ResourceAllocator: several changes:

- "D3D12_RESOURCE_DESC::Alignment" set to 0 to unlock optimizations
- Preferred block size explicitly set to 64 Mb (currently) to avoid divergence, because it's 64 Mb in D3D12MA and 256 Mb in VMA
- added "STRATEGY_MIN_MEMORY" flag

Source/D3D12/BufferD3D12.hpp

@@ -3,7 +3,6 @@
 void nri::GetResourceDesc(D3D12_RESOURCE_DESC* desc, const BufferDesc& bufferDesc) {
     *desc = {};
     desc->Dimension = D3D12_RESOURCE_DIMENSION_BUFFER;
-    desc->Alignment = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT; // 64KB
     desc->Width = bufferDesc.size;
     desc->Height = 1;
     desc->DepthOrArraySize = 1;

Source/D3D12/QueryPoolD3D12.hpp

@@ -45,7 +45,6 @@ Result QueryPoolD3D12::Create(const QueryPoolDesc& queryPoolDesc) {
 Result QueryPoolD3D12::CreateReadbackBuffer(const QueryPoolDesc& queryPoolDesc) {
     D3D12_RESOURCE_DESC resourceDesc = {};
     resourceDesc.Dimension = D3D12_RESOURCE_DIMENSION_BUFFER;
-    resourceDesc.Alignment = D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT;
     resourceDesc.Width = (uint64_t)queryPoolDesc.capacity * m_QuerySize;
     resourceDesc.Height = 1;
     resourceDesc.DepthOrArraySize = 1;

Source/D3D12/ResourceAllocatorD3D12.hpp

@@ -8,8 +8,8 @@ Result DeviceD3D12::CreateVma() {
     allocatorDesc.pDevice = m_Device;
     allocatorDesc.pAdapter = m_Adapter;
     allocatorDesc.pAllocationCallbacks = m_AllocationCallbackPtr;
-    allocatorDesc.Flags = (D3D12MA::ALLOCATOR_FLAGS)(
-        D3D12MA::ALLOCATOR_FLAG_DEFAULT_POOLS_NOT_ZEROED | D3D12MA::ALLOCATOR_FLAG_MSAA_TEXTURES_ALWAYS_COMMITTED | D3D12MA::ALLOCATOR_FLAG_DONT_PREFER_SMALL_BUFFERS_COMMITTED);
+    allocatorDesc.Flags = (D3D12MA::ALLOCATOR_FLAGS)(D3D12MA::ALLOCATOR_FLAG_DEFAULT_POOLS_NOT_ZEROED | D3D12MA::ALLOCATOR_FLAG_MSAA_TEXTURES_ALWAYS_COMMITTED | D3D12MA::ALLOCATOR_FLAG_DONT_PREFER_SMALL_BUFFERS_COMMITTED);
+    allocatorDesc.PreferredBlockSize = VMA_PREFERRED_BLOCK_SIZE;
 
     HRESULT hr = D3D12MA::CreateAllocator(&allocatorDesc, &m_Vma);
     RETURN_ON_BAD_HRESULT(this, hr, "D3D12MA::CreateAllocator");
@@ -24,7 +24,7 @@ Result BufferD3D12::Create(const AllocateBufferDesc& bufferDesc) {
 
     D3D12MA::ALLOCATION_DESC allocationDesc = {};
     allocationDesc.HeapType = GetHeapType(bufferDesc.memoryLocation);
-    allocationDesc.Flags = D3D12MA::ALLOCATION_FLAG_CAN_ALIAS;
+    allocationDesc.Flags = (D3D12MA::ALLOCATION_FLAGS)(D3D12MA::ALLOCATION_FLAG_CAN_ALIAS | D3D12MA::ALLOCATION_FLAG_STRATEGY_MIN_MEMORY);
 
 #ifdef NRI_USE_AGILITY_SDK
     if (m_Device.GetVersion() >= 10) {
@@ -80,7 +80,7 @@ Result TextureD3D12::Create(const AllocateTextureDesc& textureDesc) {
 
     D3D12MA::ALLOCATION_DESC allocationDesc = {};
     allocationDesc.HeapType = GetHeapType(textureDesc.memoryLocation);
-    allocationDesc.Flags = D3D12MA::ALLOCATION_FLAG_CAN_ALIAS;
+    allocationDesc.Flags = (D3D12MA::ALLOCATION_FLAGS)(D3D12MA::ALLOCATION_FLAG_CAN_ALIAS | D3D12MA::ALLOCATION_FLAG_STRATEGY_MIN_MEMORY);
 
 #ifdef NRI_USE_AGILITY_SDK
     if (m_Device.GetVersion() >= 10) {

Source/D3D12/TextureD3D12.hpp

@@ -4,8 +4,8 @@ void nri::GetResourceDesc(D3D12_RESOURCE_DESC* desc, const TextureDesc& textureD
     uint16_t blockWidth = (uint16_t)GetFormatProps(textureDesc.format).blockWidth;
     const DxgiFormat& dxgiFormat = GetDxgiFormat(textureDesc.format);
 
+    *desc = {};
     desc->Dimension = GetResourceDimension(textureDesc.type);
-    desc->Alignment = textureDesc.sampleNum > 1 ? 0 : D3D12_DEFAULT_RESOURCE_PLACEMENT_ALIGNMENT;
     desc->Width = Align(textureDesc.width, blockWidth);
     desc->Height = Align(std::max(textureDesc.height, (Dim_t)1), blockWidth);
     desc->DepthOrArraySize = std::max(textureDesc.type == TextureType::TEXTURE_3D ? textureDesc.depth : textureDesc.layerNum, (Dim_t)1);

Source/Shared/SharedExternal.h

@@ -44,6 +44,7 @@ constexpr uint32_t TIMEOUT_PRESENT = 1000; // 1 sec
 constexpr uint32_t TIMEOUT_FENCE = 5000;   // 5 sec
 constexpr uint64_t PRESENT_INDEX_BIT_NUM = 56ull;
 constexpr uint32_t MAX_MESSAGE_LENGTH = 2048;
+constexpr uint64_t VMA_PREFERRED_BLOCK_SIZE = 64 * 1024 * 1024;
 
 // https://learn.microsoft.com/en-us/windows/win32/direct3d12/root-signature-limits
 constexpr uint32_t ROOT_SIGNATURE_DWORD_NUM = 64;

Source/VK/ResourceAllocatorVK.hpp

@@ -20,6 +20,7 @@ Result DeviceVK::CreateVma() {
     allocatorCreateInfo.instance = m_Instance;
     allocatorCreateInfo.pVulkanFunctions = &vulkanFunctions;
     allocatorCreateInfo.pAllocationCallbacks = m_AllocationCallbackPtr;
+    allocatorCreateInfo.preferredLargeHeapBlockSize = VMA_PREFERRED_BLOCK_SIZE;
 
     if (m_IsSupported.memoryBudget)
         allocatorCreateInfo.flags |= VMA_ALLOCATOR_CREATE_EXT_MEMORY_BUDGET_BIT;
@@ -49,7 +50,7 @@ Result BufferVK::Create(const AllocateBufferDesc& bufferDesc) {
 
     // Create
     VmaAllocationCreateInfo allocationCreateInfo = {};
-    allocationCreateInfo.flags = VMA_ALLOCATION_CREATE_CAN_ALIAS_BIT;
+    allocationCreateInfo.flags = VMA_ALLOCATION_CREATE_CAN_ALIAS_BIT | VMA_ALLOCATION_CREATE_STRATEGY_MIN_MEMORY_BIT;
     allocationCreateInfo.priority = bufferDesc.memoryPriority * 0.5f + 0.5f;
     allocationCreateInfo.usage = IsHostMemory(bufferDesc.memoryLocation) ? VMA_MEMORY_USAGE_AUTO_PREFER_HOST : VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
 
@@ -115,7 +116,7 @@ Result TextureVK::Create(const AllocateTextureDesc& textureDesc) {
 
     // Create
     VmaAllocationCreateInfo allocationCreateInfo = {};
-    allocationCreateInfo.flags = VMA_ALLOCATION_CREATE_CAN_ALIAS_BIT;
+    allocationCreateInfo.flags = VMA_ALLOCATION_CREATE_CAN_ALIAS_BIT | VMA_ALLOCATION_CREATE_STRATEGY_MIN_MEMORY_BIT;
     allocationCreateInfo.priority = textureDesc.memoryPriority * 0.5f + 0.5f;
     allocationCreateInfo.usage = IsHostMemory(textureDesc.memoryLocation) ? VMA_MEMORY_USAGE_AUTO_PREFER_HOST : VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;