Realtime Compositor: Implement normalize node

This patch implements the normalize node for the realtime compositor.

Differential Revision: https://developer.blender.org/D16279

Reviewed By: Clement Foucault

source/blender/compositor/realtime_compositor/algorithms/COM_algorithm_parallel_reduction.hh

@@ -70,6 +70,17 @@ float sum_luminance_squared_difference(Context &context,
  * coefficients to compute the luminance. */
 float maximum_luminance(Context &context, GPUTexture *texture, float3 luminance_coefficients);
 
+/* Computes the maximum float of all pixels in the given float texture, limited to the given range.
+ * Values outside of the given range are ignored. If non of the pixel values are in the range, the
+ * lower bound of the range is returned. For instance, if the given range is [-10, 10] and the
+ * image contains the values {2, 5, 11}, the maximum will be 5, since 11 is outside of the range.
+ * This is particularly useful for Z Depth normalization, since Z Depth can contain near infinite
+ * values, so enforcing an upper bound is beneficial. */
+float maximum_float_in_range(Context &context,
+                             GPUTexture *texture,
+                             float lower_bound,
+                             float upper_bound);
+
 /* --------------------------------------------------------------------
  * Minimum Reductions.
  */
@@ -78,4 +89,15 @@ float maximum_luminance(Context &context, GPUTexture *texture, float3 luminance_
  * coefficients to compute the luminance. */
 float minimum_luminance(Context &context, GPUTexture *texture, float3 luminance_coefficients);
 
+/* Computes the minimum float of all pixels in the given float texture, limited to the given range.
+ * Values outside of the given range are ignored. If non of the pixel values are in the range, the
+ * upper bound of the range is returned. For instance, if the given range is [-10, 10] and the
+ * image contains the values {-11, 2, 5}, the minimum will be 2, since -11 is outside of the range.
+ * This is particularly useful for Z Depth normalization, since Z Depth can contain near infinite
+ * values, so enforcing a lower bound is beneficial. */
+float minimum_float_in_range(Context &context,
+                             GPUTexture *texture,
+                             float lower_bound,
+                             float upper_bound);
+
 }  // namespace blender::realtime_compositor

source/blender/compositor/realtime_compositor/algorithms/intern/algorithm_parallel_reduction.cc

@@ -249,6 +249,25 @@ float maximum_luminance(Context &context, GPUTexture *texture, float3 luminance_
   return maximum;
 }
 
+float maximum_float_in_range(Context &context,
+                             GPUTexture *texture,
+                             float lower_bound,
+                             float upper_bound)
+{
+  GPUShader *shader = context.shader_manager().get("compositor_maximum_float_in_range");
+  GPU_shader_bind(shader);
+
+  GPU_shader_uniform_1f(shader, "lower_bound", lower_bound);
+  GPU_shader_uniform_1f(shader, "upper_bound", upper_bound);
+
+  float *reduced_value = parallel_reduction_dispatch(context, texture, shader, GPU_R32F);
+  const float maximum = *reduced_value;
+  MEM_freeN(reduced_value);
+  GPU_shader_unbind();
+
+  return maximum;
+}
+
 /* --------------------------------------------------------------------
  * Minimum Reductions.
  */
@@ -268,4 +287,23 @@ float minimum_luminance(Context &context, GPUTexture *texture, float3 luminance_
   return minimum;
 }
 
+float minimum_float_in_range(Context &context,
+                             GPUTexture *texture,
+                             float lower_bound,
+                             float upper_bound)
+{
+  GPUShader *shader = context.shader_manager().get("compositor_minimum_float_in_range");
+  GPU_shader_bind(shader);
+
+  GPU_shader_uniform_1f(shader, "lower_bound", lower_bound);
+  GPU_shader_uniform_1f(shader, "upper_bound", upper_bound);
+
+  float *reduced_value = parallel_reduction_dispatch(context, texture, shader, GPU_R32F);
+  const float minimum = *reduced_value;
+  MEM_freeN(reduced_value);
+  GPU_shader_unbind();
+
+  return minimum;
+}
+
 }  // namespace blender::realtime_compositor

source/blender/gpu/CMakeLists.txt

@@ -349,6 +349,7 @@ set(GLSL_SRC
   shaders/compositor/compositor_morphological_distance_feather.glsl
   shaders/compositor/compositor_morphological_distance_threshold.glsl
   shaders/compositor/compositor_morphological_step.glsl
+  shaders/compositor/compositor_normalize.glsl
   shaders/compositor/compositor_parallel_reduction.glsl
   shaders/compositor/compositor_projector_lens_distortion.glsl
   shaders/compositor/compositor_realize_on_domain.glsl
@@ -633,6 +634,7 @@ set(SRC_SHADER_CREATE_INFOS
   shaders/compositor/infos/compositor_morphological_distance_info.hh
   shaders/compositor/infos/compositor_morphological_distance_threshold_info.hh
   shaders/compositor/infos/compositor_morphological_step_info.hh
+  shaders/compositor/infos/compositor_normalize_info.hh
   shaders/compositor/infos/compositor_parallel_reduction_info.hh
   shaders/compositor/infos/compositor_projector_lens_distortion_info.hh
   shaders/compositor/infos/compositor_realize_on_domain_info.hh

source/blender/gpu/shaders/compositor/compositor_normalize.glsl

@@ -0,0 +1,10 @@
+#pragma BLENDER_REQUIRE(gpu_shader_compositor_texture_utilities.glsl)
+
+void main()
+{
+  ivec2 texel = ivec2(gl_GlobalInvocationID.xy);
+  float value = texture_load(input_tx, texel).x;
+  float normalized_value = (value - minimum) * scale;
+  float clamped_value = clamp(normalized_value, 0.0, 1.0);
+  imageStore(output_img, texel, vec4(clamped_value));
+}

source/blender/gpu/shaders/compositor/infos/compositor_normalize_info.hh

@@ -0,0 +1,12 @@
+/* SPDX-License-Identifier: GPL-2.0-or-later */
+
+#include "gpu_shader_create_info.hh"
+
+GPU_SHADER_CREATE_INFO(compositor_normalize)
+    .local_group_size(16, 16)
+    .push_constant(Type::FLOAT, "minimum")
+    .push_constant(Type::FLOAT, "scale")
+    .sampler(0, ImageType::FLOAT_2D, "input_tx")
+    .image(0, GPU_R16F, Qualifier::WRITE, ImageType::FLOAT_2D, "output_img")
+    .compute_source("compositor_normalize.glsl")
+    .do_static_compilation(true);

source/blender/gpu/shaders/compositor/infos/compositor_parallel_reduction_info.hh

@@ -108,6 +108,18 @@ GPU_SHADER_CREATE_INFO(compositor_maximum_luminance)
     .define("REDUCE(lhs, rhs)", "max(lhs, rhs)")
     .do_static_compilation(true);
 
+GPU_SHADER_CREATE_INFO(compositor_maximum_float_in_range)
+    .additional_info("compositor_parallel_reduction_shared")
+    .image(0, GPU_R32F, Qualifier::WRITE, ImageType::FLOAT_2D, "output_img")
+    .push_constant(Type::FLOAT, "lower_bound")
+    .push_constant(Type::FLOAT, "upper_bound")
+    .define("TYPE", "float")
+    .define("IDENTITY", "vec4(lower_bound)")
+    .define("INITIALIZE(v)", "((v.x <= upper_bound) && (v.x >= lower_bound)) ? v.x : lower_bound")
+    .define("LOAD(value)", "value.x")
+    .define("REDUCE(lhs, rhs)", "((rhs > lhs) && (rhs <= upper_bound)) ? rhs : lhs")
+    .do_static_compilation(true);
+
 /* --------------------------------------------------------------------
  * Minimum Reductions.
  */
@@ -123,3 +135,15 @@ GPU_SHADER_CREATE_INFO(compositor_minimum_luminance)
     .define("LOAD(value)", "value.x")
     .define("REDUCE(lhs, rhs)", "min(lhs, rhs)")
     .do_static_compilation(true);
+
+GPU_SHADER_CREATE_INFO(compositor_minimum_float_in_range)
+    .additional_info("compositor_parallel_reduction_shared")
+    .image(0, GPU_R32F, Qualifier::WRITE, ImageType::FLOAT_2D, "output_img")
+    .push_constant(Type::FLOAT, "lower_bound")
+    .push_constant(Type::FLOAT, "upper_bound")
+    .define("TYPE", "float")
+    .define("IDENTITY", "vec4(upper_bound)")
+    .define("INITIALIZE(v)", "((v.x <= upper_bound) && (v.x >= lower_bound)) ? v.x : upper_bound")
+    .define("LOAD(value)", "value.x")
+    .define("REDUCE(lhs, rhs)", "((rhs < lhs) && (rhs >= lower_bound)) ? rhs : lhs")
+    .do_static_compilation(true);

source/blender/nodes/composite/nodes/node_composite_normalize.cc

@@ -5,7 +5,9 @@
  * \ingroup cmpnodes
  */
 
+#include "COM_algorithm_parallel_reduction.hh"
 #include "COM_node_operation.hh"
+#include "COM_utilities.hh"
 
 #include "node_composite_util.hh"
 
@@ -15,19 +17,60 @@ namespace blender::nodes::node_composite_normalize_cc {
 
 static void cmp_node_normalize_declare(NodeDeclarationBuilder &b)
 {
-  b.add_input<decl::Float>(N_("Value")).default_value(1.0f).min(0.0f).max(1.0f);
+  b.add_input<decl::Float>(N_("Value"))
+      .default_value(1.0f)
+      .min(0.0f)
+      .max(1.0f)
+      .compositor_domain_priority(0);
   b.add_output<decl::Float>(N_("Value"));
 }
 
 using namespace blender::realtime_compositor;
 
 class NormalizeOperation : public NodeOperation {
+ private:
+  /* The normalize operation is specifically designed to normalize Z Depth information. But since Z
+   * Depth can contain near infinite values, normalization is limited to [-range_, range], meaning
+   * that values outside of that range will be ignored when computing the maximum and minimum for
+   * normalization and will eventually be 0 or 1 if they are less than or larger than the range
+   * respectively. */
+  constexpr static float range_ = 10000.0f;
+
  public:
   using NodeOperation::NodeOperation;
 
   void execute() override
   {
-    get_input("Value").pass_through(get_result("Value"));
+    Result &input_image = get_input("Value");
+    Result &output_image = get_result("Value");
+    if (input_image.is_single_value()) {
+      input_image.pass_through(output_image);
+      return;
+    }
+
+    const float maximum = maximum_float_in_range(
+        context(), input_image.texture(), -range_, range_);
+    const float minimum = minimum_float_in_range(
+        context(), input_image.texture(), -range_, range_);
+    const float scale = (maximum != minimum) ? (1.0f / (maximum - minimum)) : 0.0f;
+
+    GPUShader *shader = shader_manager().get("compositor_normalize");
+    GPU_shader_bind(shader);
+
+    GPU_shader_uniform_1f(shader, "minimum", minimum);
+    GPU_shader_uniform_1f(shader, "scale", scale);
+
+    input_image.bind_as_texture(shader, "input_tx");
+
+    const Domain domain = compute_domain();
+    output_image.allocate_texture(domain);
+    output_image.bind_as_image(shader, "output_img");
+
+    compute_dispatch_threads_at_least(shader, domain.size);
+
+    GPU_shader_unbind();
+    output_image.unbind_as_image();
+    input_image.unbind_as_texture();
   }
 };