Distance Scrambling for for Cycles X - Sobol version

Cycles:Distance Scrambling for Cycles Sobol Sampler

This option implements micro jittering an is based on the INRIA
research paper [[ https://hal.inria.fr/hal-01325702/document | on micro jittering ]]
and work by Lukas Stockner for implementing the scrambling distance.
It works by controlling the correlation between pixels by either using
a user supplied value or an adaptive algorithm to limit the maximum
deviation of the sample values between pixels.

This is a follow up of https://developer.blender.org/D12316

The PMJ version can be found here: https://developer.blender.org/D12511

Reviewed By: leesonw

Differential Revision: https://developer.blender.org/D12318

intern/cycles/blender/addon/properties.py

@@ -342,6 +342,24 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
         default='PROGRESSIVE_MUTI_JITTER',
     )
 
+    scrambling_distance: FloatProperty(
+        name="Scrambling Distance",
+        default=1.0,
+        min=0.0, max=1.0,
+        description="Lower values give faster rendering with GPU rendering and less noise with all devices at the cost of possible artifacts if set too low",
+    )
+    preview_scrambling_distance: BoolProperty(
+        name="Scrambling Distance viewport",
+        default=False,
+        description="Uses the Scrambling Distance value for the viewport. Faster but may flicker",
+    )
+
+    adaptive_scrambling_distance: BoolProperty(
+        name="Adaptive Scrambling Distance",
+        default=False,
+        description="Uses a formula to adapt the scrambling distance strength based on the sample count",
+    )
+
     use_layer_samples: EnumProperty(
         name="Layer Samples",
         description="How to use per view layer sample settings",

intern/cycles/blender/addon/ui.py

@@ -289,6 +289,13 @@ class CYCLES_RENDER_PT_sampling_advanced(CyclesButtonsPanel, Panel):
         col = layout.column(align=True)
         col.active = not(cscene.use_adaptive_sampling)
         col.prop(cscene, "sampling_pattern", text="Pattern")
+        col = layout.column(align=True)
+        col.active = cscene.sampling_pattern == 'SOBOL' and not cscene.use_adaptive_sampling
+        col.prop(cscene, "scrambling_distance", text="Scrambling Distance Strength")
+        col.prop(cscene, "adaptive_scrambling_distance", text="Adaptive Scrambling Distance")
+        col = layout.column(align=True)
+        col.active = ((cscene.scrambling_distance < 1.0) or cscene.adaptive_scrambling_distance) and cscene.sampling_pattern == 'SOBOL' and not cscene.use_adaptive_sampling
+        col.prop(cscene, "preview_scrambling_distance", text="Viewport Scrambling Distance")
 
         layout.separator()
 

intern/cycles/blender/sync.cpp

@@ -352,6 +352,21 @@ void BlenderSync::sync_integrator(BL::ViewLayer &b_view_layer, bool background)
     integrator->set_adaptive_min_samples(get_int(cscene, "adaptive_min_samples"));
   }
 
+  int samples = get_int(cscene, "samples");
+  float scrambling_distance = get_float(cscene, "scrambling_distance");
+  bool adaptive_scrambling_distance = get_boolean(cscene, "adaptive_scrambling_distance");
+  if (adaptive_scrambling_distance) {
+    scrambling_distance *= 4.0f / sqrtf(samples);
+  }
+
+  /* only use scrambling distance in the viewport if user wants to and disable with AS */
+  bool preview_scrambling_distance = get_boolean(cscene, "preview_scrambling_distance");
+  if ((preview && !preview_scrambling_distance) || sampling_pattern != SAMPLING_PATTERN_SOBOL)
+    scrambling_distance = 1.0f;
+
+  VLOG(1) << "Used Scrambling Distance: " << scrambling_distance;
+  integrator->set_scrambling_distance(scrambling_distance);
+
   if (get_boolean(cscene, "use_fast_gi")) {
     if (preview) {
       integrator->set_ao_bounces(get_int(cscene, "ao_bounces"));

intern/cycles/integrator/path_trace_work_gpu.cpp

@@ -258,7 +258,10 @@ void PathTraceWorkGPU::render_samples(RenderStatistics &statistics,
    * schedules work in halves of available number of paths. */
   work_tile_scheduler_.set_max_num_path_states(max_num_paths_ / 8);
 
-  work_tile_scheduler_.reset(effective_buffer_params_, start_sample, samples_num);
+  work_tile_scheduler_.reset(effective_buffer_params_,
+                             start_sample,
+                             samples_num,
+                             device_scene_->data.integrator.scrambling_distance);
 
   enqueue_reset();
 

intern/cycles/integrator/tile.cpp

@@ -48,7 +48,8 @@ ccl_device_inline uint round_up_to_power_of_two(uint x)
 
 TileSize tile_calculate_best_size(const int2 &image_size,
                                   const int num_samples,
-                                  const int max_num_path_states)
+                                  const int max_num_path_states,
+                                  const float scrambling_distance)
 {
   if (max_num_path_states == 1) {
     /* Simple case: avoid any calculation, which could cause rounding issues. */
@@ -71,17 +72,54 @@ TileSize tile_calculate_best_size(const int2 &image_size,
    *  - Keep values a power of two, for more integer fit into the maximum number of paths. */
 
   TileSize tile_size;
-
-  /* Calculate tile size as if it is the most possible one to fit an entire range of samples.
-   * The idea here is to keep tiles as small as possible, and keep device occupied by scheduling
-   * multiple tiles with the same coordinates rendering different samples. */
   const int num_path_states_per_sample = max_num_path_states / num_samples;
-  if (num_path_states_per_sample != 0) {
+  if (scrambling_distance < 0.9f) {
-    tile_size.width = round_down_to_power_of_two(lround(sqrt(num_path_states_per_sample)));
+    /* Prefer large tiles for scrambling distance. */
-    tile_size.height = tile_size.width;
+    if (image_size.x * image_size.y <= num_path_states_per_sample) {
+      tile_size.width = image_size.x;
+      tile_size.height = image_size.y;
+    }
+    else {
+      /* Pick the option with the biggest tile size */
+      int heightOption = num_path_states_per_sample / image_size.x;
+      int widthOption = num_path_states_per_sample / image_size.y;
+      // Check if these options are possible
+      if ((heightOption > 0) || (widthOption > 0)) {
+        int area1 = image_size.x * heightOption;
+        int area2 = widthOption * image_size.y;
+        /* The option with the biggest pixel area */
+        if (area1 >= area2) {
+          tile_size.width = image_size.x;
+          tile_size.height = heightOption;
+        }
+        else {
+          tile_size.width = widthOption;
+          tile_size.height = image_size.y;
+        }
+      }
+      else {  // Large tiles are not an option so use square tiles
+        if (num_path_states_per_sample != 0) {
+          tile_size.width = round_down_to_power_of_two(lround(sqrt(num_path_states_per_sample)));
+          tile_size.height = tile_size.width;
+        }
+        else {
+          tile_size.width = tile_size.height = 1;
+        }
+      }
+    }
   }
   else {
-    tile_size.width = tile_size.height = 1;
+    /* Calculate tile size as if it is the most possible one to fit an entire range of samples.
+     * The idea here is to keep tiles as small as possible, and keep device occupied by scheduling
+     * multiple tiles with the same coordinates rendering different samples. */
+
+    if (num_path_states_per_sample != 0) {
+      tile_size.width = round_down_to_power_of_two(lround(sqrt(num_path_states_per_sample)));
+      tile_size.height = tile_size.width;
+    }
+    else {
+      tile_size.width = tile_size.height = 1;
+    }
   }
 
   if (num_samples == 1) {
@@ -93,7 +131,7 @@ TileSize tile_calculate_best_size(const int2 &image_size,
     tile_size.num_samples = min(round_up_to_power_of_two(lround(sqrt(num_samples / 2))),
                                 static_cast<uint>(num_samples));
 
-    const int tile_area = tile_size.width / tile_size.height;
+    const int tile_area = tile_size.width * tile_size.height;
     tile_size.num_samples = min(tile_size.num_samples, max_num_path_states / tile_area);
   }
 

intern/cycles/integrator/tile.h

@@ -51,6 +51,7 @@ std::ostream &operator<<(std::ostream &os, const TileSize &tile_size);
  * possible, and have as many threads active for every tile as possible. */
 TileSize tile_calculate_best_size(const int2 &image_size,
                                   const int num_samples,
-                                  const int max_num_path_states);
+                                  const int max_num_path_states,
+                                  const float scrambling_distance);
 
 CCL_NAMESPACE_END

intern/cycles/integrator/work_tile_scheduler.cpp

@@ -33,13 +33,17 @@ void WorkTileScheduler::set_max_num_path_states(int max_num_path_states)
   max_num_path_states_ = max_num_path_states;
 }
 
-void WorkTileScheduler::reset(const BufferParams &buffer_params, int sample_start, int samples_num)
+void WorkTileScheduler::reset(const BufferParams &buffer_params,
+                              int sample_start,
+                              int samples_num,
+                              float scrambling_distance)
 {
   /* Image buffer parameters. */
   image_full_offset_px_.x = buffer_params.full_x;
   image_full_offset_px_.y = buffer_params.full_y;
 
   image_size_px_ = make_int2(buffer_params.width, buffer_params.height);
+  scrambling_distance_ = scrambling_distance;
 
   offset_ = buffer_params.offset;
   stride_ = buffer_params.stride;
@@ -54,7 +58,8 @@ void WorkTileScheduler::reset(const BufferParams &buffer_params, int sample_star
 
 void WorkTileScheduler::reset_scheduler_state()
 {
-  tile_size_ = tile_calculate_best_size(image_size_px_, samples_num_, max_num_path_states_);
+  tile_size_ = tile_calculate_best_size(
+      image_size_px_, samples_num_, max_num_path_states_, scrambling_distance_);
 
   VLOG(3) << "Will schedule tiles of size " << tile_size_;
 

intern/cycles/integrator/work_tile_scheduler.h

@@ -38,7 +38,10 @@ class WorkTileScheduler {
   void set_max_num_path_states(int max_num_path_states);
 
   /* Scheduling will happen for pixels within a big tile denotes by its parameters. */
-  void reset(const BufferParams &buffer_params, int sample_start, int samples_num);
+  void reset(const BufferParams &buffer_params,
+             int sample_start,
+             int samples_num,
+             float scrambling_distance);
 
   /* Get work for a device.
    * Returns true if there is still work to be done and initialize the work tile to all
@@ -68,6 +71,9 @@ class WorkTileScheduler {
    * Will be passed over to the KernelWorkTile. */
   int offset_, stride_;
 
+  /* Scrambling Distance requires adapted tile size */
+  float scrambling_distance_;
+
   /* Start sample of index and number of samples which are to be rendered.
    * The scheduler will cover samples range of [start, start + num] over the entire image
    * (splitting into a smaller work tiles). */

intern/cycles/kernel/sample/pattern.h

@@ -79,7 +79,7 @@ ccl_device_forceinline float path_rng_1D(KernelGlobals kg,
    * See T38710, T50116.
    */
   uint tmp_rng = cmj_hash_simple(dimension, rng_hash);
-  shift = tmp_rng * (1.0f / (float)0xFFFFFFFF);
+  shift = tmp_rng * (kernel_data.integrator.scrambling_distance / (float)0xFFFFFFFF);
 
   return r + shift - floorf(r + shift);
 #endif

intern/cycles/kernel/types.h

@@ -1184,9 +1184,9 @@ typedef struct KernelIntegrator {
   float volume_step_rate;
 
   int has_shadow_catcher;
+  float scrambling_distance;
 
   /* padding */
-  int pad1;
 } KernelIntegrator;
 static_assert_align(KernelIntegrator, 16);
 

intern/cycles/scene/integrator.cpp

@@ -81,6 +81,7 @@ NODE_DEFINE(Integrator)
   sampling_pattern_enum.insert("sobol", SAMPLING_PATTERN_SOBOL);
   sampling_pattern_enum.insert("pmj", SAMPLING_PATTERN_PMJ);
   SOCKET_ENUM(sampling_pattern, "Sampling Pattern", sampling_pattern_enum, SAMPLING_PATTERN_SOBOL);
+  SOCKET_FLOAT(scrambling_distance, "Scrambling Distance", 1.0f);
 
   static NodeEnum denoiser_type_enum;
   denoiser_type_enum.insert("optix", DENOISER_OPTIX);
@@ -192,6 +193,7 @@ void Integrator::device_update(Device *device, DeviceScene *dscene, Scene *scene
                                            sample_clamp_indirect * 3.0f;
 
   kintegrator->sampling_pattern = new_sampling_pattern;
+  kintegrator->scrambling_distance = scrambling_distance;
 
   if (light_sampling_threshold > 0.0f) {
     kintegrator->light_inv_rr_threshold = 1.0f / light_sampling_threshold;

intern/cycles/scene/integrator.h

@@ -76,6 +76,7 @@ class Integrator : public Node {
   NODE_SOCKET_API(float, adaptive_threshold)
 
   NODE_SOCKET_API(SamplingPattern, sampling_pattern)
+  NODE_SOCKET_API(float, scrambling_distance)
 
   NODE_SOCKET_API(bool, use_denoise);
   NODE_SOCKET_API(DenoiserType, denoiser_type);

intern/cycles/test/integrator_tile_test.cpp

@@ -24,23 +24,26 @@ CCL_NAMESPACE_BEGIN
 TEST(tile_calculate_best_size, Basic)
 {
   /* Make sure CPU-like case is handled properly. */
-  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 1, 1), TileSize(1, 1, 1));
+  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 1, 1, 1.0f), TileSize(1, 1, 1));
-  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 100, 1), TileSize(1, 1, 1));
+  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 100, 1, 1.0f), TileSize(1, 1, 1));
 
   /* Enough path states to fit an entire image with all samples. */
-  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 1, 1920 * 1080),
+  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 1, 1920 * 1080, 1.0f),
             TileSize(1920, 1080, 1));
-  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 100, 1920 * 1080 * 100),
+  EXPECT_EQ(tile_calculate_best_size(make_int2(1920, 1080), 100, 1920 * 1080 * 100, 1.0f),
             TileSize(1920, 1080, 100));
 }
 
 TEST(tile_calculate_best_size, Extreme)
 {
-  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 262144, 131072), TileSize(1, 1, 512));
+  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 262144, 131072, 1.0f),
-  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 1048576, 131072), TileSize(1, 1, 1024));
+            TileSize(1, 1, 512));
-  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 10485760, 131072), TileSize(1, 1, 4096));
+  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 1048576, 131072, 1.0f),
+            TileSize(1, 1, 1024));
+  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 10485760, 131072, 1.0f),
+            TileSize(1, 1, 4096));
 
-  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 8192 * 8192 * 2, 1024),
+  EXPECT_EQ(tile_calculate_best_size(make_int2(32, 32), 8192 * 8192 * 2, 1024, 1.0f),
             TileSize(1, 1, 1024));
 }