04:46 <laanwj> marex: you might be right but not sure, if it's anything similar to the KPU on the sipeed K210 - https://github.com/laanwj/k210-sdk-stuff/blob/master/doc/kpu.md , it's a low-preciesion dense matrix multiplication engine to accelerate some specific (convolutional) layer types in nncase networks

04:53 <laanwj> something that seems very doable with OpenCL code, i'm not sure that specific peripehral would add much if it has a high end Vivante GPU, so maybe it's something else entirely

23:08 <marex> hmmm, why would dEQP-GLES2.functional.texture.vertex.2d.filtering.nearest_nearest_clamp fails with ETNA_MESA_DEBUG=nir and not with ETNA_MESA_DEBUG=nir,flush_all ?