Fix #124133: Compositor transforms are off by 0.5 pixels

The GPU compositor transforms are off by half a pixel in some cases.
That's because the realization shader was to perform transforms, and it
includes logic to move images by half a pixel if the domain and image
had different even/odd sizes. To fix this, we only move by half a pixel
if we are doing realization, while transforms are left as is.

source/blender/compositor/realtime_compositor/intern/realize_on_domain_operation.cc

@@ -2,6 +2,8 @@
  *
  * SPDX-License-Identifier: GPL-2.0-or-later */
 
+#include "BLI_math_matrix.hh"
+
 #include "COM_algorithm_realize_on_domain.hh"
 #include "COM_context.hh"
 #include "COM_domain.hh"
@@ -29,11 +31,19 @@ RealizeOnDomainOperation::RealizeOnDomainOperation(Context &context,
 
 void RealizeOnDomainOperation::execute()
 {
+  const Domain &in_domain = get_input().domain();
+
+  /* Even and odd-sized domains have different pixel locations, which produces unexpected
+   * filtering. If one is odd and the other is even (detected by testing the low bit of the xor of
+   * the sizes), shift the input by 1/2 pixel so the pixels align. */
+  const float2 translation(((in_domain.size[0] ^ domain_.size[0]) & 1) ? -0.5f : 0.0f,
+                           ((in_domain.size[1] ^ domain_.size[1]) & 1) ? -0.5f : 0.0f);
+
   realize_on_domain(context(),
                     get_input(),
                     get_result(),
                     domain_,
-                    get_input().domain().transformation,
+                    math::translate(in_domain.transformation, translation),
                     get_input().get_realization_options());
 }
 

source/blender/compositor/realtime_compositor/shaders/compositor_realize_on_domain.glsl

@@ -20,11 +20,10 @@ void main()
   /* Since an input image with an identity transformation is supposed to be centered in the domain,
    * we subtract the offset between the lower left corners of the input image and the domain, which
    * is half the difference between their sizes, because the difference in size is on both sides of
-   * the centered image. Additionally, we floor the offset to retain the 0.5 offset added above in
-   * case the difference in sizes was odd. */
+   * the centered image. */
   ivec2 domain_size = imageSize(domain_img);
   ivec2 input_size = texture_size(input_tx);
-  vec2 offset = floor(vec2(domain_size - input_size) / 2.0);
+  vec2 offset = vec2(domain_size - input_size) / 2.0;
 
   /* Subtract the offset and divide by the input image size to get the relevant coordinates into
    * the sampler's expected [0, 1] range. */