Compositor: Add OIDN prefiltering option to Denoise node

It's equivalent to the OpenImageDenoise prefiltering option in Cycles.
See D12043.

Prefilter modes:
- None: No prefiltering, use when guiding passes are noise-free.
- Fast: Denoise image and guiding passes together. Improves quality when
guiding passes are noisy using least amount of extra processing time.
- Accurate: Prefilter noisy guiding passes before denoising image.
Improves quality when guiding passes are noisy using extra
processing time.

Reviewed By: #compositing, jbakker, sergey

Differential Revision: https://developer.blender.org/D12342

7 files changed, 344 insertions, 118 deletions

diff --git a/source/blender/compositor/nodes/COM_DenoiseNode.cc b/source/blender/compositor/nodes/COM_DenoiseNode.cc
index e58a9c7ba9a..cc9328414ef 100644
--- a/source/blender/compositor/nodes/COM_DenoiseNode.cc
+++ b/source/blender/compositor/nodes/COM_DenoiseNode.cc
@@ -31,6 +31,12 @@ DenoiseNode::DenoiseNode(bNode *editorNode) : Node(editorNode)
 void DenoiseNode::convertToOperations(NodeConverter &converter,
                                       const CompositorContext & /*context*/) const
 {
+  if (!COM_is_denoise_supported()) {
+    converter.mapOutputSocket(getOutputSocket(0),
+                              converter.addInputProxy(getInputSocket(0), false));
+    return;
+  }
+
   bNode *node = this->getbNode();
   NodeDenoise *denoise = (NodeDenoise *)node->storage;
 
@@ -39,8 +45,28 @@ void DenoiseNode::convertToOperations(NodeConverter &converter,
   operation->setDenoiseSettings(denoise);
 
   converter.mapInputSocket(getInputSocket(0), operation->getInputSocket(0));
-  converter.mapInputSocket(getInputSocket(1), operation->getInputSocket(1));
-  converter.mapInputSocket(getInputSocket(2), operation->getInputSocket(2));
+  if (denoise && denoise->prefilter == CMP_NODE_DENOISE_PREFILTER_ACCURATE) {
+    {
+      DenoisePrefilterOperation *normal_prefilter = new DenoisePrefilterOperation(
+          DataType::Vector);
+      normal_prefilter->set_image_name("normal");
+      converter.addOperation(normal_prefilter);
+      converter.mapInputSocket(getInputSocket(1), normal_prefilter->getInputSocket(0));
+      converter.addLink(normal_prefilter->getOutputSocket(), operation->getInputSocket(1));
+    }
+    {
+      DenoisePrefilterOperation *albedo_prefilter = new DenoisePrefilterOperation(DataType::Color);
+      albedo_prefilter->set_image_name("albedo");
+      converter.addOperation(albedo_prefilter);
+      converter.mapInputSocket(getInputSocket(2), albedo_prefilter->getInputSocket(0));
+      converter.addLink(albedo_prefilter->getOutputSocket(), operation->getInputSocket(2));
+    }
+  }
+  else {
+    converter.mapInputSocket(getInputSocket(1), operation->getInputSocket(1));
+    converter.mapInputSocket(getInputSocket(2), operation->getInputSocket(2));
+  }
+
   converter.mapOutputSocket(getOutputSocket(0), operation->getOutputSocket(0));
 }
 
diff --git a/source/blender/compositor/operations/COM_DenoiseOperation.cc b/source/blender/compositor/operations/COM_DenoiseOperation.cc
index e7f2d5a740a..0c660e0b723 100644
--- a/source/blender/compositor/operations/COM_DenoiseOperation.cc
+++ b/source/blender/compositor/operations/COM_DenoiseOperation.cc
@@ -28,6 +28,137 @@ static pthread_mutex_t oidn_lock = BLI_MUTEX_INITIALIZER;
 
 namespace blender::compositor {
 
+bool COM_is_denoise_supported()
+{
+#ifdef WITH_OPENIMAGEDENOISE
+  /* Always supported through Accelerate framework BNNS on macOS. */
+#  ifdef __APPLE__
+  return true;
+#  else
+  return BLI_cpu_support_sse41();
+#  endif
+
+#else
+  return false;
+#endif
+}
+
+class DenoiseFilter {
+ private:
+#ifdef WITH_OPENIMAGEDENOISE
+  oidn::DeviceRef device;
+  oidn::FilterRef filter;
+#endif
+  bool initialized_ = false;
+
+ public:
+  ~DenoiseFilter()
+  {
+    BLI_assert(!initialized_);
+  }
+
+#ifdef WITH_OPENIMAGEDENOISE
+  void init_and_lock_denoiser(MemoryBuffer *output)
+  {
+    /* Since it's memory intensive, it's better to run only one instance of OIDN at a time.
+     * OpenImageDenoise is multithreaded internally and should use all available cores
+     * nonetheless. */
+    BLI_mutex_lock(&oidn_lock);
+
+    device = oidn::newDevice();
+    device.commit();
+    filter = device.newFilter("RT");
+    initialized_ = true;
+    set_image("output", output);
+  }
+
+  void deinit_and_unlock_denoiser()
+  {
+    BLI_mutex_unlock(&oidn_lock);
+    initialized_ = false;
+  }
+
+  void set_image(const StringRef name, MemoryBuffer *buffer)
+  {
+    BLI_assert(initialized_);
+    BLI_assert(!buffer->is_a_single_elem());
+    filter.setImage(name.data(),
+                    buffer->getBuffer(),
+                    oidn::Format::Float3,
+                    buffer->getWidth(),
+                    buffer->getHeight(),
+                    0,
+                    buffer->get_elem_bytes_len());
+  }
+
+  template<typename T> void set(const StringRef option_name, T value)
+  {
+    BLI_assert(initialized_);
+    filter.set(option_name.data(), value);
+  }
+
+  void execute()
+  {
+    BLI_assert(initialized_);
+    filter.commit();
+    filter.execute();
+  }
+
+#else
+  void init_and_lock_denoiser(MemoryBuffer *UNUSED(output))
+  {
+  }
+
+  void deinit_and_unlock_denoiser()
+  {
+  }
+
+  void set_image(const StringRef UNUSED(name), MemoryBuffer *UNUSED(buffer))
+  {
+  }
+
+  template<typename T> void set(const StringRef UNUSED(option_name), T UNUSED(value))
+  {
+  }
+
+  void execute()
+  {
+  }
+#endif
+};
+
+DenoiseBaseOperation::DenoiseBaseOperation()
+{
+  flags.is_fullframe_operation = true;
+  output_rendered_ = false;
+}
+
+bool DenoiseBaseOperation::determineDependingAreaOfInterest(rcti * /*input*/,
+                                                            ReadBufferOperation *readOperation,
+                                                            rcti *output)
+{
+  if (isCached()) {
+    return false;
+  }
+
+  rcti newInput;
+  newInput.xmax = this->getWidth();
+  newInput.xmin = 0;
+  newInput.ymax = this->getHeight();
+  newInput.ymin = 0;
+  return NodeOperation::determineDependingAreaOfInterest(&newInput, readOperation, output);
+}
+
+void DenoiseBaseOperation::get_area_of_interest(const int UNUSED(input_idx),
+                                                const rcti &UNUSED(output_area),
+                                                rcti &r_input_area)
+{
+  r_input_area.xmin = 0;
+  r_input_area.xmax = this->getWidth();
+  r_input_area.ymin = 0;
+  r_input_area.ymax = this->getHeight();
+}
+
 DenoiseOperation::DenoiseOperation()
 {
   this->addInputSocket(DataType::Color);
@@ -35,8 +166,6 @@ DenoiseOperation::DenoiseOperation()
   this->addInputSocket(DataType::Color);
   this->addOutputSocket(DataType::Color);
   this->m_settings = nullptr;
-  flags.is_fullframe_operation = true;
-  output_rendered_ = false;
 }
 void DenoiseOperation::initExecution()
 {
@@ -54,6 +183,25 @@ void DenoiseOperation::deinitExecution()
   SingleThreadedOperation::deinitExecution();
 }
 
+static bool are_guiding_passes_noise_free(NodeDenoise *settings)
+{
+  switch (settings->prefilter) {
+    case CMP_NODE_DENOISE_PREFILTER_NONE:
+    case CMP_NODE_DENOISE_PREFILTER_ACCURATE: /* Prefiltered with #DenoisePrefilterOperation. */
+      return true;
+    case CMP_NODE_DENOISE_PREFILTER_FAST:
+    default:
+      return false;
+  }
+}
+
+void DenoiseOperation::hash_output_params()
+{
+  if (m_settings) {
+    hash_params((int)m_settings->hdr, are_guiding_passes_noise_free(m_settings));
+  }
+}
+
 MemoryBuffer *DenoiseOperation::createMemoryBuffer(rcti *rect2)
 {
   MemoryBuffer *tileColor = (MemoryBuffer *)this->m_inputProgramColor->initializeTileData(rect2);
@@ -69,22 +217,6 @@ MemoryBuffer *DenoiseOperation::createMemoryBuffer(rcti *rect2)
   return result;
 }
 
-bool DenoiseOperation::determineDependingAreaOfInterest(rcti * /*input*/,
-                                                        ReadBufferOperation *readOperation,
-                                                        rcti *output)
-{
-  if (isCached()) {
-    return false;
-  }
-
-  rcti newInput;
-  newInput.xmax = this->getWidth();
-  newInput.xmin = 0;
-  newInput.ymax = this->getHeight();
-  newInput.ymin = 0;
-  return NodeOperation::determineDependingAreaOfInterest(&newInput, readOperation, output);
-}
-
 void DenoiseOperation::generateDenoise(MemoryBuffer *output,
                                        MemoryBuffer *input_color,
                                        MemoryBuffer *input_normal,
@@ -96,104 +228,46 @@ void DenoiseOperation::generateDenoise(MemoryBuffer *output,
     return;
   }
 
-#ifdef WITH_OPENIMAGEDENOISE
-  /* Always supported through Accelerate framework BNNS on macOS. */
-#  ifndef __APPLE__
-  if (BLI_cpu_support_sse41())
-#  endif
-  {
-    /* OpenImageDenoise needs full buffers. */
-    MemoryBuffer *buf_color = input_color->is_a_single_elem() ? input_color->inflate() :
-                                                                input_color;
-    MemoryBuffer *buf_normal = input_normal && input_normal->is_a_single_elem() ?
-                                   input_normal->inflate() :
-                                   input_normal;
-    MemoryBuffer *buf_albedo = input_albedo && input_albedo->is_a_single_elem() ?
-                                   input_albedo->inflate() :
-                                   input_albedo;
+  BLI_assert(COM_is_denoise_supported());
+  /* OpenImageDenoise needs full buffers. */
+  MemoryBuffer *buf_color = input_color->is_a_single_elem() ? input_color->inflate() : input_color;
+  MemoryBuffer *buf_normal = input_normal && input_normal->is_a_single_elem() ?
+                                 input_normal->inflate() :
+                                 input_normal;
+  MemoryBuffer *buf_albedo = input_albedo && input_albedo->is_a_single_elem() ?
+                                 input_albedo->inflate() :
+                                 input_albedo;
 
-    /* Since it's memory intensive, it's better to run only one instance of OIDN at a time.
-     * OpenImageDenoise is multithreaded internally and should use all available cores nonetheless.
-     */
-    BLI_mutex_lock(&oidn_lock);
+  DenoiseFilter filter;
+  filter.init_and_lock_denoiser(output);
 
-    oidn::DeviceRef device = oidn::newDevice();
-    device.commit();
+  filter.set_image("color", buf_color);
+  filter.set_image("normal", buf_normal);
+  filter.set_image("albedo", buf_albedo);
 
-    oidn::FilterRef filter = device.newFilter("RT");
-    filter.setImage("color",
-                    buf_color->getBuffer(),
-                    oidn::Format::Float3,
-                    buf_color->getWidth(),
-                    buf_color->getHeight(),
-                    0,
-                    sizeof(float[4]));
-    if (buf_normal && buf_normal->getBuffer()) {
-      filter.setImage("normal",
-                      buf_normal->getBuffer(),
-                      oidn::Format::Float3,
-                      buf_normal->getWidth(),
-                      buf_normal->getHeight(),
-                      0,
-                      sizeof(float[3]));
-    }
-    if (buf_albedo && buf_albedo->getBuffer()) {
-      filter.setImage("albedo",
-                      buf_albedo->getBuffer(),
-                      oidn::Format::Float3,
-                      buf_albedo->getWidth(),
-                      buf_albedo->getHeight(),
-                      0,
-                      sizeof(float[4]));
-    }
-    filter.setImage("output",
-                    output->getBuffer(),
-                    oidn::Format::Float3,
-                    buf_color->getWidth(),
-                    buf_color->getHeight(),
-                    0,
-                    sizeof(float[4]));
-
-    BLI_assert(settings);
-    if (settings) {
-      filter.set("hdr", settings->hdr);
-      filter.set("srgb", false);
-    }
-
-    filter.commit();
-    filter.execute();
-    BLI_mutex_unlock(&oidn_lock);
+  BLI_assert(settings);
+  if (settings) {
+    filter.set("hdr", settings->hdr);
+    filter.set("srgb", false);
+    filter.set("cleanAux", are_guiding_passes_noise_free(settings));
+  }
 
-    /* Copy the alpha channel, OpenImageDenoise currently only supports RGB. */
-    output->copy_from(input_color, input_color->get_rect(), 3, COM_DATA_TYPE_VALUE_CHANNELS, 3);
+  filter.execute();
+  filter.deinit_and_unlock_denoiser();
 
-    /* Delete inflated buffers. */
-    if (input_color->is_a_single_elem()) {
-      delete buf_color;
-    }
-    if (input_normal && input_normal->is_a_single_elem()) {
-      delete buf_normal;
-    }
-    if (input_albedo && input_albedo->is_a_single_elem()) {
-      delete buf_albedo;
-    }
+  /* Copy the alpha channel, OpenImageDenoise currently only supports RGB. */
+  output->copy_from(input_color, input_color->get_rect(), 3, COM_DATA_TYPE_VALUE_CHANNELS, 3);
 
-    return;
+  /* Delete inflated buffers. */
+  if (input_color->is_a_single_elem()) {
+    delete buf_color;
+  }
+  if (input_normal && input_normal->is_a_single_elem()) {
+    delete buf_normal;
+  }
+  if (input_albedo && input_albedo->is_a_single_elem()) {
+    delete buf_albedo;
   }
-#endif
-  /* If built without OIDN or running on an unsupported CPU, just pass through. */
-  UNUSED_VARS(input_albedo, input_normal, settings);
-  output->copy_from(input_color, input_color->get_rect());
-}
-
-void DenoiseOperation::get_area_of_interest(const int UNUSED(input_idx),
-                                            const rcti &UNUSED(output_area),
-                                            rcti &r_input_area)
-{
-  r_input_area.xmin = 0;
-  r_input_area.xmax = this->getWidth();
-  r_input_area.ymin = 0;
-  r_input_area.ymax = this->getHeight();
 }
 
 void DenoiseOperation::update_memory_buffer(MemoryBuffer *output,
@@ -206,4 +280,57 @@ void DenoiseOperation::update_memory_buffer(MemoryBuffer *output,
   }
 }
 
+DenoisePrefilterOperation::DenoisePrefilterOperation(DataType data_type)
+{
+  this->addInputSocket(data_type);
+  this->addOutputSocket(data_type);
+  image_name_ = "";
+}
+
+void DenoisePrefilterOperation::hash_output_params()
+{
+  hash_param(image_name_);
+}
+
+MemoryBuffer *DenoisePrefilterOperation::createMemoryBuffer(rcti *rect2)
+{
+  MemoryBuffer *input = (MemoryBuffer *)this->get_input_operation(0)->initializeTileData(rect2);
+  rcti rect;
+  BLI_rcti_init(&rect, 0, getWidth(), 0, getHeight());
+
+  MemoryBuffer *result = new MemoryBuffer(getOutputSocket()->getDataType(), rect);
+  generate_denoise(result, input);
+
+  return result;
+}
+
+void DenoisePrefilterOperation::generate_denoise(MemoryBuffer *output, MemoryBuffer *input)
+{
+  BLI_assert(COM_is_denoise_supported());
+
+  /* Denoising needs full buffers. */
+  MemoryBuffer *input_buf = input->is_a_single_elem() ? input->inflate() : input;
+
+  DenoiseFilter filter;
+  filter.init_and_lock_denoiser(output);
+  filter.set_image(image_name_, input_buf);
+  filter.execute();
+  filter.deinit_and_unlock_denoiser();
+
+  /* Delete inflated buffers. */
+  if (input->is_a_single_elem()) {
+    delete input_buf;
+  }
+}
+
+void DenoisePrefilterOperation::update_memory_buffer(MemoryBuffer *output,
+                                                     const rcti &UNUSED(area),
+                                                     Span<MemoryBuffer *> inputs)
+{
+  if (!output_rendered_) {
+    this->generate_denoise(output, inputs[0]);
+    output_rendered_ = true;
+  }
+}
+
 }  // namespace blender::compositor
diff --git a/source/blender/compositor/operations/COM_DenoiseOperation.h b/source/blender/compositor/operations/COM_DenoiseOperation.h
index 48209c3eacf..1b053b79c2d 100644
--- a/source/blender/compositor/operations/COM_DenoiseOperation.h
+++ b/source/blender/compositor/operations/COM_DenoiseOperation.h
@@ -23,7 +23,24 @@
 
 namespace blender::compositor {
 
-class DenoiseOperation : public SingleThreadedOperation {
+bool COM_is_denoise_supported();
+
+class DenoiseBaseOperation : public SingleThreadedOperation {
+ protected:
+  bool output_rendered_;
+
+ protected:
+  DenoiseBaseOperation();
+
+ public:
+  bool determineDependingAreaOfInterest(rcti *input,
+                                        ReadBufferOperation *readOperation,
+                                        rcti *output) override;
+
+  void get_area_of_interest(int input_idx, const rcti &output_area, rcti &r_input_area) override;
+};
+
+class DenoiseOperation : public DenoiseBaseOperation {
  private:
   /**
    * \brief Cached reference to the input programs
@@ -37,8 +54,6 @@ class DenoiseOperation : public SingleThreadedOperation {
    */
   NodeDenoise *m_settings;
 
-  bool output_rendered_;
-
  public:
   DenoiseOperation();
   /**
@@ -55,16 +70,13 @@ class DenoiseOperation : public SingleThreadedOperation {
   {
     this->m_settings = settings;
   }
-  bool determineDependingAreaOfInterest(rcti *input,
-                                        ReadBufferOperation *readOperation,
-                                        rcti *output) override;
 
-  void get_area_of_interest(int input_idx, const rcti &output_area, rcti &r_input_area) override;
   void update_memory_buffer(MemoryBuffer *output,
                             const rcti &area,
                             Span<MemoryBuffer *> inputs) override;
 
  protected:
+  void hash_output_params() override;
   void generateDenoise(MemoryBuffer *output,
                        MemoryBuffer *input_color,
                        MemoryBuffer *input_normal,
@@ -74,4 +86,28 @@ class DenoiseOperation : public SingleThreadedOperation {
   MemoryBuffer *createMemoryBuffer(rcti *rect) override;
 };
 
+class DenoisePrefilterOperation : public DenoiseBaseOperation {
+ private:
+  std::string image_name_;
+
+ public:
+  DenoisePrefilterOperation(DataType data_type);
+
+  void set_image_name(StringRef name)
+  {
+    image_name_ = name;
+  }
+
+  void update_memory_buffer(MemoryBuffer *output,
+                            const rcti &area,
+                            Span<MemoryBuffer *> inputs) override;
+
+ protected:
+  void hash_output_params() override;
+  MemoryBuffer *createMemoryBuffer(rcti *rect) override;
+
+ private:
+  void generate_denoise(MemoryBuffer *output, MemoryBuffer *input);
+};
+
 }  // namespace blender::compositor
diff --git a/source/blender/editors/space_node/drawnode.cc b/source/blender/editors/space_node/drawnode.cc
index 4b859de0ac9..62f40152416 100644
--- a/source/blender/editors/space_node/drawnode.cc
+++ b/source/blender/editors/space_node/drawnode.cc
@@ -2865,6 +2865,8 @@ static void node_composit_buts_denoise(uiLayout *layout, bContext *UNUSED(C), Po
 #  endif
 #endif
 
+  uiItemL(layout, IFACE_("Prefilter:"), ICON_NONE);
+  uiItemR(layout, ptr, "prefilter", DEFAULT_FLAGS, nullptr, ICON_NONE);
   uiItemR(layout, ptr, "use_hdr", DEFAULT_FLAGS, nullptr, ICON_NONE);
 }
 
diff --git a/source/blender/makesdna/DNA_node_types.h b/source/blender/makesdna/DNA_node_types.h
index ee1da6be3f2..74ed22ecafd 100644
--- a/source/blender/makesdna/DNA_node_types.h
+++ b/source/blender/makesdna/DNA_node_types.h
@@ -1167,6 +1167,7 @@ typedef struct NodeCryptomatte {
 
 typedef struct NodeDenoise {
   char hdr;
+  char prefilter;
 } NodeDenoise;
 
 typedef struct NodeAttributeClamp {
@@ -1840,6 +1841,14 @@ typedef enum CMPNodeSetAlphaMode {
   CMP_NODE_SETALPHA_MODE_REPLACE_ALPHA = 1,
 } CMPNodeSetAlphaMode;
 
+/* Denoise Node. */
+/* `NodeDenoise.prefilter` */
+typedef enum CMPNodeDenoisePrefilter {
+  CMP_NODE_DENOISE_PREFILTER_FAST = 0,
+  CMP_NODE_DENOISE_PREFILTER_NONE = 1,
+  CMP_NODE_DENOISE_PREFILTER_ACCURATE = 2
+} CMPNodeDenoisePrefilter;
+
 #define CMP_NODE_PLANETRACKDEFORM_MBLUR_SAMPLES_MAX 64
 
 /* Point Density shader node */
diff --git a/source/blender/makesrna/intern/rna_nodetree.c b/source/blender/makesrna/intern/rna_nodetree.c
index 86e134799aa..d685692c8fa 100644
--- a/source/blender/makesrna/intern/rna_nodetree.c
+++ b/source/blender/makesrna/intern/rna_nodetree.c
@@ -8890,12 +8890,37 @@ static void def_cmp_denoise(StructRNA *srna)
 {
   PropertyRNA *prop;
 
+  static const EnumPropertyItem prefilter_items[] = {
+      {CMP_NODE_DENOISE_PREFILTER_NONE,
+       "NONE",
+       0,
+       "None",
+       "No prefiltering, use when guiding passes are noise-free"},
+      {CMP_NODE_DENOISE_PREFILTER_FAST,
+       "FAST",
+       0,
+       "Fast",
+       "Denoise image and guiding passes together. Improves quality when guiding passes are noisy "
+       "using least amount of extra processing time"},
+      {CMP_NODE_DENOISE_PREFILTER_ACCURATE,
+       "ACCURATE",
+       0,
+       "Accurate",
+       "Prefilter noisy guiding passes before denoising image. Improves quality when guiding "
+       "passes are noisy using extra processing time"},
+      {0, NULL, 0, NULL, NULL}};
+
   RNA_def_struct_sdna_from(srna, "NodeDenoise", "storage");
 
   prop = RNA_def_property(srna, "use_hdr", PROP_BOOLEAN, PROP_NONE);
   RNA_def_property_boolean_sdna(prop, NULL, "hdr", 0);
   RNA_def_property_ui_text(prop, "HDR", "Process HDR images");
   RNA_def_property_update(prop, NC_NODE | NA_EDITED, "rna_Node_update");
+
+  prop = RNA_def_property(srna, "prefilter", PROP_ENUM, PROP_NONE);
+  RNA_def_property_enum_items(prop, prefilter_items);
+  RNA_def_property_ui_text(prop, "", "Denoising prefilter");
+  RNA_def_property_update(prop, NC_NODE | NA_EDITED, "rna_Node_update");
 }
 
 static void def_cmp_antialiasing(StructRNA *srna)
diff --git a/source/blender/nodes/composite/nodes/node_composite_denoise.c b/source/blender/nodes/composite/nodes/node_composite_denoise.c
index 040b350627e..e2c7c7b995f 100644
--- a/source/blender/nodes/composite/nodes/node_composite_denoise.c
+++ b/source/blender/nodes/composite/nodes/node_composite_denoise.c
@@ -36,6 +36,7 @@ static void node_composit_init_denonise(bNodeTree *UNUSED(ntree), bNode *node)
 {
   NodeDenoise *ndg = MEM_callocN(sizeof(NodeDenoise), "node denoise data");
   ndg->hdr = true;
+  ndg->prefilter = CMP_NODE_DENOISE_PREFILTER_ACCURATE;
   node->storage = ndg;
 }