05:40 tomeu: ah, np, I found the regcmd buffer inside the .rknn file
08:27 tomeu: wow, so few unknown registers: https://gitlab.freedesktop.org/tomeu/mesa/-/commit/6bb2f55e825030fe6087b493a0d6b9fe3dffca15
08:27 tomeu: this is starting to look like a nice walk in the park
13:05 phh: uh I think weight_kernel is the channel number?
13:25 phh: CPEND/bs_ow is perfectly clear.
13:55 tomeu: phh: what is CPEND?
13:56 phh: tomeu: bs_ow registers are described as cpend in the TRM
13:57 phh: I have no idea what that could be
13:57 tomeu: ah, right
13:57 tomeu: something about tensor manipulation, I think
14:02 phh: BN right after bs section should be Batch Normalisation. So bs, Batch....?
14:03 tomeu: Bull...
14:04 phh: at least based on the available understandable operations in bs (Add, Minus, RELU, MUL RELU, MUL) it's an element-wise operation
14:04 tomeu: btw, I just got confirmation that rockchip is based on NVDLA, though that was already pretty obvious
14:04 phh: cool
14:05 phh: in nvdla documentaion, after batch normalisation there is bias addition, which is kinda what's happening here (though still don't know what that S could mean)
14:06 tomeu: sum?
14:06 tomeu: maybe they translated the NVDLA docs to mandarin, then back to english?
14:06 tomeu: would explain why they changed so many terms
14:06 phh: well it shouldn't be since it's element-wise, you'd be just adding one number from the weights
14:07 tomeu: true
14:07 phh: but yeah maybe poor translations can do that indeed
14:09 phh: likewise the TRM mention RELUX activation function, which the internet just doesn't do. could be some kind of translation of parametrized ReLU? Though the TRM says "Value of the RELUX compare with", and the parameter in pReLU isnt' a value for comparaison
14:11 tomeu: I'm not overly concerned by these though, as we still have quite some work to do until we get past the simple models such as MobileNet
14:15 phh: I wouldn't be surprised that relux is a translated relu, in which case it would happen on mobilenet
14:16 tomeu: translated in what sense?
14:18 phh: ReLU is 0 when x < 0; x when x > 0; ReLUX could be 0 when x < RKNN_dpu_bn_relux_cmp_value; (x-RKNN_dpu_bn_relux_cmp_value) when x > RKNN_dpu_bn_relux_cmp_value
14:19 tomeu: sounds plausible!