[Enhancement]Replace numpy ascontiguousarray with torch contiguous to…

… speed-up (#2604)

## Motivation

Original motivation was after [MMDetection PR
#9533](open-mmlab/mmdetection#9533)

With several experiments I found out that if a ndarray is contiguous,
numpy.transpose + torch.contiguous perform better, while if not, then
use numpy.ascontiguousarray + numpy.transpose

## Modification

Replace numpy.ascontiguousarray with torch.contiguous in
[PackSegInputs](https://github.com/open-mmlab/mmsegmentation/blob/1.x/mmseg/datasets/transforms/formatting.py)

Co-authored-by: MeowZheng <meowzheng@outlook.com>

mmseg/datasets/transforms/formatting.py

@@ -63,8 +63,12 @@ def transform(self, results: dict) -> dict:
  img = results['img']
  if len(img.shape) < 3:
  img = np.expand_dims(img, -1)
- img = np.ascontiguousarray(img.transpose(2, 0, 1))
- packed_results['inputs'] = to_tensor(img)
+ if not img.flags.c_contiguous:
+ img = to_tensor(np.ascontiguousarray(img.transpose(2, 0, 1)))
+ else:
+ img = img.transpose(2, 0, 1)
+ img = to_tensor(img).contiguous()
+ packed_results['inputs'] = img
 
  data_sample = SegDataSample()
  if 'gt_seg_map' in results: