Added alignment parameter to VulkanUploadHeap::Allocate()

3 files changed, 39 insertions, 26 deletions

diff --git a/Graphics/GraphicsEngineVulkan/include/VulkanUploadHeap.h b/Graphics/GraphicsEngineVulkan/include/VulkanUploadHeap.h
index 006ccf4a..3e29b87b 100644
--- a/Graphics/GraphicsEngineVulkan/include/VulkanUploadHeap.h
+++ b/Graphics/GraphicsEngineVulkan/include/VulkanUploadHeap.h
@@ -34,21 +34,21 @@ class RenderDeviceVkImpl;
 struct VulkanUploadAllocation
 {
     VulkanUploadAllocation(){}
-    VulkanUploadAllocation(void* _CPUAddress, VkDeviceSize _Size, VkDeviceSize _Offset, VkBuffer _vkBuffer) :
-        CPUAddress(_CPUAddress),
-        Size      (_Size),
-        Offset    (_Offset),
-        vkBuffer  (_vkBuffer)
+    VulkanUploadAllocation(void* _CPUAddress, VkDeviceSize _Size, VkDeviceSize _AlignedOffset, VkBuffer _vkBuffer) :
+        CPUAddress   (_CPUAddress),
+        Size         (_Size),
+        AlignedOffset(_AlignedOffset),
+        vkBuffer     (_vkBuffer)
     {}
     VulkanUploadAllocation             (const VulkanUploadAllocation&) = delete;
     VulkanUploadAllocation& operator = (const VulkanUploadAllocation&) = delete;
     VulkanUploadAllocation             (VulkanUploadAllocation&&) = default;
     VulkanUploadAllocation& operator = (VulkanUploadAllocation&&) = default;
 
-    VkBuffer     vkBuffer   = VK_NULL_HANDLE;	    // Vulkan buffer associated with this memory.
-    void*        CPUAddress = nullptr;
-    VkDeviceSize Size       = 0;
-    VkDeviceSize Offset     = 0;
+    VkBuffer     vkBuffer      = VK_NULL_HANDLE; // Vulkan buffer associated with this memory.
+    void*        CPUAddress    = nullptr;
+    VkDeviceSize Size          = 0;
+    VkDeviceSize AlignedOffset = 0;
 };
 
 class VulkanUploadHeap
@@ -65,7 +65,7 @@ public:
 
     ~VulkanUploadHeap();
 
-    VulkanUploadAllocation Allocate(size_t SizeInBytes);
+    VulkanUploadAllocation Allocate(size_t SizeInBytes, size_t Alignment);
     void DiscardAllocations(uint64_t FenceValue);
 
     size_t GetStaleAllocationsCount()const
diff --git a/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp b/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp
index 647e210c..1a0a3d41 100644
--- a/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/DeviceContextVkImpl.cpp
@@ -1074,9 +1074,9 @@ namespace Diligent
 #endif
 
         VERIFY_EXPR( static_cast<size_t>(NumBytes) == NumBytes );
-        auto TmpSpace = m_UploadHeap.Allocate(static_cast<size_t>(NumBytes));
+        auto TmpSpace = m_UploadHeap.Allocate(static_cast<size_t>(NumBytes), 0);
 	    memcpy(TmpSpace.CPUAddress, pData, static_cast<size_t>(NumBytes));
-        UpdateBufferRegion(pBuffVk, DstOffset, NumBytes, TmpSpace.vkBuffer, TmpSpace.Offset);
+        UpdateBufferRegion(pBuffVk, DstOffset, NumBytes, TmpSpace.vkBuffer, TmpSpace.AlignedOffset);
         // The allocation will stay in the upload heap until the end of the frame at which point all upload
         // pages will be discarded
     }
@@ -1183,10 +1183,11 @@ namespace Diligent
         const auto UpdateRegionDepth  = CopyInfo.Region.MaxZ - CopyInfo.Region.MinZ;
 
         // For UpdateTextureRegion(), use UploadHeap, not dynamic heap
-        auto Allocation = m_UploadHeap.Allocate(CopyInfo.MemorySize);
+        static constexpr const size_t BufferOffsetAlignment = 4; // bufferOffset of VkBufferImageCopy must be a multiple of 4 (18.4)
+        auto Allocation = m_UploadHeap.Allocate(CopyInfo.MemorySize, BufferOffsetAlignment);
         // The allocation will stay in the upload heap until the end of the frame at which point all upload
         // pages will be discarded
-        VERIFY( (Allocation.Offset % 4) == 0, "Allocation offset must be at least 32-bit algined");
+        VERIFY( (Allocation.AlignedOffset % BufferOffsetAlignment) == 0, "Allocation offset must be at least 32-bit algined");
 
 #ifdef _DEBUG
         VERIFY(SrcStride >= CopyInfo.RowSize, "Source data stride (", SrcStride, ") is below the image row size (", CopyInfo.RowSize, ")");
@@ -1210,7 +1211,7 @@ namespace Diligent
             }
         }
         CopyBufferToTexture(Allocation.vkBuffer,
-                            static_cast<Uint32>(Allocation.Offset),
+                            static_cast<Uint32>(Allocation.AlignedOffset),
                             CopyInfo.Region,
                             TextureVk,
                             MipLevel,
diff --git a/Graphics/GraphicsEngineVulkan/src/VulkanUploadHeap.cpp b/Graphics/GraphicsEngineVulkan/src/VulkanUploadHeap.cpp
index 53d35753..aba420e8 100644
--- a/Graphics/GraphicsEngineVulkan/src/VulkanUploadHeap.cpp
+++ b/Graphics/GraphicsEngineVulkan/src/VulkanUploadHeap.cpp
@@ -79,10 +79,14 @@ VulkanUploadHeap::UploadPageInfo VulkanUploadHeap::CreateNewPage(VkDeviceSize Si
     return UploadPageInfo{std::move(MemAllocation), std::move(NewBuffer), CPUAddress};
 }
 
-VulkanUploadAllocation VulkanUploadHeap::Allocate(size_t SizeInBytes)
+VulkanUploadAllocation VulkanUploadHeap::Allocate(size_t SizeInBytes, size_t Alignment)
 {
-    static constexpr const size_t MinimumAlignment = 4;
-    SizeInBytes = (SizeInBytes + MinimumAlignment-1) & ~(MinimumAlignment-1);
+    VERIFY((Alignment & (Alignment-1)) == 0, "Alignment (", Alignment, ") must be power of two");
+    if (Alignment == 0)
+    {
+        static constexpr const size_t MinimumAlignment = 4;
+        Alignment = MinimumAlignment;
+    }
 
     VulkanUploadAllocation Allocation;
     if(SizeInBytes >= m_PageSize/2)
@@ -92,30 +96,38 @@ VulkanUploadAllocation VulkanUploadHeap::Allocate(size_t SizeInBytes)
         Allocation.vkBuffer   = NewPage.Buffer;
         Allocation.CPUAddress = NewPage.CPUAddress;
         Allocation.Size       = SizeInBytes;
-        Allocation.Offset     = 0;
-        m_CurrAllocatedSize   += NewPage.MemAllocation.Size;
+        VERIFY(Alignment < SizeInBytes, "Alignment must be smaller than the page size");
+        Allocation.AlignedOffset = 0;
+        m_CurrAllocatedSize      += NewPage.MemAllocation.Size;
         m_Pages.emplace_back(std::move(NewPage));
     }
     else
     {
-        if(m_CurrPage.AvailableSize < SizeInBytes)
+        auto AlignmentOffset = ((m_CurrPage.CurrOffset + (Alignment-1)) & ~(Alignment-1)) - m_CurrPage.CurrOffset;
+        if(m_CurrPage.AvailableSize < SizeInBytes + AlignmentOffset)
         {
+            // Allocate new page
             auto NewPage = CreateNewPage(m_PageSize);
             m_CurrPage.Reset(NewPage, m_PageSize);
             m_CurrAllocatedSize += NewPage.MemAllocation.Size;
             m_Pages.emplace_back(std::move(NewPage));
+            VERIFY_EXPR((m_CurrPage.CurrOffset & (Alignment-1)) == 0);
+            AlignmentOffset = 0;
         }
 
-        Allocation.vkBuffer   = m_CurrPage.vkBuffer;
-        Allocation.CPUAddress = m_CurrPage.CurrCPUAddress;
-        Allocation.Size       = SizeInBytes;
-        Allocation.Offset     = m_CurrPage.CurrOffset;
+        m_CurrPage.Advance(AlignmentOffset);
+        VERIFY_EXPR((m_CurrPage.CurrOffset & (Alignment-1)) == 0);
+        Allocation.vkBuffer      = m_CurrPage.vkBuffer;
+        Allocation.CPUAddress    = m_CurrPage.CurrCPUAddress;
+        Allocation.Size          = SizeInBytes;
+        Allocation.AlignedOffset = m_CurrPage.CurrOffset;
         m_CurrPage.Advance(SizeInBytes);
     }
-    m_CurrFrameSize += SizeInBytes;
+    m_CurrFrameSize += SizeInBytes; // Count unaligned size
     m_PeakFrameSize     = std::max(m_CurrFrameSize, m_PeakFrameSize);
     m_PeakAllocatedSize = std::max(m_CurrAllocatedSize, m_PeakAllocatedSize);
 
+    VERIFY_EXPR((Allocation.AlignedOffset & (Alignment-1)) == 0);
     return Allocation;
 }