Pre warm RT Handle system so that reallocations are reduced.

com.unity.render-pipelines.core/Runtime/RenderGraph/RenderGraphResourceRegistry.cs

@@ -449,7 +449,8 @@ private RenderGraphResourceRegistry()
 
         internal RenderGraphResourceRegistry(bool supportMSAA, MSAASamples initialSampleCount, RenderGraphDebugParams renderGraphDebug, RenderGraphLogger logger)
         {
-            m_RTHandleSystem.Initialize(1, 1, supportMSAA, initialSampleCount);
+            // We initialize to screen width/height to avoid multiple realloc that can lead to inflated memory usage (as releasing of memory is delayed).
+            m_RTHandleSystem.Initialize(Screen.width, Screen.height, supportMSAA, initialSampleCount);
             m_RenderGraphDebug = renderGraphDebug;
             m_Logger = logger;
         }

com.unity.render-pipelines.high-definition/CHANGELOG.md

@@ -775,6 +775,7 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 - Removed experimental namespace for ray tracing code.
 - Increase limit for max numbers of lights in UX
 - Removed direct use of BSDFData in the path tracing pass, delegated to the material instead.
+- Pre-warm the RTHandle system to reduce the amount of memory allocations and the total memory needed at all points. 
 
 ## [7.1.1] - 2019-09-05
 

com.unity.render-pipelines.high-definition/Runtime/RenderPipeline/Camera/HDCamera.cs

@@ -505,11 +505,6 @@ internal void Update(FrameSettings currentFrameSettings, HDRenderPipeline hdrp,
 
             HDRenderPipeline.UpdateVolumetricBufferParams(this, hdrp.GetFrameCount());
             HDRenderPipeline.ResizeVolumetricHistoryBuffers(this, hdrp.GetFrameCount());
-
-            // Here we use the non scaled resolution for the RTHandleSystem ref size because we assume that at some point we will need full resolution anyway.
-            // This is necessary because we assume that after post processes, we have the full size render target for debug rendering
-            // The only point of calling this here is to grow the render targets. The call in BeginRender will setup the current RTHandle viewport size.
-            RTHandles.SetReferenceSize(nonScaledViewport.x, nonScaledViewport.y, msaaSamples);
         }
 
         /// <summary>Set the RTHandle scale to the actual camera size (can be scaled)</summary>

com.unity.render-pipelines.high-definition/Runtime/RenderPipeline/HDRenderPipeline.cs

@@ -422,8 +422,8 @@ public HDRenderPipeline(HDRenderPipelineAsset asset, HDRenderPipelineAsset defau
 
             // Initial state of the RTHandle system.
             // Tells the system that we will require MSAA or not so that we can avoid wasteful render texture allocation.
-            // TODO: Might want to initialize to at least the window resolution to avoid un-necessary re-alloc in the player
-            RTHandles.Initialize(1, 1, m_Asset.currentPlatformRenderPipelineSettings.supportMSAA, m_Asset.currentPlatformRenderPipelineSettings.msaaSampleCount);
+            // We initialize to screen width/height to avoid multiple realloc that can lead to inflated memory usage (as releasing of memory is delayed).
+            RTHandles.Initialize(Screen.width, Screen.height, m_Asset.currentPlatformRenderPipelineSettings.supportMSAA, m_Asset.currentPlatformRenderPipelineSettings.msaaSampleCount);
 
             m_XRSystem = new XRSystem(asset.renderPipelineResources.shaders);
             m_GPUCopy = new GPUCopy(defaultResources.shaders.copyChannelCS);
@@ -2007,6 +2007,29 @@ ref _cullingResults
 
                     using (new ProfilingScope(null, ProfilingSampler.Get(HDProfileId.HDRenderPipelineAllRenderRequest)))
                     {
+
+                        // Warm up the RTHandle system so that it gets init to the maximum resolution available (avoiding to call multiple resizes
+                        // that can lead to high memory spike as the memory release is delayed while the creation is immediate).
+                        {
+                            Vector2Int maxSize = new Vector2Int(1, 1);
+
+                            for (int i = renderRequestIndicesToRender.Count - 1; i >= 0; --i)
+                            {
+                                var renderRequestIndex = renderRequestIndicesToRender[i];
+                                var renderRequest = renderRequests[renderRequestIndex];
+                                var hdCamera = renderRequest.hdCamera;
+
+                                maxSize.x = Math.Max((int)hdCamera.finalViewport.size.x, maxSize.x);
+                                maxSize.y = Math.Max((int)hdCamera.finalViewport.size.y, maxSize.y);
+                            }
+
+                            // Here we use the non scaled resolution for the RTHandleSystem ref size because we assume that at some point we will need full resolution anyway.
+                            // This is necessary because we assume that after post processes, we have the full size render target for debug rendering
+                            // The only point of calling this here is to grow the render targets. The call in BeginRender will setup the current RTHandle viewport size.
+                            RTHandles.SetReferenceSize(maxSize.x, maxSize.y, m_MSAASamples);
+                        }
+
+
                         // Execute render request graph, in reverse order
                         for (int i = renderRequestIndicesToRender.Count - 1; i >= 0; --i)
                         {