Added alignment parameter to allocation managers (fixed https://github.com/DiligentGraphics/DiligentCore/issues/29)

5 files changed, 18 insertions, 19 deletions

diff --git a/Graphics/GraphicsEngineD3D12/include/DynamicUploadHeap.h b/Graphics/GraphicsEngineD3D12/include/DynamicUploadHeap.h
index 8edc4c57..b4e73676 100644
--- a/Graphics/GraphicsEngineD3D12/include/DynamicUploadHeap.h
+++ b/Graphics/GraphicsEngineD3D12/include/DynamicUploadHeap.h
@@ -28,9 +28,6 @@
 namespace Diligent
 {
 
-// Constant blocks must be multiples of 16 constants @ 16 bytes each
-#define DEFAULT_ALIGN 256
-
 struct DynamicAllocation
 {
     DynamicAllocation()noexcept{}
@@ -86,9 +83,9 @@ public:
 
     ~GPURingBuffer();
 
-    DynamicAllocation Allocate(size_t SizeInBytes)
+    DynamicAllocation Allocate(size_t SizeInBytes, size_t Alignment)
     {
-        auto Offset = RingBuffer::Allocate(SizeInBytes);
+        auto Offset = RingBuffer::Allocate(SizeInBytes, Alignment);
         if (Offset != RingBuffer::InvalidOffset)
         {
             DynamicAllocation DynAlloc(m_pBuffer, Offset, SizeInBytes);
@@ -126,7 +123,7 @@ public:
     DynamicUploadHeap& operator=(const DynamicUploadHeap&)= delete;
     DynamicUploadHeap& operator=(DynamicUploadHeap&&)     = delete;
 
-    DynamicAllocation Allocate( size_t SizeInBytes, size_t Alignment = DEFAULT_ALIGN );
+    DynamicAllocation Allocate( size_t SizeInBytes, size_t Alignment);
 
     void FinishFrame(Uint64 FenceValue, Uint64 LastCompletedFenceValue);
 
diff --git a/Graphics/GraphicsEngineD3D12/src/BufferD3D12Impl.cpp b/Graphics/GraphicsEngineD3D12/src/BufferD3D12Impl.cpp
index 92b734f4..9427c271 100644
--- a/Graphics/GraphicsEngineD3D12/src/BufferD3D12Impl.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/BufferD3D12Impl.cpp
@@ -354,7 +354,8 @@ void BufferD3D12Impl :: Map(IDeviceContext* pContext, MAP_TYPE MapType, Uint32 M
             auto ContextId = pDeviceContextD3D12->GetContextId();
             if ((MapFlags & MAP_FLAG_DISCARD) != 0 || m_DynamicData[ContextId].CPUAddress == nullptr) 
             {
-                m_DynamicData[ContextId] = pCtxD3D12->AllocateDynamicSpace(m_Desc.uiSizeInBytes, 0);
+                size_t Alignment = (m_Desc.BindFlags & BIND_UNIFORM_BUFFER) ? D3D12_CONSTANT_BUFFER_DATA_PLACEMENT_ALIGNMENT : 16;
+                m_DynamicData[ContextId] = pCtxD3D12->AllocateDynamicSpace(m_Desc.uiSizeInBytes, Alignment);
             }
             else
             {
diff --git a/Graphics/GraphicsEngineD3D12/src/DescriptorHeap.cpp b/Graphics/GraphicsEngineD3D12/src/DescriptorHeap.cpp
index 6815b112..5d6b9a00 100644
--- a/Graphics/GraphicsEngineD3D12/src/DescriptorHeap.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/DescriptorHeap.cpp
@@ -95,12 +95,15 @@ DescriptorHeapAllocation DescriptorHeapAllocationManager::Allocate(uint32_t Coun
     // Methods of VariableSizeGPUAllocationsManager class are not thread safe!
 
     // Use variable-size GPU allocations manager to allocate the requested number of descriptors
-    auto DescriptorHandleOffset = m_FreeBlockManager.Allocate(Count);
+    auto Allocation = m_FreeBlockManager.Allocate(Count, 1);
+    auto DescriptorHandleOffset = Allocation.UnalignedOffset;
     if (DescriptorHandleOffset == VariableSizeGPUAllocationsManager::InvalidOffset)
     {
-        return DescriptorHeapAllocation();
+        return DescriptorHeapAllocation{};
     }
 
+    VERIFY_EXPR(Allocation.Size == Count);
+
     // Compute the first CPU and GPU descriptor handles in the allocation by
     // offseting the first CPU and GPU descriptor handle in the range
     auto CPUHandle = m_FirstCPUHandle;
diff --git a/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp b/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
index 1f7ab435..ea3b7a86 100644
--- a/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/DeviceContextD3D12Impl.cpp
@@ -692,7 +692,7 @@ namespace Diligent
    
     DynamicAllocation DeviceContextD3D12Impl::AllocateDynamicSpace(size_t NumBytes, size_t Alignment)
     {
-        return m_pUploadHeap->Allocate(NumBytes + Alignment);
+        return m_pUploadHeap->Allocate(NumBytes, Alignment);
     }
 
     void DeviceContextD3D12Impl::UpdateBufferRegion(class BufferD3D12Impl* pBuffD3D12, DynamicAllocation& Allocation, Uint64 DstOffset, Uint64 NumBytes)
@@ -711,7 +711,8 @@ namespace Diligent
     {
         VERIFY(pBuffD3D12->GetDesc().Usage != USAGE_DYNAMIC, "Dynamic buffers must be updated via Map()");
         VERIFY_EXPR( static_cast<size_t>(NumBytes) == NumBytes );
-        auto TmpSpace = m_pUploadHeap->Allocate(static_cast<size_t>(NumBytes));
+        constexpr size_t DefaultAlginment = 16;
+        auto TmpSpace = m_pUploadHeap->Allocate(static_cast<size_t>(NumBytes), DefaultAlginment);
 	    memcpy(TmpSpace.CPUAddress, pData, static_cast<size_t>(NumBytes));
         UpdateBufferRegion(pBuffD3D12, TmpSpace, DstOffset, NumBytes);
     }
diff --git a/Graphics/GraphicsEngineD3D12/src/DynamicUploadHeap.cpp b/Graphics/GraphicsEngineD3D12/src/DynamicUploadHeap.cpp
index 232c5f8e..f884579c 100644
--- a/Graphics/GraphicsEngineD3D12/src/DynamicUploadHeap.cpp
+++ b/Graphics/GraphicsEngineD3D12/src/DynamicUploadHeap.cpp
@@ -117,7 +117,7 @@ namespace Diligent
         m_RingBuffers.emplace_back(InitialSize, Allocator, pDevice->GetD3D12Device(), m_bIsCPUAccessible);
     }
 
-    DynamicAllocation DynamicUploadHeap::Allocate(size_t SizeInBytes, size_t Alignment /*= DEFAULT_ALIGN*/)
+    DynamicAllocation DynamicUploadHeap::Allocate(size_t SizeInBytes, size_t Alignment)
     {
         // Every device context has its own upload heap, so there is no need to lock
 
@@ -128,18 +128,15 @@ namespace Diligent
         //      across several frames!
         //
 
-	    const size_t AlignmentMask = Alignment - 1;
-	    // Assert that it's a power of two.
-	    VERIFY_EXPR((AlignmentMask & Alignment) == 0);
+	    VERIFY_EXPR(IsPowerOfTwo(Alignment));
 	    // Align the allocation
-	    const size_t AlignedSize = (SizeInBytes + AlignmentMask) & ~AlignmentMask;
-        auto DynAlloc = m_RingBuffers.back().Allocate(AlignedSize);
+        auto DynAlloc = m_RingBuffers.back().Allocate(SizeInBytes, Alignment);
         if (!DynAlloc.pBuffer)
         {
             auto NewMaxSize = m_RingBuffers.back().GetMaxSize() * 2;
-            while(NewMaxSize < AlignedSize)NewMaxSize*=2;
+            while(NewMaxSize < SizeInBytes)NewMaxSize*=2;
             m_RingBuffers.emplace_back(NewMaxSize, m_Allocator, m_pDeviceD3D12->GetD3D12Device(), m_bIsCPUAccessible);
-            DynAlloc = m_RingBuffers.back().Allocate(AlignedSize);
+            DynAlloc = m_RingBuffers.back().Allocate(SizeInBytes, Alignment);
         }
 #ifdef _DEBUG
 		DynAlloc.FrameNum = m_pDeviceD3D12->GetCurrentFrameNumber();