etnaviv: Implement zero run length encoding of weights
etnaviv: Implement zero run length encoding of weights
Check how much smaller can the weight+bias buffers be with different amount of bits to encode runs of zeroes and choose the smaller one.
This reduces the bandwidth considerably, which is at present the bottleneck with useful models.
On a Libre Computer Alta AML-A311D-CC, I see these improvements:
MobileNetV1: 15.650ms -> 9.991ms SSDLite MobileDet: 56.149ms -> 32.692ms