Jean-Marc Valin
Jean-Marc Valin
It's probably just some kind of difference in the build system that causes the Neon/dotprod instructions not to be used and defaulting to scalar code.
I didn't actually create that function pointer setup, so I'm not sure of all the reasoning behind it. What I'd suggest though is to make sure you run the randomized...
merged
Are you able to reproduce with opus-tools?
Can you provide the test file you use and ideally the exact options you used to build libopus, opusenc, and opus-tools?
It's the same conversion. It's just that the "little" models were trained with sparsification (which isn't yet well documented, sorry)
The number of weights in gru1_input_weights_float should be equal to 3*density*nb_inputs*nb_outputs.
Yes, we're aware of the issue and fixing it is on the todo list
Also has the same issue with __m128i_u which isn't portable.