AdjustSharpness 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
@@ -6004,7 +6004,18 @@ def test_kernel_image(self, dtype, device):
     def test_kernel_video(self):
         check_kernel(F.adjust_sharpness_video, make_video(), sharpness_factor=0.5)
 
-    @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.adjust_sharpness, make_input(), sharpness_factor=0.5)
 
@@ -6015,12 +6026,30 @@ def test_functional(self, make_input):
             (F._color._adjust_sharpness_image_pil, PIL.Image.Image),
             (F.adjust_sharpness_image, tv_tensors.Image),
             (F.adjust_sharpness_video, tv_tensors.Video),
+            pytest.param(
+                F._color._adjust_sharpness_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._adjust_sharpness_image_cvcuda:
+            input_type = _import_cvcuda().Tensor
         check_functional_kernel_signature_match(F.adjust_sharpness, 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.RandomAdjustSharpness(sharpness_factor=0.5, p=1), make_input())
 
@@ -6032,13 +6061,27 @@ def test_functional_error(self):
             F.adjust_sharpness(make_image(), sharpness_factor=-1)
 
     @pytest.mark.parametrize("sharpness_factor", [0.1, 0.5, 1.0])
+    @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.adjust_sharpness, transform_cls_to_functional(transforms.RandomAdjustSharpness, p=1)]
     )
-    def test_correctness_image(self, sharpness_factor, fn):
-        image = make_image(dtype=torch.uint8, device="cpu")
+    def test_correctness_image(self, sharpness_factor, make_input, fn):
+        image = make_input(dtype=torch.uint8, device="cpu")
 
         actual = fn(image, sharpness_factor=sharpness_factor)
+
+        if make_input == make_image_cvcuda:
+            actual = F.cvcuda_to_tensor(actual)[0].cpu()
+            image = F.cvcuda_to_tensor(image)[0].cpu()
+
         expected = F.to_image(F.adjust_sharpness(F.to_pil_image(image), sharpness_factor=sharpness_factor))
 
         assert_equal(actual, expected)

torchvision/transforms/v2/_color.py

@@ -5,11 +5,15 @@
 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_available, _is_cvcuda_tensor
 
 from ._transform import _RandomApplyTransform
 from ._utils import query_chw
 
 
+CVCUDA_AVAILABLE = _is_cvcuda_available()
+
+
 class Grayscale(Transform):
     """Convert images or videos to grayscale.
 
@@ -369,6 +373,9 @@ class RandomAdjustSharpness(_RandomApplyTransform):
 
     _v1_transform_cls = _transforms.RandomAdjustSharpness
 
+    if CVCUDA_AVAILABLE:
+        _transformed_types = _RandomApplyTransform._transformed_types + (_is_cvcuda_tensor,)
+
     def __init__(self, sharpness_factor: float, p: float = 0.5) -> None:
         super().__init__(p=p)
         self.sharpness_factor = sharpness_factor

torchvision/transforms/v2/_utils.py

@@ -16,7 +16,7 @@
 
 from torchvision.transforms.transforms import _check_sequence_input, _setup_angle, _setup_size  # noqa: F401
 from torchvision.transforms.v2.functional import get_dimensions, get_size, is_pure_tensor
-from torchvision.transforms.v2.functional._utils import _FillType, _FillTypeJIT
+from torchvision.transforms.v2.functional._utils import _FillType, _FillTypeJIT, _is_cvcuda_tensor
 
 
 def _setup_number_or_seq(arg: int | float | Sequence[int | float], name: str) -> Sequence[float]:
@@ -182,7 +182,7 @@ def query_chw(flat_inputs: list[Any]) -> tuple[int, int, int]:
     chws = {
         tuple(get_dimensions(inpt))
         for inpt in flat_inputs
-        if check_type(inpt, (is_pure_tensor, tv_tensors.Image, PIL.Image.Image, tv_tensors.Video))
+        if check_type(inpt, (is_pure_tensor, tv_tensors.Image, PIL.Image.Image, tv_tensors.Video, _is_cvcuda_tensor))
     }
     if not chws:
         raise TypeError("No image or video was found in the sample")
@@ -207,6 +207,7 @@ def query_size(flat_inputs: list[Any]) -> tuple[int, int]:
                 tv_tensors.Mask,
                 tv_tensors.BoundingBoxes,
                 tv_tensors.KeyPoints,
+                _is_cvcuda_tensor,
             ),
         )
     }

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 +11,15 @@
 
 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]
+if CVCUDA_AVAILABLE:
+    cvcuda = _import_cvcuda()  # noqa: F811
 
 
 def rgb_to_grayscale(inpt: torch.Tensor, num_output_channels: int = 1) -> torch.Tensor:
@@ -286,6 +296,88 @@ def adjust_sharpness_video(video: torch.Tensor, sharpness_factor: float) -> torc
     return adjust_sharpness_image(video, sharpness_factor=sharpness_factor)
 
 
+_max_value_map: dict["cvcuda.Type", float | int] = {}
+_dtype_to_format: dict[tuple["cvcuda.Type", int], "cvcuda.Format"] = {}
+
+
+def _adjust_sharpness_image_cvcuda(
+    image: "cvcuda.Tensor",
+    sharpness_factor: float,
+) -> "cvcuda.Tensor":
+    cvcuda = _import_cvcuda()
+
+    if len(_max_value_map) == 0:
+        _max_value_map[cvcuda.Type.U8] = 255
+        _max_value_map[cvcuda.Type.F32] = 1.0
+    if len(_dtype_to_format) == 0:
+        _dtype_to_format[(cvcuda.Type.U8, 1)] = cvcuda.Format.U8
+        _dtype_to_format[(cvcuda.Type.U8, 3)] = cvcuda.Format.RGB8
+        _dtype_to_format[(cvcuda.Type.F32, 1)] = cvcuda.Format.F32
+        _dtype_to_format[(cvcuda.Type.F32, 3)] = cvcuda.Format.RGBf32
+
+    if sharpness_factor < 0:
+        raise ValueError(f"sharpness_factor ({sharpness_factor}) is not non-negative.")
+
+    n, h, w, c = image.shape
+    if c not in (1, 3):
+        raise TypeError(f"Input image tensor can have 1 or 3 channels, but found {c}")
+
+    if h <= 2 or w <= 2:
+        return image
+
+    # grab the constants like in the torchvision
+    bound = _max_value_map[image.dtype]
+    fp = image.dtype == cvcuda.Type.F32
+    img_format = _dtype_to_format.get((image.dtype, c))
+    if img_format is None:
+        raise TypeError(f"Unsupported dtype/channel combination: {image.dtype}, {c} channels")
+
+    # conv2d requires ImageBatchVarShape, so we split the batch into individual images
+    # CV-CUDA has no split, so use zero-copy and torch
+    batch = cvcuda.ImageBatchVarShape(capacity=n)
+    for tensor in torch.as_tensor(image.cuda()).split(1, dim=0):
+        cv_image = cvcuda.as_image(tensor, format=img_format)
+        batch.pushback(cv_image)
+
+    # create kernel same as adjust_sharpness_image
+    a, b = 1.0 / 13.0, 5.0 / 13.0
+    torch_kernel = torch.tensor([[a, a, a], [a, b, a], [a, a, a]], dtype=torch.float32, device="cuda")
+    kernel_batch = cvcuda.ImageBatchVarShape(capacity=n)
+    for _ in range(n):
+        kernel_batch.pushback(cvcuda.as_image(torch_kernel, format=cvcuda.Format.F32))
+
+    # anchors of kernel for cvcuda, [-1, -1] means center of kernel
+    anchor_data = torch.tensor([[-1, -1]] * n, dtype=torch.int32, device="cuda")
+    anchor = cvcuda.as_tensor(anchor_data, "NC")
+
+    # run the sharpen operator using cvcuda.conv2d
+    sharpened_batch = cvcuda.conv2d(batch, kernel=kernel_batch, kernel_anchor=anchor, border=cvcuda.Border.REPLICATE)
+    sharpened_list = []
+    for sharpened_img in sharpened_batch:
+        tensor = cvcuda.as_tensor(sharpened_img.cuda(), cvcuda.TensorLayout.HWC)
+        sharpened_list.append(tensor)
+    sharpened = cvcuda.stack(sharpened_list)
+
+    # handle the final blend operations using zero-copy from the adjust_sharpness_image
+    blurred_degenerate = torch.as_tensor(sharpened.cuda())
+    output = torch.as_tensor(image.cuda()).to(dtype=torch.float32, copy=True)
+    if not fp:
+        blurred_degenerate = blurred_degenerate.round()
+    view = output[:, 1:-1, 1:-1, :]
+    blurred_inner = blurred_degenerate[:, 1:-1, 1:-1, :]
+    view.add_(blurred_inner.sub(view), alpha=(1.0 - sharpness_factor))
+    output = output.clamp_(0, bound)
+    if not fp:
+        output = output.to(torch.uint8)
+
+    # convert back to cvcuda.Tensor
+    return cvcuda.as_tensor(output.contiguous(), cvcuda.TensorLayout.NHWC)
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(adjust_sharpness, _import_cvcuda().Tensor)(_adjust_sharpness_image_cvcuda)
+
+
 def adjust_hue(inpt: torch.Tensor, hue_factor: float) -> torch.Tensor:
     """Adjust hue"""
     if torch.jit.is_scripting():

torchvision/transforms/v2/functional/_meta.py

@@ -51,6 +51,16 @@ def get_dimensions_video(video: torch.Tensor) -> list[int]:
     return get_dimensions_image(video)
 
 
+def get_dimensions_image_cvcuda(image: "cvcuda.Tensor") -> list[int]:
+    # CV-CUDA tensor is always in NHWC layout
+    # get_dimensions is CHW
+    return [image.shape[3], image.shape[1], image.shape[2]]
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(get_dimensions, cvcuda.Tensor)(get_dimensions_image_cvcuda)
+
+
 def get_num_channels(inpt: torch.Tensor) -> int:
     if torch.jit.is_scripting():
         return get_num_channels_image(inpt)
@@ -87,6 +97,16 @@ def get_num_channels_video(video: torch.Tensor) -> int:
 get_image_num_channels = get_num_channels
 
 
+def get_num_channels_image_cvcuda(image: "cvcuda.Tensor") -> int:
+    # CV-CUDA tensor is always in NHWC layout
+    # get_num_channels is C
+    return image.shape[3]
+
+
+if CVCUDA_AVAILABLE:
+    _register_kernel_internal(get_num_channels, cvcuda.Tensor)(get_num_channels_image_cvcuda)
+
+
 def get_size(inpt: torch.Tensor) -> list[int]:
     if torch.jit.is_scripting():
         return get_size_image(inpt)
@@ -125,7 +145,7 @@ def get_size_image_cvcuda(image: "cvcuda.Tensor") -> list[int]:
 
 
 if CVCUDA_AVAILABLE:
-    _get_size_image_cvcuda = _register_kernel_internal(get_size, cvcuda.Tensor)(get_size_image_cvcuda)
+    _register_kernel_internal(get_size, _import_cvcuda().Tensor)(get_size_image_cvcuda)
 
 
 @_register_kernel_internal(get_size, tv_tensors.Video, tv_tensor_wrapper=False)