[nomerge] Potential perf opt on bbox kernels

[datumbox]

No description provided.

[datumbox]

A quick check of:

def single(inpt, left, top):
    output = inpt.clone()
    output[..., 0] += left
    output[..., 1] += top
    output[..., 2] += left
    output[..., 3] += top

    return output


def concurrent(inpt, left, top):
    output = inpt.clone()
    output.add_(torch.tensor([left, top, left, top]))

    return output

Shows:

[-------- Pad cpu torch.float32 --------]
               |  single   |  concurrent 
1 threads: ------------------------------
      (16, 4)  |     60    |       20    
6 threads: ------------------------------
      (16, 4)  |     58    |       15    

Times are in microseconds (us).

There might be multiple opportunities for speed up like this.