16:19tomeu: finally! got ssdlite mobiledet working on the rk3588
16:20tomeu: 32ms in a single 2 TOPS core
16:22phh: cool
16:22tomeu: the blob takes twice as much, don't know why...
16:23phh: does that include exec() library loading, or is it like ioctl-to-ioctl?
16:23tomeu: hmm, what do you mean by these?
16:24tomeu: this is my Mesa driver, calling into their kernel driver
16:24phh: I mean how do you measure the time taken by the blob?
16:25tomeu: just measuring wall clock time around
16:25tomeu: outputs = rknn.inference(inputs=[input], data_type="uint8")
16:26phh: after previous runs in case they "compile" the model or something?
16:29tomeu: ah, I would expect it to be ready to inference, but I haven't studied they API in depth (do they even have docs?)
16:30phh: that's my expectation as well, but I think we've already established the blob isn't super reasonable...
16:31tomeu: yeah...
18:56f_: >finally! got ssdlite mobiledet working on the rk3588
18:56f_: tomeu: \o/ !!
22:21lvrp16_: wow...
22:22lvrp16_: can't wait for the rest of tRK3588 to catch up