Initial work for GPU upload heap & UMA.

UnleashedRecomp/CMakeLists.txt

@@ -310,7 +310,11 @@ endif()
 if (UNLEASHED_RECOMP_D3D12)
     find_package(directx-headers CONFIG REQUIRED)
     find_package(directx12-agility CONFIG REQUIRED)
-    target_compile_definitions(UnleashedRecomp PRIVATE UNLEASHED_RECOMP_D3D12)
+    target_compile_definitions(UnleashedRecomp PRIVATE 
+        UNLEASHED_RECOMP_D3D12
+        D3D12MA_USING_DIRECTX_HEADERS
+        D3D12MA_OPTIONS16_SUPPORTED
+    )
 endif()
 
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")

UnleashedRecomp/gpu/rhi/plume_d3d12.cpp

@@ -442,6 +442,8 @@ namespace plume {
             return D3D12_HEAP_TYPE_UPLOAD;
         case RenderHeapType::READBACK:
             return D3D12_HEAP_TYPE_READBACK;
+        case RenderHeapType::GPU_UPLOAD:
+            return D3D12_HEAP_TYPE_GPU_UPLOAD;
         default:
             assert(false && "Unknown heap type.");
             return D3D12_HEAP_TYPE_DEFAULT;
@@ -3391,12 +3393,14 @@ namespace plume {
                 triangleFanSupportOption = d3d12Options15.TriangleFanSupported;
             }
 
-            // Check if dynamic depth bias is supported.
+            // Check if dynamic depth bias and GPU upload heap are supported.
             bool dynamicDepthBiasOption = false;
+            bool gpuUploadHeapOption = false;
             D3D12_FEATURE_DATA_D3D12_OPTIONS16 d3d12Options16 = {};
             res = deviceOption->CheckFeatureSupport(D3D12_FEATURE_D3D12_OPTIONS16, &d3d12Options16, sizeof(d3d12Options16));
             if (SUCCEEDED(res)) {
                 dynamicDepthBiasOption = d3d12Options16.DynamicDepthBiasSupported;
+                gpuUploadHeapOption = d3d12Options16.GPUUploadHeapSupported;
             }
 
             // Check if the architecture has UMA.
@@ -3431,6 +3435,7 @@ namespace plume {
                 capabilities.triangleFan = triangleFanSupportOption;
                 capabilities.dynamicDepthBias = dynamicDepthBiasOption;
                 capabilities.uma = uma;
+                capabilities.gpuUploadHeap = gpuUploadHeapOption;
                 description.name = deviceName;
                 description.dedicatedVideoMemory = adapterDesc.DedicatedVideoMemory;
                 description.vendor = RenderDeviceVendor(adapterDesc.VendorId);

UnleashedRecomp/gpu/rhi/plume_render_interface_types.h

@@ -351,7 +351,8 @@ namespace plume {
         UNKNOWN,
         DEFAULT,
         UPLOAD,
-        READBACK
+        READBACK,
+        GPU_UPLOAD
     };
 
     enum class RenderTextureArrangement {
@@ -1807,6 +1808,9 @@ namespace plume {
 
         // UMA.
         bool uma = false;
+
+        // GPU Upload heap.
+        bool gpuUploadHeap = false;
     };
 
     struct RenderInterfaceCapabilities {

UnleashedRecomp/gpu/rhi/plume_vulkan.cpp

@@ -808,6 +808,12 @@ namespace plume {
             bufferInfo.usage |= VK_BUFFER_USAGE_TRANSFER_DST_BIT;
             createInfo.flags |= VMA_ALLOCATION_CREATE_HOST_ACCESS_RANDOM_BIT;
             break;
+        case RenderHeapType::GPU_UPLOAD:
+            bufferInfo.usage |= VK_BUFFER_USAGE_TRANSFER_SRC_BIT;
+            bufferInfo.usage |= VK_BUFFER_USAGE_TRANSFER_DST_BIT;
+            createInfo.flags |= VMA_ALLOCATION_CREATE_HOST_ACCESS_SEQUENTIAL_WRITE_BIT;
+            createInfo.requiredFlags |= VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT | VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT;
+            break;
         default:
             assert(false && "Unknown heap type.");
             break;
@@ -3907,6 +3913,7 @@ namespace plume {
         capabilities.preferHDR = memoryHeapSize > (512 * 1024 * 1024);
         capabilities.triangleFan = true;
         capabilities.dynamicDepthBias = true;
+        capabilities.gpuUploadHeap = true; // TODO: Do a test buffer allocation with the required flags to set this.
 
         // Fill Vulkan-only capabilities.
         loadStoreOpNoneSupported = supportedOptionalExtensions.find(VK_EXT_LOAD_STORE_OP_NONE_EXTENSION_NAME) != supportedOptionalExtensions.end();

UnleashedRecomp/gpu/video.cpp

@@ -2112,36 +2112,46 @@ static void* LockVertexBuffer(GuestBuffer* buffer, uint32_t, uint32_t, uint32_t
 template<typename T>
 static void UnlockBuffer(GuestBuffer* buffer, bool useCopyQueue)
 {
-    auto uploadBuffer = g_device->createBuffer(RenderBufferDesc::UploadBuffer(buffer->dataSize));
+    auto copyBuffer = [&](T* dest)
+        {
+            auto src = reinterpret_cast<const T*>(buffer->mappedMemory);
 
-    auto dest = reinterpret_cast<T*>(uploadBuffer->map());
+            for (size_t i = 0; i < buffer->dataSize; i += sizeof(T))
-    auto src = reinterpret_cast<const T*>(buffer->mappedMemory);
-
-    for (size_t i = 0; i < buffer->dataSize; i += sizeof(T))
-    {
-        *dest = ByteSwap(*src);
-        ++dest;
-        ++src;
-    }
-
-    uploadBuffer->unmap();
-
-    if (useCopyQueue)
-    {
-        ExecuteCopyCommandList([&]
             {
-                g_copyCommandList->copyBufferRegion(buffer->buffer->at(0), uploadBuffer->at(0), buffer->dataSize);
+                *dest = ByteSwap(*src);
-            });
+                ++dest;
+                ++src;
+            }
+        };
+
+    if (useCopyQueue && (g_capabilities.uma || g_capabilities.gpuUploadHeap))
+    {
+        copyBuffer(reinterpret_cast<T*>(buffer->buffer->map()));
+        buffer->buffer->unmap();
     }
     else
     {
-        auto& commandList = g_commandLists[g_frame];
+        auto uploadBuffer = g_device->createBuffer(RenderBufferDesc::UploadBuffer(buffer->dataSize));
+        copyBuffer(reinterpret_cast<T*>(uploadBuffer->map()));
+        uploadBuffer->unmap();
 
-        commandList->barriers(RenderBarrierStage::COPY, RenderBufferBarrier(buffer->buffer.get(), RenderBufferAccess::WRITE));
+        if (useCopyQueue)
-        commandList->copyBufferRegion(buffer->buffer->at(0), uploadBuffer->at(0), buffer->dataSize);
+        {
-        commandList->barriers(RenderBarrierStage::GRAPHICS, RenderBufferBarrier(buffer->buffer.get(), RenderBufferAccess::READ));
+            ExecuteCopyCommandList([&]
+                {
+                    g_copyCommandList->copyBufferRegion(buffer->buffer->at(0), uploadBuffer->at(0), buffer->dataSize);
+                });
+        }
+        else
+        {
+            auto& commandList = g_commandLists[g_frame];
 
-        g_tempBuffers[g_frame].emplace_back(std::move(uploadBuffer));
+            commandList->barriers(RenderBarrierStage::COPY, RenderBufferBarrier(buffer->buffer.get(), RenderBufferAccess::WRITE));
+            commandList->copyBufferRegion(buffer->buffer->at(0), uploadBuffer->at(0), buffer->dataSize);
+            commandList->barriers(RenderBarrierStage::GRAPHICS, RenderBufferBarrier(buffer->buffer.get(), RenderBufferAccess::READ));
+
+            g_tempBuffers[g_frame].emplace_back(std::move(uploadBuffer));
+        }
     }
 }
 
@@ -2339,6 +2349,7 @@ static void DrawProfiler()
         ImGui::Text("Device Type: %s", DeviceTypeName(g_device->getDescription().type));
         ImGui::Text("VRAM: %.2f MiB", (double)(g_device->getDescription().dedicatedVideoMemory) / (1024.0 * 1024.0));
         ImGui::Text("UMA: %s", g_capabilities.uma ? "Supported" : "Unsupported");
+        ImGui::Text("GPU Upload Heap: %s", g_capabilities.gpuUploadHeap ? "Supported" : "Unsupported");
 
         const char* sdlVideoDriver = SDL_GetCurrentVideoDriver();
         if (sdlVideoDriver != nullptr)
@@ -3019,10 +3030,15 @@ static GuestTexture* CreateTexture(uint32_t width, uint32_t height, uint32_t dep
     return texture;
 }
 
+static RenderHeapType GetBufferHeapType()
+{
+    return g_capabilities.gpuUploadHeap ? RenderHeapType::GPU_UPLOAD : RenderHeapType::DEFAULT;
+}
+
 static GuestBuffer* CreateVertexBuffer(uint32_t length) 
 {
     auto buffer = g_userHeap.AllocPhysical<GuestBuffer>(ResourceType::VertexBuffer);
-    buffer->buffer = g_device->createBuffer(RenderBufferDesc::VertexBuffer(length, RenderHeapType::DEFAULT, RenderBufferFlag::INDEX));
+    buffer->buffer = g_device->createBuffer(RenderBufferDesc::VertexBuffer(length, GetBufferHeapType(), RenderBufferFlag::INDEX));
     buffer->dataSize = length;
 #ifdef _DEBUG 
     buffer->buffer->setName(fmt::format("Vertex Buffer {:X}", g_memory.MapVirtual(buffer)));
@@ -3033,7 +3049,7 @@ static GuestBuffer* CreateVertexBuffer(uint32_t length)
 static GuestBuffer* CreateIndexBuffer(uint32_t length, uint32_t, uint32_t format)
 {
     auto buffer = g_userHeap.AllocPhysical<GuestBuffer>(ResourceType::IndexBuffer);
-    buffer->buffer = g_device->createBuffer(RenderBufferDesc::IndexBuffer(length, RenderHeapType::DEFAULT));
+    buffer->buffer = g_device->createBuffer(RenderBufferDesc::IndexBuffer(length, GetBufferHeapType()));
     buffer->dataSize = length;
     buffer->format = ConvertFormat(format);
     buffer->guestFormat = format;