|
onnx2versal
|
Vector implementation for 1x1 QLinearConv, streams weights requires data to be reshaped from (M,C,1,1) to (M,C') where C' is padded to next multiple of 16, requires bias to be shifted, i.e. tbias - tw_1x1.reshape(M,-1).sum(1) * X_zero_point, requires KH==KW==1, INP_W%16=0, OUT_W_PAD%16=0, STEP_H==1|2, STEP_W==1|2,.
#include <qlinearconv.h>
Static Public Member Functions | |
| static void | registerKernelClass () |