16:14tomeu: uff, finally managed extracting the registers from the TRM: https://gitlab.freedesktop.org/tomeu/mesa/-/commit/8aaef982265062ee7f9dfd059effe49b75887f02#32988b192fc568726cdf1f177590799cd1184535
16:15tomeu: tomorrow I hope to get the HW and can start crafting a simple convolution operation
16:15tomeu: it's so great that rockchip documented the registers, will make everything so much quicker
16:19phh: ah, hex byte offsets, much better to read :D
16:37tomeu: also got a decoder script using that same xml
16:38tomeu: phh: but in my tests here, I don't see the blob allocating a regcmd buffer for my single-convolution models
16:38tomeu: do you know where the registers might be in that case?
19:10phh: tomeu: like you list and dump the gems and nothing look like a command buffer?
20:18tomeu: no, rather that I intercept the ioctls that create, map, etc buffers, and for the smaller models I don't see regcmd buffers created
20:21tomeu: also, the blob writes such messages when it does:
20:21tomeu: D RKNN: [21:18:52.457] allocated memory, name: regcmd, virt addr: 0xfffee7cd8000, dma addr: 0x9, obj addr: 0x7, size: 270336, aligned size: 270336, fd: 0, handle: 1, flags: 0x3, gem name: 0
20:28phh: well i'm guessing for simple operations they don't need ping/pong, so they just registers mmap directly
20:29phh: let me check if vendor kernel drivers can do that
20:36phh: I see no such capability in vendor driver :/
20:46phh: tomeu: got a strace just to see if there is an ioctl I missed?
21:02phh: (please don't be /dev/mem)