From 38249c362d4a7130e45dad35f6fd0d123900eabe Mon Sep 17 00:00:00 2001
From: Wojtek Figat <wojtek@figat.pl>
Date: Thu, 29 Jul 2021 12:39:21 +0200
Subject: [PATCH] Optimize memory allocations in Vulkan backend

---
 .../GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp | 59 +++----------------
 .../GraphicsDevice/Vulkan/GPUDeviceVulkan.h   | 19 ++----
 .../GraphicsDevice/Vulkan/QueueVulkan.cpp     |  2 +-
 3 files changed, 12 insertions(+), 68 deletions(-)

diff --git a/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp b/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp
index 27c2b4dc8..5cc1440d4 100644
--- a/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp
+++ b/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.cpp
@@ -882,22 +882,6 @@ void HelperResourcesVulkan::Dispose()
     }
 }
 
-StagingManagerVulkan::PendingItems* StagingManagerVulkan::PendingItemsPerCmdBuffer::FindOrAdd(uint64 fence)
-{
-    for (int32 i = 0; i < Items.Count(); i++)
-    {
-        if (Items[i].FenceCounter == fence)
-        {
-            return &Items[i];
-        }
-    }
-
-    const auto item = &Items.AddOne();
-    item->FenceCounter = fence;
-    item->Resources.Clear();
-    return item;
-}
-
 StagingManagerVulkan::StagingManagerVulkan(GPUDeviceVulkan* device)
     : _device(device)
 {
@@ -951,9 +935,10 @@ void StagingManagerVulkan::ReleaseBuffer(CmdBufferVulkan* cmdBuffer, GPUBuffer*&
     if (cmdBuffer)
     {
         // Return to pending pool (need to wait until command buffer will be executed and buffer will be reusable)
-        auto items = FindOrAdd(cmdBuffer);
-        auto item = items->FindOrAdd(cmdBuffer->GetFenceSignaledCounter());
-        item->Resources.Add(buffer);
+        auto& item = _pendingBuffers.AddOne();
+        item.Buffer = buffer;
+        item.CmdBuffer = cmdBuffer;
+        item.FenceCounter = cmdBuffer->GetFenceSignaledCounter();
     }
     else
     {
@@ -973,24 +958,10 @@ void StagingManagerVulkan::ProcessPendingFree()
     for (int32 i = _pendingBuffers.Count() - 1; i >= 0; i--)
     {
         auto& e = _pendingBuffers[i];
-
-        for (int32 fenceIndex = e.Items.Count() - 1; fenceIndex >= 0; fenceIndex--)
-        {
-            auto& items = e.Items[fenceIndex];
-            if (items.FenceCounter < e.CmdBuffer->GetFenceSignaledCounter())
-            {
-                for (auto buffer : items.Resources)
-                {
-                    // Return to pool
-                    _freeBuffers.Add({ buffer, Engine::FrameCount });
-                }
-
-                e.Items.RemoveAt(fenceIndex);
-            }
-        }
-
-        if (e.Items.IsEmpty())
+        if (e.FenceCounter < e.CmdBuffer->GetFenceSignaledCounter())
         {
+            // Return to pool
+            _freeBuffers.Add({ e.Buffer, Engine::FrameCount });
             _pendingBuffers.RemoveAt(i);
         }
     }
@@ -1039,22 +1010,6 @@ void StagingManagerVulkan::Dispose()
     _pendingBuffers.Resize(0);
 }
 
-StagingManagerVulkan::PendingItemsPerCmdBuffer* StagingManagerVulkan::FindOrAdd(CmdBufferVulkan* cmdBuffer)
-{
-    for (int32 i = 0; i < _pendingBuffers.Count(); i++)
-    {
-        if (_pendingBuffers[i].CmdBuffer == cmdBuffer)
-        {
-            return &_pendingBuffers[i];
-        }
-    }
-
-    const auto item = &_pendingBuffers.AddOne();
-    item->CmdBuffer = cmdBuffer;
-    item->Items.Clear();
-    return item;
-}
-
 GPUDeviceVulkan::GPUDeviceVulkan(ShaderProfile shaderProfile, GPUAdapterVulkan* adapter)
     : GPUDevice(RendererType::Vulkan, shaderProfile)
     , _renderPasses(512)
diff --git a/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.h b/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.h
index 32f64491e..ea0d4edff 100644
--- a/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.h
+++ b/Source/Engine/GraphicsDevice/Vulkan/GPUDeviceVulkan.h
@@ -413,18 +413,11 @@ class StagingManagerVulkan
 {
 private:
 
-    struct PendingItems
-    {
-        uint64 FenceCounter;
-        Array<GPUBuffer*> Resources;
-    };
-
-    struct PendingItemsPerCmdBuffer
+    struct PendingEntry
     {
+        GPUBuffer* Buffer;
         CmdBufferVulkan* CmdBuffer;
-        Array<PendingItems> Items;
-
-        inline PendingItems* FindOrAdd(uint64 fence);
+        uint64 FenceCounter;
     };
 
     struct FreeEntry
@@ -437,7 +430,7 @@ private:
     CriticalSection _locker;
     Array<GPUBuffer*> _allBuffers;
     Array<FreeEntry> _freeBuffers;
-    Array<PendingItemsPerCmdBuffer> _pendingBuffers;
+    Array<PendingEntry> _pendingBuffers;
 #if !BUILD_RELEASE
     uint64 _allBuffersTotalSize = 0;
     uint64 _allBuffersPeekSize = 0;
@@ -452,10 +445,6 @@ public:
     void ReleaseBuffer(CmdBufferVulkan* cmdBuffer, GPUBuffer*& buffer);
     void ProcessPendingFree();
     void Dispose();
-
-private:
-
-    PendingItemsPerCmdBuffer* FindOrAdd(CmdBufferVulkan* cmdBuffer);
 };
 
 /// <summary>
diff --git a/Source/Engine/GraphicsDevice/Vulkan/QueueVulkan.cpp b/Source/Engine/GraphicsDevice/Vulkan/QueueVulkan.cpp
index aef60c55f..9673f9adf 100644
--- a/Source/Engine/GraphicsDevice/Vulkan/QueueVulkan.cpp
+++ b/Source/Engine/GraphicsDevice/Vulkan/QueueVulkan.cpp
@@ -35,7 +35,7 @@ void QueueVulkan::Submit(CmdBufferVulkan* cmdBuffer, uint32 numSignalSemaphores,
     submitInfo.signalSemaphoreCount = numSignalSemaphores;
     submitInfo.pSignalSemaphores = signalSemaphores;
 
-    Array<VkSemaphore> waitSemaphores;
+    Array<VkSemaphore, InlinedAllocation<8>> waitSemaphores;
     if (cmdBuffer->_waitSemaphores.HasItems())
     {
         waitSemaphores.EnsureCapacity((uint32)cmdBuffer->_waitSemaphores.Count());