Back-propagation is too slow with propagate_adjoints

[creotiv]

Hi guys, im new to Halide, and have problem with understanding why back-propagation is so slow in comparing with forward run.

Im trying to make operation for pytorch.
Here is my code: https://github.com/creotiv/halide_bilateral_slice_apply

Would be very appreciate if someone can say where is a problem.
Thanks

[BachiLi]

You need to schedule the backward pass. You can use the autoschedulers to do so.

[creotiv]

But i used it:

output
                   .fuse(x, y, xy)
                   .fuse(xy, c, xyc)
                   .fuse(xyc, n, xycn)
                   .gpu_tile(xycn, tx, 4)

[BachiLi]

Sorry I missed it. Those are not very efficient schedules. Have you tried the autoschedulers?