[proto] Speed-up crop on bboxes and tests

[vfdev-5]

[-------- crop_bounding_box cpu BoundingBoxFormat.XYXY --------]
            |  crop_bounding_box_old v2  |  crop_bounding_box v2
1 threads: -----------------------------------------------------
      (4,)  |             50             |           20         
6 threads: -----------------------------------------------------
      (4,)  |             50             |           20         

Times are in microseconds (us).

[-------- crop_bounding_box cpu BoundingBoxFormat.XYWH --------]
            |  crop_bounding_box_old v2  |  crop_bounding_box v2
1 threads: -----------------------------------------------------
      (4,)  |             95             |           20         
6 threads: -----------------------------------------------------
      (4,)  |            106             |           20         

Times are in microseconds (us).

[------- crop_bounding_box cpu BoundingBoxFormat.CXCYWH -------]
            |  crop_bounding_box_old v2  |  crop_bounding_box v2
1 threads: -----------------------------------------------------
      (4,)  |            100             |           20         
6 threads: -----------------------------------------------------
      (4,)  |            100             |           20         

Times are in microseconds (us).

[------- crop_bounding_box cuda BoundingBoxFormat.XYXY --------]
            |  crop_bounding_box_old v2  |  crop_bounding_box v2
1 threads: -----------------------------------------------------
      (4,)  |            100             |           72         
6 threads: -----------------------------------------------------
      (4,)  |            100             |           71         

Times are in microseconds (us).

[------- crop_bounding_box cuda BoundingBoxFormat.XYWH --------]
            |  crop_bounding_box_old v2  |  crop_bounding_box v2
1 threads: -----------------------------------------------------
      (4,)  |            174             |          71.5        
6 threads: -----------------------------------------------------
      (4,)  |            170             |          70.9        

Times are in microseconds (us).

[------ crop_bounding_box cuda BoundingBoxFormat.CXCYWH -------]
            |  crop_bounding_box_old v2  |  crop_bounding_box v2
1 threads: -----------------------------------------------------
      (4,)  |            260             |           71         
6 threads: -----------------------------------------------------
      (4,)  |            264             |           71         

Times are in microseconds (us).

cc @datumbox @bjuncek @pmeier

[datumbox]

LGTM, just one minor comment.

[vfdev-5]

This is important to assign to bounding_box the output of bounding_box.sub_(sub) otherwise, bounding_box is still BoundingBox with wrong spatial_size

[github-actions]

Hey @vfdev-5!

You merged this PR, but no labels were added. The list of valid labels is available at https://github.com/pytorch/vision/blob/main/.github/process_commit.py

[pmeier]

Vasilis found that it is generally faster to do one regular operation instead of cloning and a single inplace operation. @vfdev-5 Have you benchmarked that as well?

[vfdev-5]

Yes, it faster by ~2us to go directly with bounding_box = bounding_box.sub(sub) instead of cloning and sub_.

[pmeier]

Given that we have used the idiom Vasilis found everywhere else, I think it would be good to also use it here especially if it is (marginally) faster. Do you want to send a PR or should I?

[vfdev-5]

I'll update in pad PR (#6890). Less lines of cleaner code is better then current code :)

[vfdev-5]

Funny, full benchmark shows ~10us improvements :)