search

NNgen / nngen

NNgen: A Fully-Customizable Hardware Synthesis Compiler for Deep Neural Network
Apache License 2.0
339 stars 46 forks source link

fix verify matmul #36

Closed RyusukeYamano closed 2 years ago

RyusukeYamano commented 3 years ago

The eval mode is broken in the case of non transposing B in matmul A@B operation. Similarly, the case of transposing A (A^T@B) seems to be broken, but I haven't been able to track down the details of this case.

RyusukeYamano commented 2 years ago

The rounding process also differed from the hardware implementation, so this has been corrected.

RyusukeYamano commented 2 years ago

I have confirmed that hallo_nngen.py passes. image