GPU Upload Heap & UMA. (#1421)

* Initial work for GPU upload heap & UMA.

* Finish D3D12 Support.

* Rework the logic for the GPU Upload Heap fallback.

* Only enable UMA on Vulkan on integrated GPUs.

* Fix D3D12 fallback condition.

---------

Co-authored-by: Dario <dariosamo@gmail.com>

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;
@@ -2385,7 +2387,7 @@ namespace plume {
             range.End = readRange->end;
         }
 
-        void *outputData;
+        void *outputData = nullptr;
         d3d->Map(subresource, (readRange != nullptr) ? &range : nullptr, &outputData);
         return outputData;
     }
@@ -2629,14 +2631,22 @@ namespace plume {
 
     // D3D12Pool
 
-    D3D12Pool::D3D12Pool(D3D12Device *device, const RenderPoolDesc &desc) {
+    D3D12Pool::D3D12Pool(D3D12Device *device, const RenderPoolDesc &desc, bool gpuUploadHeapFallback) {
         assert(device != nullptr);
 
         this->device = device;
         this->desc = desc;
 
         D3D12MA::POOL_DESC poolDesc = {};
+
+        // When using an UMA architecture without explicit support for GPU Upload heaps, we instead just make a custom heap with the same properties as Upload heaps.
+        if ((desc.heapType == RenderHeapType::GPU_UPLOAD) && gpuUploadHeapFallback) {
+            poolDesc.HeapProperties = device->d3d->GetCustomHeapProperties(0, D3D12_HEAP_TYPE_UPLOAD);
+        }
+        else {
             poolDesc.HeapProperties.Type = toD3D12(desc.heapType);
+        }
+
         poolDesc.MinBlockCount = desc.minBlockCount;
         poolDesc.MaxBlockCount = desc.maxBlockCount;
         poolDesc.Flags |= desc.useLinearAlgorithm ? D3D12MA::POOL_FLAG_ALGORITHM_LINEAR : D3D12MA::POOL_FLAG_NONE;
@@ -3391,12 +3401,14 @@ 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 +3443,11 @@ namespace plume {
                 capabilities.triangleFan = triangleFanSupportOption;
                 capabilities.dynamicDepthBias = dynamicDepthBiasOption;
                 capabilities.uma = uma;
+
+                // Pretend GPU Upload heaps are supported if UMA is supported, as the backend has a workaround using a custom pool for it.
+                capabilities.gpuUploadHeap = uma || gpuUploadHeapOption;
+                gpuUploadHeapFallback = uma && !gpuUploadHeapOption;
+
                 description.name = deviceName;
                 description.dedicatedVideoMemory = adapterDesc.DedicatedVideoMemory;
                 description.vendor = RenderDeviceVendor(adapterDesc.VendorId);
@@ -3528,6 +3545,13 @@ namespace plume {
         colorTargetHeapAllocator = std::make_unique<D3D12DescriptorHeapAllocator>(this, TargetDescriptorHeapSize, D3D12_DESCRIPTOR_HEAP_TYPE_RTV);
         depthTargetHeapAllocator = std::make_unique<D3D12DescriptorHeapAllocator>(this, TargetDescriptorHeapSize, D3D12_DESCRIPTOR_HEAP_TYPE_DSV);
 
+        // Create the custom upload pool that will be used as the fallback when using an UMA architecture without explicit support for GPU Upload heaps.
+        if (gpuUploadHeapFallback) {
+            RenderPoolDesc poolDesc;
+            poolDesc.heapType = RenderHeapType::GPU_UPLOAD;
+            customUploadPool = std::make_unique<D3D12Pool>(this, poolDesc, true);
+        }
+
         // Create a command queue only for retrieving the timestamp frequency. Delete it immediately afterwards.
         std::unique_ptr<D3D12CommandQueue> timestampCommandQueue = std::make_unique<D3D12CommandQueue>(this, RenderCommandListType::DIRECT);
         res = timestampCommandQueue->d3d->GetTimestampFrequency(&timestampFrequency);
@@ -3577,8 +3601,13 @@ namespace plume {
     }
     
     std::unique_ptr<RenderBuffer> D3D12Device::createBuffer(const RenderBufferDesc &desc) {
+        if ((desc.heapType == RenderHeapType::GPU_UPLOAD) && gpuUploadHeapFallback) {
+            return std::make_unique<D3D12Buffer>(this, customUploadPool.get(), desc);
+        }
+        else {
             return std::make_unique<D3D12Buffer>(this, nullptr, desc);
         }
+    }
 
     std::unique_ptr<RenderTexture> D3D12Device::createTexture(const RenderTextureDesc &desc) {
         return std::make_unique<D3D12Texture>(this, nullptr, desc);
@@ -3589,7 +3618,7 @@ namespace plume {
     }
 
     std::unique_ptr<RenderPool> D3D12Device::createPool(const RenderPoolDesc &desc) {
-        return std::make_unique<D3D12Pool>(this, desc);
+        return std::make_unique<D3D12Pool>(this, desc, gpuUploadHeapFallback);
     }
 
     std::unique_ptr<RenderPipelineLayout> D3D12Device::createPipelineLayout(const RenderPipelineLayoutDesc &desc) {

UnleashedRecomp/gpu/rhi/plume_d3d12.h

@@ -329,7 +329,7 @@ namespace plume {
         D3D12Device *device = nullptr;
         RenderPoolDesc desc;
 
-        D3D12Pool(D3D12Device *device, const RenderPoolDesc &desc);
+        D3D12Pool(D3D12Device *device, const RenderPoolDesc &desc, bool gpuUploadHeapFallback);
         ~D3D12Pool() override;
         std::unique_ptr<RenderBuffer> createBuffer(const RenderBufferDesc &desc) override;
         std::unique_ptr<RenderTexture> createTexture(const RenderTextureDesc &desc) override;
@@ -430,9 +430,11 @@ namespace plume {
         std::unique_ptr<D3D12DescriptorHeapAllocator> samplerHeapAllocator;
         std::unique_ptr<D3D12DescriptorHeapAllocator> colorTargetHeapAllocator;
         std::unique_ptr<D3D12DescriptorHeapAllocator> depthTargetHeapAllocator;
+        std::unique_ptr<D3D12Pool> customUploadPool;
         RenderDeviceCapabilities capabilities;
         RenderDeviceDescription description;
         uint64_t timestampFrequency = 1;
+        bool gpuUploadHeapFallback = false;
 
         D3D12Device(D3D12Interface *renderInterface, const std::string &preferredDeviceName);
         ~D3D12Device() override;

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;
@@ -833,7 +839,7 @@ namespace plume {
         }
 
         if (res != VK_SUCCESS) {
-            fprintf(stderr, "vkCreateBuffer failed with error code 0x%X.\n", res);
+            fprintf(stderr, "vmaCreateBuffer failed with error code 0x%X.\n", res);
             return;
         }
     }
@@ -3887,6 +3893,15 @@ namespace plume {
         VkDeviceSize memoryHeapSize = 0;
         const VkPhysicalDeviceMemoryProperties *memoryProps = nullptr;
         vmaGetMemoryProperties(allocator, &memoryProps);
+
+        constexpr VkMemoryPropertyFlags uploadHeapPropertyFlags = VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT | VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
+        bool hasHostVisibleDeviceLocalMemory = false;
+        for (uint32_t i = 0; i < memoryProps->memoryTypeCount; i++) {
+            if ((memoryProps->memoryTypes[i].propertyFlags & uploadHeapPropertyFlags) == uploadHeapPropertyFlags) {
+                hasHostVisibleDeviceLocalMemory = true;
+            }
+        }
+
         for (uint32_t i = 0; i < memoryProps->memoryHeapCount; i++) {
             if (memoryProps->memoryHeaps[i].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
                 memoryHeapSize = std::max(memoryProps->memoryHeaps[i].size, memoryHeapSize);
@@ -3907,6 +3922,8 @@ namespace plume {
         capabilities.preferHDR = memoryHeapSize > (512 * 1024 * 1024);
         capabilities.triangleFan = true;
         capabilities.dynamicDepthBias = true;
+        capabilities.uma = (description.type == RenderDeviceType::INTEGRATED) && hasHostVisibleDeviceLocalMemory;
+        capabilities.gpuUploadHeap = capabilities.uma;
 
         // Fill Vulkan-only capabilities.
         loadStoreOpNoneSupported = supportedOptionalExtensions.find(VK_EXT_LOAD_STORE_OP_NONE_EXTENSION_NAME) != supportedOptionalExtensions.end();

UnleashedRecomp/gpu/video.cpp

@@ -2114,12 +2114,13 @@ static void* LockVertexBuffer(GuestBuffer* buffer, uint32_t, uint32_t, uint32_t
     return LockBuffer(buffer, flags);
 }
 
+static std::atomic<uint32_t> g_bufferUploadCount = 0;
+
 template<typename T>
 static void UnlockBuffer(GuestBuffer* buffer, bool useCopyQueue)
 {
-    auto uploadBuffer = g_device->createBuffer(RenderBufferDesc::UploadBuffer(buffer->dataSize));
+    auto copyBuffer = [&](T* dest)
-
+        {
-    auto dest = reinterpret_cast<T*>(uploadBuffer->map());
             auto src = reinterpret_cast<const T*>(buffer->mappedMemory);
 
             for (size_t i = 0; i < buffer->dataSize; i += sizeof(T))
@@ -2128,7 +2129,17 @@ static void UnlockBuffer(GuestBuffer* buffer, bool useCopyQueue)
                 ++dest;
                 ++src;
             }
+        };
 
+    if (useCopyQueue && g_capabilities.gpuUploadHeap)
+    {
+        copyBuffer(reinterpret_cast<T*>(buffer->buffer->map()));
+        buffer->buffer->unmap();
+    }
+    else
+    {
+        auto uploadBuffer = g_device->createBuffer(RenderBufferDesc::UploadBuffer(buffer->dataSize));
+        copyBuffer(reinterpret_cast<T*>(uploadBuffer->map()));
         uploadBuffer->unmap();
 
         if (useCopyQueue)
@@ -2148,6 +2159,9 @@ static void UnlockBuffer(GuestBuffer* buffer, bool useCopyQueue)
 
             g_tempBuffers[g_frame].emplace_back(std::move(uploadBuffer));
         }
+    }
+
+    g_bufferUploadCount++;
 }
 
 template<typename T>
@@ -2329,6 +2343,7 @@ static void DrawProfiler()
         ImGui::Text("Heap Allocated: %d MB", int32_t(diagnostics.allocated / (1024 * 1024)));
         ImGui::Text("Physical Heap Allocated: %d MB", int32_t(physicalDiagnostics.allocated / (1024 * 1024)));
         ImGui::Text("GPU Waits: %d", int32_t(g_waitForGPUCount));
+        ImGui::Text("Buffer Uploads: %d", int32_t(g_bufferUploadCount));
         ImGui::NewLine();
 
         ImGui::Text("Present Wait: %s", g_capabilities.presentWait ? "Supported" : "Unsupported");
@@ -2344,6 +2359,7 @@ 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)
@@ -3024,10 +3040,15 @@ 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)));
@@ -3038,7 +3059,7 @@ 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;