[proto] Small optim for perspective op on images

[vfdev-5]

• Small optim for perspective op on images
• reverted concat+single matmul trick on bboxes as it does not give any speed up and brings a slowdown for images.

[----------------------------------- Perspective_image_tensor cpu ----------------------------------]
                                   |  perspective_image_tensor_old v2  |  perspective_image_tensor v2
1 threads: ------------------------------------------------------------------------------------------
      (3, 400, 500) torch.uint8    |                4.8                |              4.64           
      (3, 400, 500) torch.float32  |                4.2                |              4.14           
6 threads: ------------------------------------------------------------------------------------------
      (3, 400, 500) torch.uint8    |                3.7                |              3.67           
      (3, 400, 500) torch.float32  |                3.6                |              3.52           

Times are in milliseconds (ms).

[---------------------------------- Perspective_image_tensor cuda ----------------------------------]
                                   |  perspective_image_tensor_old v2  |  perspective_image_tensor v2
1 threads: ------------------------------------------------------------------------------------------
      (3, 400, 500) torch.uint8    |                633                |              585            
      (3, 400, 500) torch.float32  |                549                |              543            
6 threads: ------------------------------------------------------------------------------------------
      (3, 400, 500) torch.uint8    |                594                |              586            
      (3, 400, 500) torch.float32  |                551                |              545            

Times are in microseconds (us).

cc @datumbox @bjuncek @pmeier

[datumbox]

Thanks @vfdev-5 just 2 questions. Looks good overall.

[datumbox]

Can we do in-place division?

[datumbox]

Just to understand, previously when we measured this specific change for bboxes it didn't improve the speed?

[vfdev-5]

I measured all changes together: concat+single matmul + inplace + aminmax

[pmeier]

IIUC, it improved the speed for bounding boxes, but the PR didn't measure the impact on images. Since bounding box tensors are usually a lot smaller than images, the perf regression for images was not noticed there.

[vfdev-5]

There were no perf regression at all as it was using previous implementation. For bboxes shape=(1000, 4) "concat+single matmul" only trick does not bring any speed up on cpu. While working on images (this PR) I see that same trick brings even a slowdown.

[datumbox]

OK if you measured it now and you know it doesn't slow us down it's fine by me.

[datumbox]

LGTM, feel free to merge after testing the in-place division.