What does ConvTranspose2d's output_padding fill

vision
1

ConvTranspose2d’s output_padding not fill zeor-padding.
I tried to read source code to find it，but failed. I want to know how pytorch caculate it