Cycles: Add sample offset option

This patch exposes the sampling offset option to Blender. It is located in the "Sampling > Advanced" panel.
For example, this can be useful to parallelize rendering and distribute different chunks of samples for each computer to render.

---

I also had to add this option to `RenderWork` and `RenderScheduler` classes so that the sample count in the status string can be calculated correctly.

Reviewed By: leesonw

Differential Revision: https://developer.blender.org/D13086

intern/cycles/blender/addon/properties.py

@@ -325,6 +325,13 @@ class CyclesRenderSettings(bpy.types.PropertyGroup):
         default=1024,
     )
 
+    sample_offset: IntProperty(
+        name="Sample Offset",
+        description="Number of samples to skip when starting render",
+        min=0, max=(1 << 24),
+        default=0,
+    )
+
     time_limit: FloatProperty(
         name="Time Limit",
         description="Limit the render time (excluding synchronization time)."

intern/cycles/blender/addon/ui.py

@@ -290,6 +290,9 @@ class CYCLES_RENDER_PT_sampling_advanced(CyclesButtonsPanel, Panel):
         col.active = not (cscene.use_adaptive_sampling and cscene.use_preview_adaptive_sampling)
         col.prop(cscene, "sampling_pattern", text="Pattern")
 
+        col = layout.column(align=True)
+        col.prop(cscene, "sample_offset")
+
         layout.separator()
 
         col = layout.column(align=True)

intern/cycles/blender/sync.cpp

@@ -835,18 +835,25 @@ SessionParams BlenderSync::get_session_params(BL::RenderEngine &b_engine,
   /* samples */
   int samples = get_int(cscene, "samples");
   int preview_samples = get_int(cscene, "preview_samples");
+  int sample_offset = get_int(cscene, "sample_offset");
 
   if (background) {
     params.samples = samples;
+    params.sample_offset = sample_offset;
   }
   else {
     params.samples = preview_samples;
-    if (params.samples == 0)
+    if (params.samples == 0) {
       params.samples = INT_MAX;
+    }
+    params.sample_offset = 0;
   }
 
+  /* Clamp sample offset. */
+  params.sample_offset = clamp(params.sample_offset, 0, Integrator::MAX_SAMPLES);
+
   /* Clamp samples. */
-  params.samples = min(params.samples, Integrator::MAX_SAMPLES);
+  params.samples = clamp(params.samples, 0, Integrator::MAX_SAMPLES - params.sample_offset);
 
   /* Viewport Performance */
   params.pixel_size = b_engine.get_preview_pixel_size(b_scene);

intern/cycles/integrator/path_trace.cpp

@@ -380,7 +380,10 @@ void PathTrace::path_trace(RenderWork &render_work)
     PathTraceWork *path_trace_work = path_trace_works_[i].get();
 
     PathTraceWork::RenderStatistics statistics;
-    path_trace_work->render_samples(statistics, render_work.path_trace.start_sample, num_samples);
+    path_trace_work->render_samples(statistics,
+                                    render_work.path_trace.start_sample,
+                                    num_samples,
+                                    render_work.path_trace.sample_offset);
 
     const double work_time = time_dt() - work_start_time;
     work_balance_infos_[i].time_spent += work_time;
@@ -849,7 +852,8 @@ void PathTrace::progress_update_if_needed(const RenderWork &render_work)
     const int2 tile_size = get_render_tile_size();
     const int num_samples_added = tile_size.x * tile_size.y * render_work.path_trace.num_samples;
     const int current_sample = render_work.path_trace.start_sample +
-                               render_work.path_trace.num_samples;
+                               render_work.path_trace.num_samples -
+                               render_work.path_trace.sample_offset;
     progress_->add_samples(num_samples_added, current_sample);
   }
 

intern/cycles/integrator/path_trace_work.h

@@ -75,7 +75,10 @@ class PathTraceWork {
 
   /* Render given number of samples as a synchronous blocking call.
    * The samples are added to the render buffer associated with this work. */
-  virtual void render_samples(RenderStatistics &statistics, int start_sample, int samples_num) = 0;
+  virtual void render_samples(RenderStatistics &statistics,
+                              int start_sample,
+                              int samples_num,
+                              int sample_offset) = 0;
 
   /* Copy render result from this work to the corresponding place of the GPU display.
    *

intern/cycles/integrator/path_trace_work_cpu.cpp

@@ -71,7 +71,8 @@ void PathTraceWorkCPU::init_execution()
 
 void PathTraceWorkCPU::render_samples(RenderStatistics &statistics,
                                       int start_sample,
-                                      int samples_num)
+                                      int samples_num,
+                                      int sample_offset)
 {
   const int64_t image_width = effective_buffer_params_.width;
   const int64_t image_height = effective_buffer_params_.height;
@@ -97,6 +98,7 @@ void PathTraceWorkCPU::render_samples(RenderStatistics &statistics,
       work_tile.w = 1;
       work_tile.h = 1;
       work_tile.start_sample = start_sample;
+      work_tile.sample_offset = sample_offset;
       work_tile.num_samples = 1;
       work_tile.offset = effective_buffer_params_.offset;
       work_tile.stride = effective_buffer_params_.stride;

intern/cycles/integrator/path_trace_work_cpu.h

@@ -48,7 +48,8 @@ class PathTraceWorkCPU : public PathTraceWork {
 
   virtual void render_samples(RenderStatistics &statistics,
                               int start_sample,
-                              int samples_num) override;
+                              int samples_num,
+                              int sample_offset) override;
 
   virtual void copy_to_display(PathTraceDisplay *display,
                                PassMode pass_mode,

intern/cycles/integrator/path_trace_work_gpu.cpp

@@ -250,7 +250,8 @@ void PathTraceWorkGPU::init_execution()
 
 void PathTraceWorkGPU::render_samples(RenderStatistics &statistics,
                                       int start_sample,
-                                      int samples_num)
+                                      int samples_num,
+                                      int sample_offset)
 {
   /* Limit number of states for the tile and rely on a greedy scheduling of tiles. This allows to
    * add more work (because tiles are smaller, so there is higher chance that more paths will
@@ -261,6 +262,7 @@ void PathTraceWorkGPU::render_samples(RenderStatistics &statistics,
   work_tile_scheduler_.reset(effective_buffer_params_,
                              start_sample,
                              samples_num,
+                             sample_offset,
                              device_scene_->data.integrator.scrambling_distance);
 
   enqueue_reset();

intern/cycles/integrator/path_trace_work_gpu.h

@@ -46,7 +46,8 @@ class PathTraceWorkGPU : public PathTraceWork {
 
   virtual void render_samples(RenderStatistics &statistics,
                               int start_sample,
-                              int samples_num) override;
+                              int samples_num,
+                              int sample_offset) override;
 
   virtual void copy_to_display(PathTraceDisplay *display,
                                PassMode pass_mode,

intern/cycles/integrator/render_scheduler.cpp

@@ -88,6 +88,16 @@ int RenderScheduler::get_num_samples() const
   return num_samples_;
 }
 
+void RenderScheduler::set_sample_offset(int sample_offset)
+{
+  sample_offset_ = sample_offset;
+}
+
+int RenderScheduler::get_sample_offset() const
+{
+  return sample_offset_;
+}
+
 void RenderScheduler::set_time_limit(double time_limit)
 {
   time_limit_ = time_limit;
@@ -110,13 +120,15 @@ int RenderScheduler::get_num_rendered_samples() const
   return state_.num_rendered_samples;
 }
 
-void RenderScheduler::reset(const BufferParams &buffer_params, int num_samples)
+void RenderScheduler::reset(const BufferParams &buffer_params, int num_samples, int sample_offset)
 {
   buffer_params_ = buffer_params;
 
   update_start_resolution_divider();
 
   set_num_samples(num_samples);
+  set_start_sample(sample_offset);
+  set_sample_offset(sample_offset);
 
   /* In background mode never do lower resolution render preview, as it is not really supported
    * by the software. */
@@ -171,7 +183,7 @@ void RenderScheduler::reset(const BufferParams &buffer_params, int num_samples)
 
 void RenderScheduler::reset_for_next_tile()
 {
-  reset(buffer_params_, num_samples_);
+  reset(buffer_params_, num_samples_, sample_offset_);
 }
 
 bool RenderScheduler::render_work_reschedule_on_converge(RenderWork &render_work)
@@ -317,6 +329,7 @@ RenderWork RenderScheduler::get_render_work()
 
   render_work.path_trace.start_sample = get_start_sample_to_path_trace();
   render_work.path_trace.num_samples = get_num_samples_to_path_trace();
+  render_work.path_trace.sample_offset = get_sample_offset();
 
   render_work.init_render_buffers = (render_work.path_trace.start_sample == get_start_sample());
 

intern/cycles/integrator/render_scheduler.h

@@ -39,6 +39,7 @@ class RenderWork {
   struct {
     int start_sample = 0;
     int num_samples = 0;
+    int sample_offset = 0;
   } path_trace;
 
   struct {
@@ -125,6 +126,9 @@ class RenderScheduler {
   void set_num_samples(int num_samples);
   int get_num_samples() const;
 
+  void set_sample_offset(int sample_offset);
+  int get_sample_offset() const;
+
   /* Time limit for the path tracing tasks, in minutes.
    * Zero disables the limit. */
   void set_time_limit(double time_limit);
@@ -150,7 +154,7 @@ class RenderScheduler {
 
   /* Reset scheduler, indicating that rendering will happen from scratch.
    * Resets current rendered state, as well as scheduling information. */
-  void reset(const BufferParams &buffer_params, int num_samples);
+  void reset(const BufferParams &buffer_params, int num_samples, int sample_offset);
 
   /* Reset scheduler upon switching to a next tile.
    * Will keep the same number of samples and full-frame render parameters, but will reset progress
@@ -419,6 +423,8 @@ class RenderScheduler {
   int start_sample_ = 0;
   int num_samples_ = 0;
 
+  int sample_offset_ = 0;
+
   /* Limit in seconds for how long path tracing is allowed to happen.
    * Zero means no limit is applied. */
   double time_limit_ = 0.0;

intern/cycles/integrator/work_tile_scheduler.cpp

@@ -36,6 +36,7 @@ void WorkTileScheduler::set_max_num_path_states(int max_num_path_states)
 void WorkTileScheduler::reset(const BufferParams &buffer_params,
                               int sample_start,
                               int samples_num,
+                              int sample_offset,
                               float scrambling_distance)
 {
   /* Image buffer parameters. */
@@ -51,6 +52,7 @@ void WorkTileScheduler::reset(const BufferParams &buffer_params,
   /* Samples parameters. */
   sample_start_ = sample_start;
   samples_num_ = samples_num;
+  sample_offset_ = sample_offset;
 
   /* Initialize new scheduling. */
   reset_scheduler_state();
@@ -111,6 +113,7 @@ bool WorkTileScheduler::get_work(KernelWorkTile *work_tile_, const int max_work_
   work_tile.h = tile_size_.height;
   work_tile.start_sample = sample_start_ + start_sample;
   work_tile.num_samples = min(tile_size_.num_samples, samples_num_ - start_sample);
+  work_tile.sample_offset = sample_offset_;
   work_tile.offset = offset_;
   work_tile.stride = stride_;
 

intern/cycles/integrator/work_tile_scheduler.h

@@ -41,6 +41,7 @@ class WorkTileScheduler {
   void reset(const BufferParams &buffer_params,
              int sample_start,
              int samples_num,
+             int sample_offset,
              float scrambling_distance);
 
   /* Get work for a device.
@@ -79,6 +80,7 @@ class WorkTileScheduler {
    * (splitting into a smaller work tiles). */
   int sample_start_ = 0;
   int samples_num_ = 0;
+  int sample_offset_ = 0;
 
   /* Tile size which be scheduled for rendering. */
   TileSize tile_size_;

intern/cycles/kernel/film/accumulate.h

@@ -141,7 +141,8 @@ ccl_device_forceinline ccl_global float *kernel_accum_pixel_render_buffer(
 ccl_device_inline int kernel_accum_sample(KernelGlobals kg,
                                           ConstIntegratorState state,
                                           ccl_global float *ccl_restrict render_buffer,
-                                          int sample)
+                                          int sample,
+                                          int sample_offset)
 {
   if (kernel_data.film.pass_sample_count == PASS_UNUSED) {
     return sample;
@@ -149,7 +150,8 @@ ccl_device_inline int kernel_accum_sample(KernelGlobals kg,
 
   ccl_global float *buffer = kernel_accum_pixel_render_buffer(kg, state, render_buffer);
 
-  return atomic_fetch_and_add_uint32((uint *)(buffer) + kernel_data.film.pass_sample_count, 1);
+  return atomic_fetch_and_add_uint32((uint *)(buffer) + kernel_data.film.pass_sample_count, 1) +
+         sample_offset;
 }
 
 ccl_device void kernel_accum_adaptive_buffer(KernelGlobals kg,

intern/cycles/kernel/integrator/init_from_bake.h

@@ -65,7 +65,8 @@ ccl_device bool integrator_init_from_bake(KernelGlobals kg,
   }
 
   /* Always count the sample, even if the camera sample will reject the ray. */
-  const int sample = kernel_accum_sample(kg, state, render_buffer, scheduled_sample);
+  const int sample = kernel_accum_sample(
+      kg, state, render_buffer, scheduled_sample, tile->sample_offset);
 
   /* Setup render buffers. */
   const int index = INTEGRATOR_STATE(state, path, render_pixel_index);

intern/cycles/kernel/integrator/init_from_camera.h

@@ -89,7 +89,8 @@ ccl_device bool integrator_init_from_camera(KernelGlobals kg,
    * This logic allows to both count actual number of samples per pixel, and to add samples to this
    * pixel after it was converged and samples were added somewhere else (in which case the
    * `scheduled_sample` will be different from actual number of samples in this pixel). */
-  const int sample = kernel_accum_sample(kg, state, render_buffer, scheduled_sample);
+  const int sample = kernel_accum_sample(
+      kg, state, render_buffer, scheduled_sample, tile->sample_offset);
 
   /* Initialize random number seed for path. */
   const uint rng_hash = path_rng_hash_init(kg, sample, x, y);

intern/cycles/kernel/types.h

@@ -1410,6 +1410,7 @@ typedef struct KernelWorkTile {
 
   uint start_sample;
   uint num_samples;
+  uint sample_offset;
 
   int offset;
   uint stride;

intern/cycles/session/session.cpp

@@ -262,6 +262,7 @@ RenderWork Session::run_update_for_next_iteration()
   }
 
   render_scheduler_.set_num_samples(params.samples);
+  render_scheduler_.set_start_sample(params.sample_offset);
   render_scheduler_.set_time_limit(params.time_limit);
 
   while (have_tiles) {
@@ -397,7 +398,7 @@ void Session::do_delayed_reset()
 
   /* Tile and work scheduling. */
   tile_manager_.reset_scheduling(buffer_params_, get_effective_tile_size());
-  render_scheduler_.reset(buffer_params_, params.samples);
+  render_scheduler_.reset(buffer_params_, params.samples, params.sample_offset);
 
   /* Passes. */
   /* When multiple tiles are used SAMPLE_COUNT pass is used to keep track of possible partial

intern/cycles/session/session.h

@@ -54,6 +54,7 @@ class SessionParams {
 
   bool experimental;
   int samples;
+  int sample_offset;
   int pixel_size;
   int threads;
 
@@ -75,6 +76,7 @@ class SessionParams {
 
     experimental = false;
     samples = 1024;
+    sample_offset = 0;
     pixel_size = 1;
     threads = 0;
     time_limit = 0.0;