From e3b98c902bd60a0f96c16d859abc4baa859a4dc6 Mon Sep 17 00:00:00 2001
From: Wojtek Figat <wojtek@figat.pl>
Date: Mon, 30 Aug 2021 20:30:52 +0200
Subject: [PATCH] Optimize CPU particles impl parts

---
 .../CPU/ParticleEmitterGraph.CPU.ParticleModules.cpp      | 6 ++++--
 .../Graph/CPU/ParticleEmitterGraph.CPU.Particles.cpp      | 2 +-
 .../Particles/Graph/CPU/ParticleEmitterGraph.CPU.cpp      | 8 ++++----
 .../Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.h | 3 ++-
 4 files changed, 11 insertions(+), 8 deletions(-)

diff --git a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.ParticleModules.cpp b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.ParticleModules.cpp
index 24a584852..de06e6435 100644
--- a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.ParticleModules.cpp
+++ b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.ParticleModules.cpp
@@ -610,10 +610,11 @@ void ParticleEmitterGraphCPUExecutor::ProcessModule(ParticleEmitterGraphCPUNode*
         ValueType type(GetVariantType(attribute.ValueType));
         if (node->UsePerParticleDataResolve())
         {
+            Value value;
             for (int32 particleIndex = particlesStart; particleIndex < particlesEnd; particleIndex++)
             {
                 context.ParticleIndex = particleIndex;
-                const Value value = GetValue(box, 4).Cast(type);
+                value = GetValue(box, 4).Cast(type);
                 Platform::MemoryCopy(dataPtr, &value.AsPointer, dataSize);
                 dataPtr += stride;
             }
@@ -667,10 +668,11 @@ void ParticleEmitterGraphCPUExecutor::ProcessModule(ParticleEmitterGraphCPUNode*
         ValueType type(GetVariantType(attribute.ValueType));
         if (node->UsePerParticleDataResolve())
         {
+            Value value;
             for (int32 particleIndex = particlesStart; particleIndex < particlesEnd; particleIndex++)
             {
                 context.ParticleIndex = particleIndex;
-                const Value value = GetValue(box, 2).Cast(type);
+                value = GetValue(box, 2).Cast(type);
                 Platform::MemoryCopy(dataPtr, &value.AsPointer, dataSize);
                 dataPtr += stride;
             }
diff --git a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.Particles.cpp b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.Particles.cpp
index 6d9ba3722..ee71f8d09 100644
--- a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.Particles.cpp
+++ b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.Particles.cpp
@@ -419,7 +419,7 @@ void ParticleEmitterGraphCPUExecutor::ProcessGroupFunction(Box* box, Node* node,
         Node* functionCallNode = nullptr;
         ASSERT(context.GraphStack.Count() >= 2);
         Graph* graph;
-        for (int32 i = context.CallStack.Count() - 1; i >= 0; i--)
+        for (int32 i = context.CallStackSize - 1; i >= 0; i--)
         {
             if (context.CallStack[i]->Type == GRAPH_NODE_MAKE_TYPE(14, 300) && context.Functions.TryGet(context.CallStack[i], graph) && context.GraphStack[context.GraphStack.Count() - 1] == graph)
             {
diff --git a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.cpp b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.cpp
index ba22564bd..0c459ec96 100644
--- a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.cpp
+++ b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.cpp
@@ -131,7 +131,7 @@ void ParticleEmitterGraphCPUExecutor::Init(ParticleEmitter* emitter, ParticleEff
     context.DeltaTime = dt;
     context.ParticleIndex = 0;
     context.ViewTask = effect->GetRenderTask();
-    context.CallStack.Clear();
+    context.CallStackSize = 0;
     context.Functions.Clear();
 }
 
@@ -588,7 +588,7 @@ VisjectExecutor::Value ParticleEmitterGraphCPUExecutor::eatBox(Node* caller, Box
 {
     // Check if graph is looped or is too deep
     auto& context = Context.Get();
-    if (context.CallStack.Count() >= PARTICLE_EMITTER_MAX_CALL_STACK)
+    if (context.CallStackSize >= PARTICLE_EMITTER_MAX_CALL_STACK)
     {
         OnError(caller, box, TEXT("Graph is looped or too deep!"));
         return Value::Zero;
@@ -602,7 +602,7 @@ VisjectExecutor::Value ParticleEmitterGraphCPUExecutor::eatBox(Node* caller, Box
 #endif
 
     // Add to the calling stack
-    context.CallStack.Add(caller);
+    context.CallStack[context.CallStackSize++] = caller;
 
     // Call per group custom processing event
     Value value;
@@ -611,7 +611,7 @@ VisjectExecutor::Value ParticleEmitterGraphCPUExecutor::eatBox(Node* caller, Box
     (this->*func)(box, parentNode, value);
 
     // Remove from the calling stack
-    context.CallStack.RemoveLast();
+    context.CallStackSize--;
 
     return value;
 }
diff --git a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.h b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.h
index 34834a2cd..a1e06f850 100644
--- a/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.h
+++ b/Source/Engine/Particles/Graph/CPU/ParticleEmitterGraph.CPU.h
@@ -120,9 +120,10 @@ struct ParticleEmitterGraphCPUContext
     ParticleEmitter* Emitter;
     ParticleEffect* Effect;
     class SceneRenderTask* ViewTask;
-    Array<VisjectExecutor::Node*, FixedAllocation<PARTICLE_EMITTER_MAX_CALL_STACK>> CallStack;
     Array<VisjectExecutor::Graph*, FixedAllocation<32>> GraphStack;
     Dictionary<VisjectExecutor::Node*, VisjectExecutor::Graph*> Functions;
+    int32 CallStackSize = 0;
+    VisjectExecutor::Node* CallStack[PARTICLE_EMITTER_MAX_CALL_STACK];
 };
 
 /// <summary>