04:57 tomeu: pH5: have pushed the script to https://gitlab.freedesktop.org/tomeu/mesa/-/commit/56ba2feaa4c3e70be50317a271270fb1f6cdd512
04:57 tomeu: seems to work for both NPUs in the very limited testing I have done
07:57 pH5: wait, so same encoding after all? that's awesome!
08:27 daniels: amazing work :)
09:48 pH5: tomeu: with https://gitlab.freedesktop.org/pH5/mesa/-/commit/640852d5c9f4531a26a46c2cd93408e235610c2b it matches most of my test cases
09:49 pH5: https://gitlab.freedesktop.org/pH5/ask-tim/-/blob/encoder-test/src/test7.h is the odd one out, that one doesn't use the zero run length table.
16:01 tomeu: <pH5> "https://gitlab.freedesktop.org/..." <- ah, that's interesting. so we could ignore zrl encoding if we wanted
16:01 tomeu: tomeu: but it is probably better to skip that step, as IMO it is very much required for performance on this NPUs, as they don't support natively strided and depthwise convolutions