Questions about the window-major feature map organization

[jacksonsc007]

I was deeply impressed by the window-major feature map organization proposed in the paper and I checked the implementation.

However, It seems that Pytorch does the window-major feature map organization automatically for us when we want to perform window attention (through the copy and re-organization built in the reshape function for non-contiguous tensor) and I could not come up with a way to calculate window attention with row-major feature map organization.

What I want is to write a code to make clear efficiency comparison between these two organization schemes. Is there any code available? Or any suggestion?

[jacksonsc007]

I checked the code of VitDet and I think I kind of figure it out.