EEVEE-Next: Add Temporal-AntiAliasing

The improvements over the old implementation are:
- Improved history reprojection filter (catmull-rom)
- Use proper velocity for history reprojection.
- History clipping is now done in YCoCg color space using better algorithm.
- Velocity is dilated to keep correct edge anti-aliasing on moving objects.

As a result, the 3x3 blocks that made the image smoother in the previous
implementation are no longer visible is replaced by correct antialiasing.

This removes the velocity resolve pass in order to reduce the bandwidth
usage. The velocities are just resolved as they are loadded in the film
pass.

1 files changed, 30 insertions, 14 deletions

diff --git a/source/blender/draw/engines/eevee_next/eevee_film.cc b/source/blender/draw/engines/eevee_next/eevee_film.cc
index 1fd4c278c88..a1becaed9c4 100644
--- a/source/blender/draw/engines/eevee_next/eevee_film.cc
+++ b/source/blender/draw/engines/eevee_next/eevee_film.cc
@@ -163,11 +163,13 @@ inline bool operator!=(const FilmData &a, const FilmData &b)
 
 void Film::init(const int2 &extent, const rcti *output_rect)
 {
+  Sampling &sampling = inst_.sampling;
+
   init_aovs();
 
   {
     /* Enable passes that need to be rendered. */
-    eViewLayerEEVEEPassType render_passes;
+    eViewLayerEEVEEPassType render_passes = eViewLayerEEVEEPassType(0);
 
     if (inst_.is_viewport()) {
       /* Viewport Case. */
@@ -178,6 +180,8 @@ void Film::init(const int2 &extent, const rcti *output_rect)
          * Using the render pass ensure we store the center depth. */
         render_passes |= EEVEE_RENDER_PASS_Z;
       }
+      /* TEST */
+      render_passes |= EEVEE_RENDER_PASS_VECTOR;
     }
     else {
       /* Render Case. */
@@ -211,7 +215,7 @@ void Film::init(const int2 &extent, const rcti *output_rect)
 
     /* TODO(@fclem): Can't we rely on depsgraph update notification? */
     if (assign_if_different(enabled_passes_, render_passes)) {
-      inst_.sampling.reset();
+      sampling.reset();
     }
   }
   {
@@ -224,14 +228,18 @@ void Film::init(const int2 &extent, const rcti *output_rect)
     FilmData data = data_;
     data.extent = int2(BLI_rcti_size_x(output_rect), BLI_rcti_size_y(output_rect));
     data.offset = int2(output_rect->xmin, output_rect->ymin);
-    data.filter_size = clamp_f(inst_.scene->r.gauss, 0.0f, 100.0f);
+    data.extent_inv = 1.0f / float2(data.extent);
+    /* Disable filtering if sample count is 1. */
+    data.filter_size = (sampling.sample_count() == 1) ?
+                           0.0f :
+                           clamp_f(inst_.scene->r.gauss, 0.0f, 100.0f);
     /* TODO(fclem): parameter hidden in experimental.
      * We need to figure out LOD bias first in order to preserve texture crispiness. */
     data.scaling_factor = 1;
 
     FilmData &data_prev_ = data_;
     if (assign_if_different(data_prev_, data)) {
-      inst_.sampling.reset();
+      sampling.reset();
     }
 
     const eViewLayerEEVEEPassType data_passes = EEVEE_RENDER_PASS_Z | EEVEE_RENDER_PASS_NORMAL |
@@ -325,7 +333,7 @@ void Film::init(const int2 &extent, const rcti *output_rect)
                                              (data_.value_len > 0) ? data_.value_len : 1);
 
     if (reset > 0) {
-      inst_.sampling.reset();
+      sampling.reset();
       data_.use_history = 0;
       data_.use_reprojection = 0;
 
@@ -349,12 +357,22 @@ void Film::sync()
   /* TODO(fclem): Shader variation for panoramic & scaled resolution. */
 
   RenderBuffers &rbuffers = inst_.render_buffers;
+  VelocityModule &velocity = inst_.velocity;
+
+  eGPUSamplerState filter = GPU_SAMPLER_FILTER;
+
+  /* For viewport, only previous motion is supported.
+   * Still bind previous step to avoid undefined behavior. */
+  eVelocityStep step_next = inst_.is_viewport() ? STEP_PREVIOUS : STEP_NEXT;
 
   DRWState state = DRW_STATE_WRITE_COLOR | DRW_STATE_WRITE_DEPTH | DRW_STATE_DEPTH_ALWAYS;
   accumulate_ps_ = DRW_pass_create("Film.Accumulate", state);
   GPUShader *sh = inst_.shaders.static_shader_get(shader);
   DRWShadingGroup *grp = DRW_shgroup_create(sh, accumulate_ps_);
   DRW_shgroup_uniform_block_ref(grp, "film_buf", &data_);
+  DRW_shgroup_uniform_block_ref(grp, "camera_prev", &(*velocity.camera_steps[STEP_PREVIOUS]));
+  DRW_shgroup_uniform_block_ref(grp, "camera_curr", &(*velocity.camera_steps[STEP_CURRENT]));
+  DRW_shgroup_uniform_block_ref(grp, "camera_next", &(*velocity.camera_steps[step_next]));
   DRW_shgroup_uniform_texture_ref(grp, "depth_tx", &rbuffers.depth_tx);
   DRW_shgroup_uniform_texture_ref(grp, "combined_tx", &rbuffers.combined_tx);
   DRW_shgroup_uniform_texture_ref(grp, "normal_tx", &rbuffers.normal_tx);
@@ -375,7 +393,7 @@ void Film::sync()
    * use image binding instead. */
   DRW_shgroup_uniform_image_ref(grp, "in_weight_img", &weight_tx_.current());
   DRW_shgroup_uniform_image_ref(grp, "out_weight_img", &weight_tx_.next());
-  DRW_shgroup_uniform_image_ref(grp, "in_combined_img", &combined_tx_.current());
+  DRW_shgroup_uniform_texture_ref_ex(grp, "in_combined_tx", &combined_tx_.current(), filter);
   DRW_shgroup_uniform_image_ref(grp, "out_combined_img", &combined_tx_.next());
   DRW_shgroup_uniform_image_ref(grp, "depth_img", &depth_tx_);
   DRW_shgroup_uniform_image_ref(grp, "color_accum_img", &color_accum_tx_);
@@ -395,13 +413,7 @@ void Film::sync()
 
 void Film::end_sync()
 {
-  if (inst_.sampling.is_reset()) {
-    data_.use_history = 0;
-  }
-
-  // if (camera.changed_type) {
-  //   data_.use_reprojection = false;
-  // }
+  data_.use_reprojection = inst_.sampling.interactive_mode();
 
   aovs_info.push_update();
 
@@ -429,6 +441,11 @@ float2 Film::pixel_jitter_get() const
   return jitter;
 }
 
+eViewLayerEEVEEPassType Film::enabled_passes_get() const
+{
+  return enabled_passes_;
+}
+
 void Film::update_sample_table()
 {
   data_.subpixel_offset = pixel_jitter_get();
@@ -528,7 +545,6 @@ void Film::accumulate(const DRWView *view)
   /* Use history after first sample. */
   if (data_.use_history == 0) {
     data_.use_history = 1;
-    data_.use_reprojection = 1;
   }
 }