[prototype] Speed up adjust_contrast_image_tensor

[datumbox]

Related to #6818

Small performance improvement for uint8 images by avoiding the double casting from floats to ints and back to floats.

Floats remain unaffected, a 5% boost on uint8s:

[----------- adjust_contrast_image_tensor cpu torch.float32 -----------]
                         |  adjust_contrast_image_tensor old  |  fn2 new
1 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |               13600                |   13500 
      (3, 400, 400)      |                 538                |     535 
6 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |               14300                |   14300 
      (3, 400, 400)      |                 838                |     834 

Times are in microseconds (us).

[---------- adjust_contrast_image_tensor cuda torch.float32 -----------]
                         |  adjust_contrast_image_tensor old  |  fn2 new
1 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |                196                 |    196  
      (3, 400, 400)      |                 64                 |     62  
6 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |                197                 |    197  
      (3, 400, 400)      |                 64                 |     62  

Times are in microseconds (us).

[------------ adjust_contrast_image_tensor cpu torch.uint8 ------------]
                         |  adjust_contrast_image_tensor old  |  fn2 new
1 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |               26100                |   25000 
      (3, 400, 400)      |                1109                |     999 
6 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |               28300                |   30000 
      (3, 400, 400)      |                1700                |    1550 

Times are in microseconds (us).

[----------- adjust_contrast_image_tensor cuda torch.uint8 ------------]
                         |  adjust_contrast_image_tensor old  |  fn2 new
1 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |               257.3                |    242  
      (3, 400, 400)      |                89.7                |     79  
6 threads: -------------------------------------------------------------
      (16, 3, 400, 400)  |               258.1                |    242  
      (3, 400, 400)      |                90.2                |     80  

Times are in microseconds (us).

cc @vfdev-5 @bjuncek @pmeier

[datumbox]

@pmeier In an early iteration of the PR, I have missed this floor_() call which is necessary to reproduce identical results with stable. Unfortunately the tests passed without catching the issue. Might be worth checking that our reference tests check both on floats and ints.

[pmeier]

This is properly tested with

vision/test/prototype_transforms_kernel_infos.py

Lines 1845 to 1850 in f32600b

	def reference_inputs_adjust_contrast_image_tensor():
	for image_loader, contrast_factor in itertools.product(
	make_image_loaders(color_spaces=(features.ColorSpace.GRAY, features.ColorSpace.RGB), extra_dims=[()]),
	_ADJUST_CONTRAST_FACTORS,
	):
	yield ArgsKwargs(image_loader, contrast_factor=contrast_factor)

but it seems our tolerances are to high to pick up on this

vision/test/prototype_transforms_kernel_infos.py

Line 1866 in f32600b

closeness_kwargs=DEFAULT_PIL_REFERENCE_CLOSENESS_KWARGS,

vision/test/prototype_transforms_kernel_infos.py

Lines 64 to 67 in f32600b

	DEFAULT_PIL_REFERENCE_CLOSENESS_KWARGS = {
	(("TestKernels", "test_against_reference"), torch.float32, "cpu"): dict(atol=1e-5, rtol=0, agg_method="mean"),
	(("TestKernels", "test_against_reference"), torch.uint8, "cpu"): dict(atol=1e-5, rtol=0, agg_method="mean"),
	}

However, we can't remove the tolerances completely here, since the output differs even with the current implementation.

[datumbox]

Hmm... 99% of the performance optimizations were tested for training and we know they are correct. So there might be rounding error issues. I think we should adjust the tolerances to the degree that will catch BC issues but not go ballistic for the 8th decimal. That might be a hard exercise though.

[vfdev-5]

Looks good, thanks @datumbox

[datumbox]

@vfdev-5 thanks. It seems there is an unrelated flaky test about resizing bboxes:
https://github.com/pytorch/vision/actions/runs/3427555648/jobs/5710704975

Could you have a look?

FAILED test/test_prototype_transforms_functional.py::TestKernels::test_against_reference[resize_bounding_box-55] - AssertionError: Tensor-likes are not equal!

Mismatched elements: 2 / 16 (12.5%)
Greatest absolute difference: 1 at index (2, 2)
Greatest relative difference: 0.25 at index (2, 3)

The failure occurred for item [0]
FAILED test/test_prototype_transforms_functional.py::TestKernels::test_against_reference[resize_bounding_box-56] - AssertionError: Tensor-likes are not equal!

Mismatched elements: 2 / 16 (12.5%)
Greatest absolute difference: 1 at index (2, 2)
Greatest relative difference: 0.25 at index (2, 3)

The failure occurred for item [0]
FAILED test/test_prototype_transforms_functional.py::TestKernels::test_against_reference[resize_bounding_box-57] - AssertionError: Tensor-likes are not equal!

Mismatched elements: 2 / 16 (12.5%)
Greatest absolute difference: 1 at index (2, 2)
Greatest relative difference: 0.25 at index (2, 3)

The failure occurred for item [0]