Invert CV-CUDA Backend

test/test_transforms_v2.py

@@ -21,6 +21,7 @@
 import torchvision.transforms.v2 as transforms
 
 from common_utils import (
+    assert_close,
     assert_equal,
     cache,
     cpu_and_cuda,
@@ -41,7 +42,6 @@
 )
 
 from torch import nn
-from torch.testing import assert_close
 from torch.utils._pytree import tree_flatten, tree_map
 from torch.utils.data import DataLoader, default_collate
 from torchvision import tv_tensors
@@ -5824,7 +5824,18 @@ def test_kernel_image(self, dtype, device):
     def test_kernel_video(self):
         check_kernel(F.invert_video, make_video())
 
-    @pytest.mark.parametrize("make_input", [make_image_tensor, make_image, make_image_pil, make_video])
+    @pytest.mark.parametrize(
+        "make_input",
+        [
+            make_image_tensor,
+            make_image,
+            make_image_pil,
+            make_video,
+            pytest.param(
+                make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="test requires CVCUDA")
+            ),
+        ],
+    )
     def test_functional(self, make_input):
         check_functional(F.invert, make_input())
 
@@ -5835,20 +5846,51 @@ def test_functional(self, make_input):
             (F._color._invert_image_pil, PIL.Image.Image),
             (F.invert_image, tv_tensors.Image),
             (F.invert_video, tv_tensors.Video),
+            pytest.param(
+                F._color._invert_image_cvcuda,
+                None,
+                marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="test requires CVCUDA"),
+            ),
         ],
     )
     def test_functional_signature(self, kernel, input_type):
+        if kernel is F._color._invert_image_cvcuda:
+            input_type = _import_cvcuda().Tensor
         check_functional_kernel_signature_match(F.invert, kernel=kernel, input_type=input_type)
 
-    @pytest.mark.parametrize("make_input", [make_image_tensor, make_image_pil, make_image, make_video])
+    @pytest.mark.parametrize(
+        "make_input",
+        [
+            make_image_tensor,
+            make_image_pil,
+            make_image,
+            make_video,
+            pytest.param(
+                make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="test requires CVCUDA")
+            ),
+        ],
+    )
     def test_transform(self, make_input):
         check_transform(transforms.RandomInvert(p=1), make_input())
 
+    @pytest.mark.parametrize(
+        "make_input",
+        [
+            make_image,
+            pytest.param(
+                make_image_cvcuda, marks=pytest.mark.skipif(not CVCUDA_AVAILABLE, reason="test requires CVCUDA")
+            ),
+        ],
+    )
     @pytest.mark.parametrize("fn", [F.invert, transform_cls_to_functional(transforms.RandomInvert, p=1)])
-    def test_correctness_image(self, fn):
-        image = make_image(dtype=torch.uint8, device="cpu")
+    def test_correctness_image(self, make_input, fn):
+        image = make_input(dtype=torch.uint8, device="cpu")
 
         actual = fn(image)
+
+        if make_input is make_image_cvcuda:
+            image = F.cvcuda_to_tensor(image)[0].cpu()
+
         expected = F.to_image(F.invert(F.to_pil_image(image)))
 
         assert_equal(actual, expected)

torchvision/transforms/v2/_color.py

@@ -5,6 +5,7 @@
 import torch
 from torchvision import transforms as _transforms
 from torchvision.transforms.v2 import functional as F, Transform
+from torchvision.transforms.v2.functional._utils import _is_cvcuda_tensor
 
 from ._transform import _RandomApplyTransform
 from ._utils import query_chw
@@ -282,6 +283,8 @@ class RandomInvert(_RandomApplyTransform):
 
     _v1_transform_cls = _transforms.RandomInvert
 
+    _transformed_types = _RandomApplyTransform._transformed_types + (_is_cvcuda_tensor,)
+
     def transform(self, inpt: Any, params: dict[str, Any]) -> Any:
         return self._call_kernel(F.invert, inpt)
 

torchvision/transforms/v2/functional/_color.py

@@ -1,3 +1,5 @@
+from typing import TYPE_CHECKING
+
 import PIL.Image
 import torch
 from torch.nn.functional import conv2d
@@ -9,7 +11,13 @@
 
 from ._misc import _num_value_bits, to_dtype_image
 from ._type_conversion import pil_to_tensor, to_pil_image
-from ._utils import _get_kernel, _register_kernel_internal
+from ._utils import _get_kernel, _import_cvcuda, _is_cvcuda_available, _register_kernel_internal
+
+
+CVCUDA_AVAILABLE = _is_cvcuda_available()
+
+if TYPE_CHECKING:
+    import cvcuda  # type: ignore[import-not-found]
 
 
 def rgb_to_grayscale(inpt: torch.Tensor, num_output_channels: int = 1) -> torch.Tensor:
@@ -680,6 +688,43 @@ def invert_video(video: torch.Tensor) -> torch.Tensor:
     return invert_image(video)
 
 
+_invert_cvcuda_tensors: dict[str, "cvcuda.Tensor"] = {}
+
+
+def _invert_image_cvcuda(image: "cvcuda.Tensor") -> "cvcuda.Tensor":
+    cvcuda = _import_cvcuda()
+
+    # save the tensors into a dictionary only if CV-CUDA is actually used
+    # we save these here, since they are static and small in size
+    if "base" not in _invert_cvcuda_tensors:
+        _invert_cvcuda_tensors["base"] = cvcuda.as_tensor(
+            torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32, device="cuda").reshape(1, 1, 1, 3).contiguous(), "NHWC"
+        )
+    if "scale" not in _invert_cvcuda_tensors:
+        _invert_cvcuda_tensors["scale"] = cvcuda.as_tensor(
+            torch.tensor([-1.0, -1.0, -1.0], dtype=torch.float32, device="cuda").reshape(1, 1, 1, 3).contiguous(),
+            "NHWC",
+        )
+
+    base = _invert_cvcuda_tensors["base"]
+    scale = _invert_cvcuda_tensors["scale"]
+
+    if image.dtype == cvcuda.Type.U8:
+        shift = 255.0
+    elif image.dtype == cvcuda.Type.F32:
+        shift = 1.0
+    else:
+        raise ValueError(f"Input image dtype must be uint8 or float32, got {image.dtype}")
+
+    # Use normalize to invert: output = (input - base) * scale * global_scale + shift
+    # For inversion: output = (input - 0) * (-1) * 1 + shift = shift - input
+    return cvcuda.normalize(image, base=base, scale=scale, globalscale=1.0, globalshift=shift)
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(invert, _import_cvcuda().Tensor)(_invert_image_cvcuda)
+
+
 def permute_channels(inpt: torch.Tensor, permutation: list[int]) -> torch.Tensor:
     """Permute the channels of the input according to the given permutation.