another round of perf improvements for equalize (#6776)

* perf improvements for equalize

Co-authored-by: lezcano <[email protected]>

* improve reference tests

* add extensive comments and minor fixes to the kernel

* improve comments

Co-authored-by: lezcano <[email protected]>

Author: pmeier

test/prototype_transforms_kernel_infos.py

@@ -13,6 +13,8 @@
 from datasets_utils import combinations_grid
 from prototype_common_utils import (
     ArgsKwargs,
+    get_num_channels,
+    ImageLoader,
     InfoBase,
     make_bounding_box_loaders,
     make_image_loader,
@@ -1359,9 +1361,43 @@ def sample_inputs_equalize_image_tensor():
 
 
 def reference_inputs_equalize_image_tensor():
-    for image_loader in make_image_loaders(
-        extra_dims=[()], color_spaces=(features.ColorSpace.GRAY, features.ColorSpace.RGB), dtypes=[torch.uint8]
+    # We are not using `make_image_loaders` here since that uniformly samples the values over the whole value range.
+    # Since the whole point of this kernel is to transform an arbitrary distribution of values into a uniform one,
+    # the information gain is low if we already provide something really close to the expected value.
+    spatial_size = (256, 256)
+    for fn, color_space in itertools.product(
+        [
+            *[
+                lambda shape, dtype, device, low=low, high=high: torch.randint(
+                    low, high, shape, dtype=dtype, device=device
+                )
+                for low, high in [
+                    (0, 1),
+                    (255, 256),
+                    (0, 64),
+                    (64, 192),
+                    (192, 256),
+                ]
+            ],
+            *[
+                lambda shape, dtype, device, alpha=alpha, beta=beta: torch.distributions.Beta(alpha, beta)
+                .sample(shape)
+                .mul_(255)
+                .round_()
+                .to(dtype=dtype, device=device)
+                for alpha, beta in [
+                    (0.5, 0.5),
+                    (2, 2),
+                    (2, 5),
+                    (5, 2),
+                ]
+            ],
+        ],
+        [features.ColorSpace.GRAY, features.ColorSpace.RGB],
     ):
+        image_loader = ImageLoader(
+            fn, shape=(get_num_channels(color_space), *spatial_size), dtype=torch.uint8, color_space=color_space
+        )
         yield ArgsKwargs(image_loader)
 
 

torchvision/prototype/transforms/functional/_color.py

@@ -228,39 +228,6 @@ def autocontrast(inpt: features.InputTypeJIT) -> features.InputTypeJIT:
         return autocontrast_image_pil(inpt)
 
 
-def _equalize_image_tensor_vec(image: torch.Tensor) -> torch.Tensor:
-    # input image shape should be [N, H, W]
-    shape = image.shape
-    # Compute image histogram:
-    flat_image = image.flatten(start_dim=1).to(torch.long)  # -> [N, H * W]
-    hist = flat_image.new_zeros(shape[0], 256)
-    hist.scatter_add_(dim=1, index=flat_image, src=flat_image.new_ones(1).expand_as(flat_image))
-
-    # Compute image cdf
-    chist = hist.cumsum_(dim=1)
-    # Compute steps, where step per channel is nonzero_hist[:-1].sum() // 255
-    # Trick: nonzero_hist[:-1].sum() == chist[idx - 1], where idx = chist.argmax()
-    idx = chist.argmax(dim=1).sub_(1)
-    # If histogram is degenerate (hist of zero image), index is -1
-    neg_idx_mask = idx < 0
-    idx.clamp_(min=0)
-    step = chist.gather(dim=1, index=idx.unsqueeze(1))
-    step[neg_idx_mask] = 0
-    step.div_(255, rounding_mode="floor")
-
-    # Compute batched Look-up-table:
-    # Necessary to avoid an integer division by zero, which raises
-    clamped_step = step.clamp(min=1)
-    chist.add_(torch.div(step, 2, rounding_mode="floor")).div_(clamped_step, rounding_mode="floor").clamp_(0, 255)
-    lut = chist.to(torch.uint8)  # [N, 256]
-
-    # Pad lut with zeros
-    zeros = lut.new_zeros((1, 1)).expand(shape[0], 1)
-    lut = torch.cat([zeros, lut[:, :-1]], dim=1)
-
-    return torch.where((step == 0).unsqueeze(-1), image, lut.gather(dim=1, index=flat_image).reshape_as(image))
-
-
 def equalize_image_tensor(image: torch.Tensor) -> torch.Tensor:
     if image.dtype != torch.uint8:
         raise TypeError(f"Only torch.uint8 image tensors are supported, but found {image.dtype}")
@@ -272,7 +239,60 @@ def equalize_image_tensor(image: torch.Tensor) -> torch.Tensor:
     if image.numel() == 0:
         return image
 
-    return _equalize_image_tensor_vec(image.reshape(-1, height, width)).reshape(image.shape)
+    batch_shape = image.shape[:-2]
+    flat_image = image.flatten(start_dim=-2).to(torch.long)
+
+    # The algorithm for histogram equalization is mirrored from PIL:
+    # https://github.com/python-pillow/Pillow/blob/eb59cb61d5239ee69cbbf12709a0c6fd7314e6d7/src/PIL/ImageOps.py#L368-L385
+
+    # Although PyTorch has builtin functionality for histograms, it doesn't support batches. Since we deal with uint8
+    # images here and thus the values are already binned, the computation is trivial. The histogram is computed by using
+    # the flattened image as index. For example, a pixel value of 127 in the image corresponds to adding 1 to index 127
+    # in the histogram.
+    hist = flat_image.new_zeros(batch_shape + (256,), dtype=torch.int32)
+    hist.scatter_add_(dim=-1, index=flat_image, src=hist.new_ones(1).expand_as(flat_image))
+    cum_hist = hist.cumsum(dim=-1)
+
+    # The simplest form of lookup-table (LUT) that also achieves histogram equalization is
+    # `lut = cum_hist / flat_image.shape[-1] * 255`
+    # However, PIL uses a more elaborate scheme:
+    # `lut = ((cum_hist + num_non_max_pixels // (2 * 255)) // num_non_max_pixels) * 255`
+
+    # The last non-zero element in the histogram is the first element in the cumulative histogram with the maximum
+    # value. Thus, the "max" in `num_non_max_pixels` does not refer to 255 as the maximum value of uint8 images, but
+    # rather the maximum value in the image, which might be or not be 255.
+    index = cum_hist.argmax(dim=-1)
+    num_non_max_pixels = flat_image.shape[-1] - hist.gather(dim=-1, index=index.unsqueeze_(-1))
+
+    # This is performance optimization that saves us one multiplication later. With this, the LUT computation simplifies
+    # to `lut = (cum_hist + step // 2) // step` and thus saving the final multiplication by 255 while keeping the
+    # division count the same. PIL uses the variable name `step` for this, so we keep that for easier comparison.
+    step = num_non_max_pixels.div_(255, rounding_mode="floor")
+
+    # Although it looks like we could return early if we find `step == 0` like PIL does, that is unfortunately not as
+    # easy due to our support for batched images. We can only return early if `(step == 0).all()` holds. If it doesn't,
+    # we have to go through the computation below anyway. Since `step == 0` is an edge case anyway, it makes no sense to
+    # pay the runtime cost for checking it every time.
+    no_equalization = step.eq(0).unsqueeze_(-1)
+
+    # `lut[k]` is computed with `cum_hist[k-1]` with `lut[0] == (step // 2) // step == 0`. Thus, we perform the
+    # computation only for `lut[1:]` with `cum_hist[:-1]` and add `lut[0] == 0` afterwards.
+    cum_hist = cum_hist[..., :-1]
+    (
+        cum_hist.add_(step // 2)
+        # We need the `clamp_`(min=1) call here to avoid zero division since they fail for integer dtypes. This has no
+        # effect on the returned result of this kernel since images inside the batch with `step == 0` are returned as is
+        # instead of equalized version.
+        .div_(step.clamp_(min=1), rounding_mode="floor")
+        # We need the `clamp_` call here since PILs LUT computation scheme can produce values outside the valid value
+        # range of uint8 images
+        .clamp_(0, 255)
+    )
+    lut = cum_hist.to(torch.uint8)
+    lut = torch.cat([lut.new_zeros(1).expand(batch_shape + (1,)), lut], dim=-1)
+    equalized_image = lut.gather(dim=-1, index=flat_image).view_as(image)
+
+    return torch.where(no_equalization, image, equalized_image)
 
 
 equalize_image_pil = _FP.equalize