Optimize ProbesRenderer to use time-slicing for cubemap faces rendering and filtering

Source/Engine/Renderer/ProbesRenderer.cpp

@@ -106,6 +106,8 @@ private:
     Array<ProbeEntry> _probesToBake;
 
     ProbeEntry _current;
+    int32 _workStep;
+    float _customCullingNear;
 
     AssetReference<Shader> _shader;
     GPUPipelineState* _psFilterFace = nullptr;
@@ -134,6 +136,7 @@ ProbesRendererService ProbesRendererServiceInstance;
 
 TimeSpan ProbesRenderer::UpdateDelay(0, 0, 0, 0, 100);
 TimeSpan ProbesRenderer::ReleaseTimeout(0, 0, 0, 30);
+int32 ProbesRenderer::MaxWorkPerFrame = 1;
 Delegate<Actor*> ProbesRenderer::OnRegisterBake;
 Delegate<Actor*> ProbesRenderer::OnFinishBake;
 
@@ -293,6 +296,7 @@ void ProbesRendererService::Update()
 
         // Clear flag
         _updateFrameNumber = 0;
+        _workStep = 0;
         _current.Type = ProbeEntry::Types::Invalid;
     }
     else if (_current.Type == ProbeEntry::Types::Invalid && timeSinceUpdate > ProbesRenderer::UpdateDelay)
@@ -321,6 +325,7 @@ void ProbesRendererService::Update()
             _probesToBake.RemoveAtKeepOrder(firstValidEntryIndex);
             _task->Enabled = true;
             _updateFrameNumber = 0;
+            _workStep = 0;
             _lastProbeUpdate = timeNow;
         }
         // Check if need to release data
@@ -408,9 +413,11 @@ void ProbesRendererService::OnRender(RenderTask* task, GPUContext* context)
     PROFILE_GPU("Render Probe");
 
     // Init
-    float customCullingNear = -1;
     const int32 probeResolution = _current.GetResolution();
     const PixelFormat probeFormat = _current.GetFormat();
+    if (_workStep == 0)
+    {
+        _customCullingNear = -1;
         if (_current.Type == ProbeEntry::Types::EnvProbe)
         {
             auto envProbe = (EnvironmentProbe*)_current.Actor.Get();
@@ -434,13 +441,12 @@ void ProbesRendererService::OnRender(RenderTask* task, GPUContext* context)
             Vector3 position = skyLight->GetPosition();
             float nearPlane = 10.0f;
             float farPlane = Math::Max(nearPlane + 1000.0f, skyLight->SkyDistanceThreshold * 2.0f);
-        customCullingNear = skyLight->SkyDistanceThreshold;
+            _customCullingNear = skyLight->SkyDistanceThreshold;
 
             // Setup view
             LargeWorlds::UpdateOrigin(_task->View.Origin, position);
             _task->View.SetUpCube(nearPlane, farPlane, position - _task->View.Origin);
         }
-    _task->CameraCut();
 
         // Resize buffers
         bool resizeFailed = _output->Resize(probeResolution, probeResolution, probeFormat);
@@ -449,31 +455,34 @@ void ProbesRendererService::OnRender(RenderTask* task, GPUContext* context)
         resizeFailed |= _task->Resize(probeResolution, probeResolution);
         if (resizeFailed)
             LOG(Error, "Failed to resize probe");
+    }
 
     // Disable actor during baking (it cannot influence own results)
     const bool isActorActive = _current.Actor->GetIsActive();
     _current.Actor->SetIsActive(false);
 
     // Lower quality when rendering probes in-game to gain performance
-    _task->View.MaxShadowsQuality = Engine::IsPlayMode() ? Quality::Low : Quality::Ultra;
+    _task->View.MaxShadowsQuality = Engine::IsPlayMode() || probeResolution <= 128 ? Quality::Low : Quality::Ultra;
 
     // Render scene for all faces
-    for (int32 faceIndex = 0; faceIndex < 6; faceIndex++)
+    int32 workLeft = ProbesRenderer::MaxWorkPerFrame;
+    const int32 lastFace = Math::Min(_workStep + workLeft, 6);
+    for (int32 faceIndex = _workStep; faceIndex < lastFace; faceIndex++)
     {
+        _task->CameraCut();
         _task->View.SetFace(faceIndex);
 
         // Handle custom frustum for the culling (used to skip objects near the camera)
-        if (customCullingNear > 0)
+        if (_customCullingNear > 0)
         {
             Matrix p;
-            Matrix::PerspectiveFov(PI_OVER_2, 1.0f, customCullingNear, _task->View.Far, p);
+            Matrix::PerspectiveFov(PI_OVER_2, 1.0f, _customCullingNear, _task->View.Far, p);
             _task->View.CullingFrustum.SetMatrix(_task->View.View, p);
         }
 
         // Render frame
         Renderer::Render(_task);
         context->ClearState();
-        _task->CameraCut();
 
         // Copy frame to cube face
         {
@@ -483,12 +492,17 @@ void ProbesRendererService::OnRender(RenderTask* task, GPUContext* context)
             context->Draw(_output->View());
             context->ResetRenderTarget();
         }
+
+        // Move to the next face
+        _workStep++;
+        workLeft--;
     }
 
     // Enable actor back
     _current.Actor->SetIsActive(isActorActive);
 
     // Filter all lower mip levels
+    if (workLeft > 0)
     {
         PROFILE_GPU("Filtering");
         Data data;
@@ -520,11 +534,18 @@ void ProbesRendererService::OnRender(RenderTask* task, GPUContext* context)
                 context->Draw(_tmpFace->View(0, mipIndex));
             }
         }
+
+        // End
+        workLeft--;
+        _workStep++;
     }
 
     // Cleanup
     context->ClearState();
 
+    if (_workStep < 7)
+        return; // Continue rendering next frame
+
     // Mark as rendered
     _updateFrameNumber = Engine::FrameCount;
     _task->Enabled = false;

Source/Engine/Renderer/ProbesRenderer.h

@@ -23,6 +23,11 @@ public:
     /// </summary>
     static TimeSpan ReleaseTimeout;
 
+    /// <summary>
+    /// Maximum amount of cubemap faces or filtering passes that can be performed per-frame (in total). Set it to 7 to perform whole cubemap capture within a single frame, lower values spread the work across multiple frames.
+    /// </summary>
+    static int32 MaxWorkPerFrame;
+
     static Delegate<Actor*> OnRegisterBake;
 
     static Delegate<Actor*> OnFinishBake;