05:41 tomeu: phh: that would be great, yeah
05:43 tomeu: build/src/gallium/targets/teflon/libteflon.so once you build mesa, you can load that as an external delegate in, say, tf's label_image.py or the benchmark tool
05:45 tomeu: will send you the model I'm using to test
09:00 tomeu: hmm, there are calls to RKNPU_GET_DRV_VERSION before some syncs, so I guess something changed in the kernel driver that affects what syncs are required
09:35 tomeu: ok, found it
09:36 tomeu: I wasn't ORing the target in each register write with 0x1
09:36 tomeu: and my cunning decoder script was hiding that, as the target is largely redundant...
10:47 tomeu: uff, finally got some tests passing
10:48 tomeu: let's see if I can pull some data conversion code from the NVDLA user mode driver
10:57 phh: 🍾
10:59 tomeu: 🚀
16:34 tomeu: awesome :)
16:34 tomeu: Pass: 70, ExpectedFail: 78, Skip: 12, Duration: 2, Remaining: 0
16:35 tomeu: a good start
16:35 phh: did any actual computation passed with result checked?
16:37 tomeu: of course!
16:37 phh: cool
16:37 tomeu: that's a bunch of convolutions, but with input channels 1 or 32, and with just 1 output channel
16:37 phh: so you already have a working matmul?
16:37 tomeu: but with strides, padding and depthwise
16:38 phh: (I don't know if conv can get you reasonable "free" matmul)
16:38 tomeu: well, I'm targeting tflite's Conv2D and DepthwiseConv2D operations for now
16:38 phh: ok
16:40 tomeu: maybe, I think the question is whether the frontend won't cause troubles there
16:40 tomeu: but the MAC cells are square, iiuic