Cycles: Rework tile scheduling for denoising

This fixes denoising being delayed until after all rendering has finished. Instead, tile-based
denoising is now part of the "RENDER" task again, so that it is all in one task and does not
cause issues with dedicated task pools where tasks are serialized.

Reviewed By: brecht

Differential Revision: https://developer.blender.org/D6940

12 files changed, 124 insertions, 116 deletions

diff --git a/intern/cycles/device/cuda/device_cuda_impl.cpp b/intern/cycles/device/cuda/device_cuda_impl.cpp
index f6a4f93a690..4a7c45d8b93 100644
--- a/intern/cycles/device/cuda/device_cuda_impl.cpp
+++ b/intern/cycles/device/cuda/device_cuda_impl.cpp
@@ -2144,7 +2144,7 @@ void CUDADevice::thread_run(DeviceTask *task)
 {
   CUDAContextScope scope(this);
 
-  if (task->type == DeviceTask::RENDER || task->type == DeviceTask::DENOISE) {
+  if (task->type == DeviceTask::RENDER) {
     DeviceRequestedFeatures requested_features;
     if (use_split_kernel()) {
       if (split_kernel == NULL) {
@@ -2159,7 +2159,7 @@ void CUDADevice::thread_run(DeviceTask *task)
     RenderTile tile;
     DenoisingTask denoising(this, *task);
 
-    while (task->acquire_tile(this, tile)) {
+    while (task->acquire_tile(this, tile, task->tile_types)) {
       if (tile.task == RenderTile::PATH_TRACE) {
         if (use_split_kernel()) {
           device_only_memory<uchar> void_buffer(this, "void_buffer");
diff --git a/intern/cycles/device/device_cpu.cpp b/intern/cycles/device/device_cpu.cpp
index 795781ee072..1c9d2227ac3 100644
--- a/intern/cycles/device/device_cpu.cpp
+++ b/intern/cycles/device/device_cpu.cpp
@@ -511,7 +511,7 @@ class CPUDevice : public Device {
 
   void thread_run(DeviceTask *task)
   {
-    if (task->type == DeviceTask::RENDER || task->type == DeviceTask::DENOISE)
+    if (task->type == DeviceTask::RENDER)
       thread_render(*task);
     else if (task->type == DeviceTask::SHADER)
       thread_shader(*task);
@@ -927,7 +927,7 @@ class CPUDevice : public Device {
     DenoisingTask denoising(this, task);
     denoising.profiler = &kg->profiler;
 
-    while (task.acquire_tile(this, tile)) {
+    while (task.acquire_tile(this, tile, task.tile_types)) {
       if (tile.task == RenderTile::PATH_TRACE) {
         if (use_split_kernel) {
           device_only_memory<uchar> void_buffer(this, "void_buffer");
diff --git a/intern/cycles/device/device_multi.cpp b/intern/cycles/device/device_multi.cpp
index 8226221ea08..0044610eeb4 100644
--- a/intern/cycles/device/device_multi.cpp
+++ b/intern/cycles/device/device_multi.cpp
@@ -482,11 +482,24 @@ class MultiDevice : public Device {
 
   void task_add(DeviceTask &task)
   {
-    list<SubDevice> &task_devices = denoising_devices.empty() ||
-                                            (task.type != DeviceTask::DENOISE &&
-                                             task.type != DeviceTask::DENOISE_BUFFER) ?
-                                        devices :
-                                        denoising_devices;
+    list<SubDevice> task_devices = devices;
+    if (!denoising_devices.empty()) {
+      if (task.type == DeviceTask::DENOISE_BUFFER) {
+        /* Denoising tasks should be redirected to the denoising devices entirely. */
+        task_devices = denoising_devices;
+      }
+      else if (task.type == DeviceTask::RENDER && (task.tile_types & RenderTile::DENOISE)) {
+        const uint tile_types = task.tile_types;
+        /* For normal rendering tasks only redirect the denoising part to the denoising devices.
+         * Do not need to split the task here, since they all run through 'acquire_tile'. */
+        task.tile_types = RenderTile::DENOISE;
+        foreach (SubDevice &sub, denoising_devices) {
+          sub.device->task_add(task);
+        }
+        /* Rendering itself should still be executed on the rendering devices. */
+        task.tile_types = tile_types ^ RenderTile::DENOISE;
+      }
+    }
 
     list<DeviceTask> tasks;
     task.split(tasks, task_devices.size());
diff --git a/intern/cycles/device/device_optix.cpp b/intern/cycles/device/device_optix.cpp
index 0121b89e9d8..61a5c74f69e 100644
--- a/intern/cycles/device/device_optix.cpp
+++ b/intern/cycles/device/device_optix.cpp
@@ -569,9 +569,14 @@ class OptiXDevice : public CUDADevice {
     if (have_error())
       return;  // Abort early if there was an error previously
 
-    if (task.type == DeviceTask::RENDER || task.type == DeviceTask::DENOISE) {
+    if (task.type == DeviceTask::RENDER) {
+      if (thread_index != 0) {
+        // Only execute denoising in a single thread (see also 'task_add')
+        task.tile_types &= ~RenderTile::DENOISE;
+      }
+
       RenderTile tile;
-      while (task.acquire_tile(this, tile)) {
+      while (task.acquire_tile(this, tile, task.tile_types)) {
         if (tile.task == RenderTile::PATH_TRACE)
           launch_render(task, tile, thread_index);
         else if (tile.task == RenderTile::DENOISE)
@@ -1451,7 +1456,7 @@ class OptiXDevice : public CUDADevice {
       return;
     }
 
-    if (task.type == DeviceTask::DENOISE || task.type == DeviceTask::DENOISE_BUFFER) {
+    if (task.type == DeviceTask::DENOISE_BUFFER) {
       // Execute denoising in a single thread (e.g. to avoid race conditions during creation)
       task_pool.push(new OptiXDeviceTask(this, task, 0));
       return;
diff --git a/intern/cycles/device/device_task.cpp b/intern/cycles/device/device_task.cpp
index 8f15e8c8c1e..36522b874ab 100644
--- a/intern/cycles/device/device_task.cpp
+++ b/intern/cycles/device/device_task.cpp
@@ -68,7 +68,7 @@ int DeviceTask::get_subtask_count(int num, int max_size)
   if (type == SHADER) {
     num = min(shader_w, num);
   }
-  else if (type == RENDER || type == DENOISE) {
+  else if (type == RENDER) {
   }
   else {
     num = min(h, num);
@@ -94,7 +94,7 @@ void DeviceTask::split(list<DeviceTask> &tasks, int num, int max_size)
       tasks.push_back(task);
     }
   }
-  else if (type == RENDER || type == DENOISE) {
+  else if (type == RENDER) {
     for (int i = 0; i < num; i++)
       tasks.push_back(*this);
   }
diff --git a/intern/cycles/device/device_task.h b/intern/cycles/device/device_task.h
index 0f718528b86..972f6131092 100644
--- a/intern/cycles/device/device_task.h
+++ b/intern/cycles/device/device_task.h
@@ -64,7 +64,7 @@ class DenoiseParams {
 
 class DeviceTask : public Task {
  public:
-  typedef enum { RENDER, DENOISE, DENOISE_BUFFER, FILM_CONVERT, SHADER } Type;
+  typedef enum { RENDER, FILM_CONVERT, SHADER, DENOISE_BUFFER } Type;
   Type type;
 
   int x, y, w, h;
@@ -90,7 +90,7 @@ class DeviceTask : public Task {
 
   void update_progress(RenderTile *rtile, int pixel_samples = -1);
 
-  function<bool(Device *device, RenderTile &)> acquire_tile;
+  function<bool(Device *device, RenderTile &, uint)> acquire_tile;
   function<void(long, int)> update_progress_sample;
   function<void(RenderTile &)> update_tile_sample;
   function<void(RenderTile &)> release_tile;
@@ -98,6 +98,7 @@ class DeviceTask : public Task {
   function<void(RenderTile *, Device *)> map_neighbor_tiles;
   function<void(RenderTile *, Device *)> unmap_neighbor_tiles;
 
+  uint tile_types;
   DenoiseParams denoising;
   bool denoising_from_render;
   vector<int> denoising_frames;
diff --git a/intern/cycles/device/opencl/device_opencl_impl.cpp b/intern/cycles/device/opencl/device_opencl_impl.cpp
index 012f6dbe114..68cdfd5238c 100644
--- a/intern/cycles/device/opencl/device_opencl_impl.cpp
+++ b/intern/cycles/device/opencl/device_opencl_impl.cpp
@@ -1308,7 +1308,7 @@ void OpenCLDevice::thread_run(DeviceTask *task)
 {
   flush_texture_buffers();
 
-  if (task->type == DeviceTask::RENDER || task->type == DeviceTask::DENOISE) {
+  if (task->type == DeviceTask::RENDER) {
     RenderTile tile;
     DenoisingTask denoising(this, *task);
 
@@ -1317,7 +1317,7 @@ void OpenCLDevice::thread_run(DeviceTask *task)
     kgbuffer.alloc_to_device(1);
 
     /* Keep rendering tiles until done. */
-    while (task->acquire_tile(this, tile)) {
+    while (task->acquire_tile(this, tile, task->tile_types)) {
       if (tile.task == RenderTile::PATH_TRACE) {
         assert(tile.task == RenderTile::PATH_TRACE);
         scoped_timer timer(&tile.buffers->render_time);
diff --git a/intern/cycles/render/buffers.h b/intern/cycles/render/buffers.h
index 1042b42810f..42efb031843 100644
--- a/intern/cycles/render/buffers.h
+++ b/intern/cycles/render/buffers.h
@@ -130,7 +130,7 @@ class DisplayBuffer {
 
 class RenderTile {
  public:
-  typedef enum { PATH_TRACE, DENOISE } Task;
+  typedef enum { PATH_TRACE = (1 << 0), DENOISE = (1 << 1) } Task;
 
   Task task;
   int x, y, w, h;
diff --git a/intern/cycles/render/session.cpp b/intern/cycles/render/session.cpp
index f45e6d68c97..0d1f8df3610 100644
--- a/intern/cycles/render/session.cpp
+++ b/intern/cycles/render/session.cpp
@@ -301,12 +301,7 @@ void Session::run_gpu()
       update_status_time();
 
       /* render */
-      render();
-
-      /* denoise */
-      if (need_denoise) {
-        denoise();
-      }
+      render(need_denoise);
 
       device->task_wait();
 
@@ -384,7 +379,7 @@ bool Session::draw_cpu(BufferParams &buffer_params, DeviceDrawParams &draw_param
   return false;
 }
 
-bool Session::acquire_tile(Device *tile_device, RenderTile &rtile, RenderTile::Task task)
+bool Session::acquire_tile(RenderTile &rtile, Device *tile_device, uint tile_types)
 {
   if (progress.get_cancel()) {
     if (params.progressive_refine == false) {
@@ -399,9 +394,9 @@ bool Session::acquire_tile(Device *tile_device, RenderTile &rtile, RenderTile::T
   Tile *tile;
   int device_num = device->device_number(tile_device);
 
-  while (!tile_manager.next_tile(tile, device_num, task == RenderTile::DENOISE)) {
+  while (!tile_manager.next_tile(tile, device_num, tile_types)) {
     /* Wait for denoising tiles to become available */
-    if (task == RenderTile::DENOISE && !progress.get_cancel() && tile_manager.has_tiles()) {
+    if ((tile_types & RenderTile::DENOISE) && !progress.get_cancel() && tile_manager.has_tiles()) {
       denoising_cond.wait(tile_lock);
       continue;
     }
@@ -417,7 +412,7 @@ bool Session::acquire_tile(Device *tile_device, RenderTile &rtile, RenderTile::T
   rtile.num_samples = tile_manager.state.num_samples;
   rtile.resolution = tile_manager.state.resolution_divider;
   rtile.tile_index = tile->index;
-  rtile.task = task;
+  rtile.task = tile->state == Tile::DENOISE ? RenderTile::DENOISE : RenderTile::PATH_TRACE;
 
   tile_lock.unlock();
 
@@ -700,12 +695,7 @@ void Session::run_cpu()
       update_status_time();
 
       /* render */
-      render();
-
-      /* denoise */
-      if (need_denoise) {
-        denoise();
-      }
+      render(need_denoise);
 
       /* update status and timing */
       update_status_time();
@@ -1089,100 +1079,91 @@ void Session::update_status_time(bool show_pause, bool show_done)
   progress.set_status(status, substatus);
 }
 
-void Session::render()
+void Session::render(bool with_denoising)
 {
-  /* Clear buffers. */
   if (buffers && tile_manager.state.sample == tile_manager.range_start_sample) {
+    /* Clear buffers. */
     buffers->zero();
   }
 
+  if (tile_manager.state.buffer.width == 0 || tile_manager.state.buffer.height == 0) {
+    return; /* Avoid empty launches. */
+  }
+
   /* Add path trace task. */
   DeviceTask task(DeviceTask::RENDER);
 
-  task.acquire_tile = function_bind(&Session::acquire_tile, this, _1, _2, RenderTile::PATH_TRACE);
+  task.acquire_tile = function_bind(&Session::acquire_tile, this, _2, _1, _3);
   task.release_tile = function_bind(&Session::release_tile, this, _1);
+  task.map_neighbor_tiles = function_bind(&Session::map_neighbor_tiles, this, _1, _2);
+  task.unmap_neighbor_tiles = function_bind(&Session::unmap_neighbor_tiles, this, _1, _2);
   task.get_cancel = function_bind(&Progress::get_cancel, &this->progress);
   task.update_tile_sample = function_bind(&Session::update_tile_sample, this, _1);
   task.update_progress_sample = function_bind(&Progress::add_samples, &this->progress, _1, _2);
   task.need_finish_queue = params.progressive_refine;
   task.integrator_branched = scene->integrator->method == Integrator::BRANCHED_PATH;
 
-  device->task_add(task);
-}
+  /* Acquire render tiles by default. */
+  task.tile_types = RenderTile::PATH_TRACE;
 
-void Session::denoise()
-{
-  if (!params.run_denoising) {
-    return;
-  }
+  with_denoising = params.run_denoising && with_denoising;
+  if (with_denoising) {
+    /* Do not denoise viewport until the sample at which denoising should start is reached. */
+    if (!params.background && tile_manager.state.sample < params.denoising_start_sample) {
+      with_denoising = false;
+    }
 
-  /* Do not denoise viewport until the sample at which denoising should start is reached. */
-  if (!params.background && tile_manager.state.sample < params.denoising_start_sample) {
-    return;
-  }
+    /* Cannot denoise with resolution divider and separate denoising devices.
+     * It breaks the copy in 'MultiDevice::map_neighbor_tiles' (which operates on the full buffer
+     * dimensions and not the scaled ones). */
+    if (!params.device.denoising_devices.empty() && tile_manager.state.resolution_divider > 1) {
+      with_denoising = false;
+    }
 
-  /* Cannot denoise with resolution divider and separate denoising devices.
-   * It breaks the copy in 'MultiDevice::map_neighbor_tiles' (which operates on the full buffer
-   * dimensions and not the scaled ones). */
-  if (!params.device.denoising_devices.empty() && tile_manager.state.resolution_divider > 1) {
-    return;
+    /* It can happen that denoising was already enabled, but the scene still needs an update. */
+    if (scene->film->need_update || !scene->film->denoising_data_offset) {
+      with_denoising = false;
+    }
   }
 
-  /* It can happen that denoising was already enabled, but the scene still needs an update. */
-  if (scene->film->need_update || !scene->film->denoising_data_offset) {
-    return;
-  }
+  if (with_denoising) {
+    task.denoising = params.denoising;
 
-  /* Add separate denoising task. */
-  DeviceTask task(DeviceTask::DENOISE);
+    task.pass_stride = scene->film->pass_stride;
+    task.target_pass_stride = task.pass_stride;
+    task.pass_denoising_data = scene->film->denoising_data_offset;
+    task.pass_denoising_clean = scene->film->denoising_clean_offset;
 
-  if (tile_manager.schedule_denoising) {
-    /* Run denoising on each tile. */
-    task.acquire_tile = function_bind(&Session::acquire_tile, this, _1, _2, RenderTile::DENOISE);
-    task.release_tile = function_bind(&Session::release_tile, this, _1);
-    task.update_tile_sample = function_bind(&Session::update_tile_sample, this, _1);
-    task.update_progress_sample = function_bind(&Progress::add_samples, &this->progress, _1, _2);
-  }
-  else {
-    assert(buffers);
+    task.denoising_from_render = true;
+    task.denoising_do_filter = params.full_denoising;
+    task.denoising_use_optix = params.optix_denoising;
+    task.denoising_write_passes = params.write_denoising_passes;
 
-    if (tile_manager.state.buffer.width == 0 || tile_manager.state.buffer.height == 0) {
-      return; /* Avoid empty launches. */
+    if (tile_manager.schedule_denoising) {
+      /* Acquire denoising tiles during rendering. */
+      task.tile_types |= RenderTile::DENOISE;
     }
+    else {
+      assert(buffers);
 
-    /* Wait for rendering to finish. */
-    device->task_wait();
+      /* Schedule rendering and wait for it to finish. */
+      device->task_add(task);
+      device->task_wait();
 
-    /* Run denoising on the whole image at once. */
-    task.type = DeviceTask::DENOISE_BUFFER;
-    task.x = tile_manager.state.buffer.full_x;
-    task.y = tile_manager.state.buffer.full_y;
-    task.w = tile_manager.state.buffer.width;
-    task.h = tile_manager.state.buffer.height;
-    task.buffer = buffers->buffer.device_pointer;
-    task.sample = tile_manager.state.sample;
-    task.num_samples = tile_manager.state.num_samples;
-    tile_manager.state.buffer.get_offset_stride(task.offset, task.stride);
-    task.buffers = buffers;
+      /* Then run denoising on the whole image at once. */
+      task.type = DeviceTask::DENOISE_BUFFER;
+      task.x = tile_manager.state.buffer.full_x;
+      task.y = tile_manager.state.buffer.full_y;
+      task.w = tile_manager.state.buffer.width;
+      task.h = tile_manager.state.buffer.height;
+      task.buffer = buffers->buffer.device_pointer;
+      task.sample = tile_manager.state.sample;
+      task.num_samples = tile_manager.state.num_samples;
+      tile_manager.state.buffer.get_offset_stride(task.offset, task.stride);
+      task.buffers = buffers;
+    }
   }
 
-  task.get_cancel = function_bind(&Progress::get_cancel, &this->progress);
-  task.need_finish_queue = params.progressive_refine;
-  task.map_neighbor_tiles = function_bind(&Session::map_neighbor_tiles, this, _1, _2);
-  task.unmap_neighbor_tiles = function_bind(&Session::unmap_neighbor_tiles, this, _1, _2);
-
-  task.denoising = params.denoising;
-
-  task.pass_stride = scene->film->pass_stride;
-  task.target_pass_stride = task.pass_stride;
-  task.pass_denoising_data = scene->film->denoising_data_offset;
-  task.pass_denoising_clean = scene->film->denoising_clean_offset;
-
-  task.denoising_from_render = true;
-  task.denoising_do_filter = params.full_denoising;
-  task.denoising_use_optix = params.optix_denoising;
-  task.denoising_write_passes = params.write_denoising_passes;
-
   device->task_add(task);
 }
 
diff --git a/intern/cycles/render/session.h b/intern/cycles/render/session.h
index 40ec3979afd..7f3614ccb19 100644
--- a/intern/cycles/render/session.h
+++ b/intern/cycles/render/session.h
@@ -183,8 +183,7 @@ class Session {
 
   void update_status_time(bool show_pause = false, bool show_done = false);
 
-  void render();
-  void denoise();
+  void render(bool with_denoising);
   void copy_to_display_buffer(int sample);
 
   void reset_(BufferParams &params, int samples);
@@ -197,7 +196,7 @@ class Session {
   bool draw_gpu(BufferParams &params, DeviceDrawParams &draw_params);
   void reset_gpu(BufferParams &params, int samples);
 
-  bool acquire_tile(Device *tile_device, RenderTile &tile, RenderTile::Task task);
+  bool acquire_tile(RenderTile &tile, Device *tile_device, uint tile_types);
   void update_tile_sample(RenderTile &tile);
   void release_tile(RenderTile &tile);
 
diff --git a/intern/cycles/render/tile.cpp b/intern/cycles/render/tile.cpp
index 4ddfd56cd01..1480b6d1aab 100644
--- a/intern/cycles/render/tile.cpp
+++ b/intern/cycles/render/tile.cpp
@@ -495,20 +495,20 @@ bool TileManager::finish_tile(int index, bool &delete_tile)
   }
 }
 
-bool TileManager::next_tile(Tile *&tile, int device, bool denoising)
+bool TileManager::next_tile(Tile *&tile, int device, uint tile_types)
 {
   /* Preserve device if requested, unless this is a separate denoising device that just wants to
    * grab any available tile. */
   const bool preserve_device = preserve_tile_device && device < num_devices;
 
-  int tile_index = -1;
-  int logical_device = preserve_device ? device : 0;
+  if (tile_types & RenderTile::DENOISE) {
+    int tile_index = -1;
+    int logical_device = preserve_device ? device : 0;
 
-  if (denoising) {
     while (logical_device < state.denoising_tiles.size()) {
       if (state.denoising_tiles[logical_device].empty()) {
         if (preserve_device) {
-          return false;
+          break;
         }
         else {
           logical_device++;
@@ -520,12 +520,21 @@ bool TileManager::next_tile(Tile *&tile, int device, bool denoising)
       state.denoising_tiles[logical_device].pop_front();
       break;
     }
+
+    if (tile_index >= 0) {
+      tile = &state.tiles[tile_index];
+      return true;
+    }
   }
-  else {
+
+  if (tile_types & RenderTile::PATH_TRACE) {
+    int tile_index = -1;
+    int logical_device = preserve_device ? device : 0;
+
     while (logical_device < state.render_tiles.size()) {
       if (state.render_tiles[logical_device].empty()) {
         if (preserve_device) {
-          return false;
+          break;
         }
         else {
           logical_device++;
@@ -537,11 +546,11 @@ bool TileManager::next_tile(Tile *&tile, int device, bool denoising)
       state.render_tiles[logical_device].pop_front();
       break;
     }
-  }
 
-  if (tile_index >= 0) {
-    tile = &state.tiles[tile_index];
-    return true;
+    if (tile_index >= 0) {
+      tile = &state.tiles[tile_index];
+      return true;
+    }
   }
 
   return false;
diff --git a/intern/cycles/render/tile.h b/intern/cycles/render/tile.h
index 14c693683c4..9fb9c1ca782 100644
--- a/intern/cycles/render/tile.h
+++ b/intern/cycles/render/tile.h
@@ -106,7 +106,7 @@ class TileManager {
   void reset(BufferParams &params, int num_samples);
   void set_samples(int num_samples);
   bool next();
-  bool next_tile(Tile *&tile, int device, bool denoising);
+  bool next_tile(Tile *&tile, int device, uint tile_types);
   bool finish_tile(int index, bool &delete_tile);
   bool done();
   bool has_tiles();