06:00 tomeu: pH5: I'm surprised you needed to hide some code behind VERSION == VIPPICO_V3_PID0X99 , because I had been testing for VIP8000NANOSI_PLUS_PID0X9F
06:02 tomeu: in any case, I guess the correct thing to do would be to skip the rest of the 64-byte aligned header
06:13 tomeu: and the change in get_symbol breaks the weights I'm testing with
06:14 tomeu: guess I haven't been testing enough situations yet, but I will be expanding testing soon
07:23 tomeu: and I got my first generated kernel running on the s905d3 \o/
07:24 tomeu: pH5: your huffman encoder worked great as-is
07:24 tomeu: it's just the in-memory format that has changed a bit
07:25 tomeu: to me it seems as if amlogic complained about performance being still bad when they first tested the v7-based IP, and then VSI giving them a half-cooked huffman-based format
07:26 tomeu: then the simulator and the c-based flipflop code in galcore remaining in the v7 format for the s905d3, as VSI did only change what was strictly needed to get things working
07:27 pH5: tomeu: \o/
07:29 pH5: there seems to be a difference in the run length table indexing though, on imx8mp anything that uses rlt entries other than the first doesn't decode correctly.
07:29 pH5: so very plausible that this is some kind of in-between version
07:29 tomeu: yeah
09:03 tomeu: and the symbol map that your code generates works fine, just I guess it's not as efficient in compressing
09:03 tomeu: and 0, 1, 0, 1 as the rlt seems to work as well fine
09:08 tomeu: Pass: 5, Duration: 0, Remaining: 0
09:08 tomeu: nice start :)