How does "conv2d" divide image into different GPUs while using -unified-gpus

[wangw42]

It seems during the Im2col process in conv2d, the pages belonging to im2colMatirx are accessed mostly by only one GPU, however, a small part of pages are accessed by 2 GPUs. This happens even when I set the N the same as the number of GPUs.
So I wonder whether the images or matrix are divided into multiple parts and allocated to different GPUs. Or It doesn't divide data during initialization, just pinned the data to the GPU that first requests the data?

Thank you! :D