[prototype] Speed up adjust_sharpness_image_tensor (#6930)

* Speed up `adjust_sharpness_image_tensor`

* Add a comment

Author: datumbox

torchvision/prototype/transforms/functional/_color.py

@@ -1,4 +1,5 @@
 import torch
+from torch.nn.functional import conv2d
 from torchvision.prototype import features
 from torchvision.transforms import functional_pil as _FP, functional_tensor as _FT
 
@@ -111,6 +112,8 @@ def adjust_sharpness_image_tensor(image: torch.Tensor, sharpness_factor: float)
     if image.numel() == 0 or height <= 2 or width <= 2:
         return image
 
+    bound = _FT._max_value(image.dtype)
+    fp = image.is_floating_point()
     shape = image.shape
 
     if image.ndim > 4:
@@ -119,7 +122,30 @@ def adjust_sharpness_image_tensor(image: torch.Tensor, sharpness_factor: float)
     else:
         needs_unsquash = False
 
-    output = _blend(image, _FT._blurred_degenerate_image(image), sharpness_factor)
+    # The following is a normalized 3x3 kernel with 1s in the edges and a 5 in the middle.
+    kernel_dtype = image.dtype if fp else torch.float32
+    a, b = 1.0 / 13.0, 5.0 / 13.0
+    kernel = torch.tensor([[a, a, a], [a, b, a], [a, a, a]], dtype=kernel_dtype, device=image.device)
+    kernel = kernel.expand(num_channels, 1, 3, 3)
+
+    # We copy and cast at the same time to avoid modifications on the original data
+    output = image.to(dtype=kernel_dtype, copy=True)
+    blurred_degenerate = conv2d(output, kernel, groups=num_channels)
+    if not fp:
+        # it is better to round before cast
+        blurred_degenerate = blurred_degenerate.round_()
+
+    # Create a view on the underlying output while pointing at the same data. We do this to avoid indexing twice.
+    view = output[..., 1:-1, 1:-1]
+
+    # We speed up blending by minimizing flops and doing in-place. The 2 blend options are mathematically equivalent:
+    # x+(1-r)*(y-x) = x + (1-r)*y - (1-r)*x = x*r + y*(1-r)
+    view.add_(blurred_degenerate.sub_(view), alpha=(1.0 - sharpness_factor))
+
+    # The actual data of ouput have been modified by the above. We only need to clamp and cast now.
+    output = output.clamp_(0, bound)
+    if not fp:
+        output = output.to(image.dtype)
 
     if needs_unsquash:
         output = output.reshape(shape)