08:02 <robmur01> which would presumably also answer the question of whether the command set is really rich enough to natively run the firmware which emulates the command set...

08:22 <icecream95> robmur01: So far the only computational instruction I've seen is ADD. Maybe that's all that is natively supported?

08:22 <icecream95> Probably there is subtract as well, as ADD is 64-bit (I think) but has a 32-bit immediate

10:05 <icecream95> Hmm.. after Valgrind instrumentation of the Valhall blob, I see that it only uses four different locations from the CS user register pages

10:05 <icecream95> This sure beats using mprotect() to catch when the blob tries to read/write GPU memory

19:44 <cphealy> Are there any tests that can be run with ARM GPUs using Panfrost IOCTLs via libdrm (without Mesa)? I looked in the libdrm source and don't see anything like AMD or Vivante have, but perhaps someone is aware of some not yet upstream code?

19:45 <cphealy> I'm looking for some basic tests that exercise the IOCTLs and simple command execution on the GPU.

20:12 <robclark> cphealy: i-g-t is probably the place to look

20:17 <cphealy> robclark: for exercising Panfrost GPU ioctls?

20:18 <cphealy> I thought IGT was more about DRM/KMS drivers

20:34 <robclark> gpu stuff too (although those are driver specific tests)

20:35 <cphealy> sounds good, I'll take a look.

21:16 <anholt_> yeah, igt would be the place for such tests to be put if they existed (but they likely don't)

22:46 <robclark> there are a few panfrost tests in igt.. not sure if any of 'em actually run anything on gpu or if only ioctl tests