Add timestamps to D3D12.

UnleashedRecomp/gpu/rhi/plume_d3d12.cpp

@@ -1437,6 +1437,52 @@ namespace plume {
         return height;
     }
 
+    // D3D12QueryPool
+
+    D3D12QueryPool::D3D12QueryPool(D3D12Device *device, uint32_t queryCount) {
+        assert(device != nullptr);
+        assert(queryCount > 0);
+
+        this->device = device;
+
+        D3D12_QUERY_HEAP_DESC queryHeapDesc = {};
+        queryHeapDesc.Type = D3D12_QUERY_HEAP_TYPE_TIMESTAMP;
+        queryHeapDesc.Count = queryCount;
+
+        HRESULT res = device->d3d->CreateQueryHeap(&queryHeapDesc, IID_PPV_ARGS(&d3d));
+        if (FAILED(res)) {
+            fprintf(stderr, "CreateQueryHeap failed with error code 0x%lX.\n", res);
+            return;
+        }
+
+        readbackBuffer = device->createBuffer(RenderBufferDesc::ReadbackBuffer(sizeof(uint64_t) * queryCount));
+        results.resize(queryCount);
+    }
+
+    D3D12QueryPool::~D3D12QueryPool() {
+        if (d3d != nullptr) {
+            d3d->Release();
+        }
+    }
+
+    void D3D12QueryPool::queryResults() {
+        void *readbackData = readbackBuffer->map();
+        memcpy(results.data(), readbackData, sizeof(uint64_t) * results.size());
+        readbackBuffer->unmap();
+
+        for (uint64_t &result : results) {
+            result = result / double(device->timestampFrequency) * 1000000000.0;
+        }
+    }
+
+    const uint64_t *D3D12QueryPool::getResults() const {
+        return results.data();
+    }
+
+    uint32_t D3D12QueryPool::getCount() const {
+        return uint32_t(results.size());
+    }
+
     // D3D12CommandList
 
     D3D12CommandList::D3D12CommandList(D3D12Device *device, RenderCommandListType type) {
@@ -2004,6 +2050,19 @@ namespace plume {
         d3d->DiscardResource(interfaceTexture->d3d, nullptr);
     }
 
+    void D3D12CommandList::resetQueryPool(const RenderQueryPool *queryPool, uint32_t queryFirstIndex, uint32_t queryCount) {
+        // Do nothing.
+    }
+
+    void D3D12CommandList::writeTimestamp(const RenderQueryPool *queryPool, uint32_t queryIndex) {
+        assert(queryPool != nullptr);
+
+        const D3D12QueryPool *interfaceQueryPool = static_cast<const D3D12QueryPool *>(queryPool);
+        const D3D12Buffer *readbackBuffer = static_cast<const D3D12Buffer *>(interfaceQueryPool->readbackBuffer.get());
+        d3d->EndQuery(interfaceQueryPool->d3d, D3D12_QUERY_TYPE_TIMESTAMP, queryIndex);
+        d3d->ResolveQueryData(interfaceQueryPool->d3d, D3D12_QUERY_TYPE_TIMESTAMP, queryIndex, 1, readbackBuffer->d3d, queryIndex * sizeof(uint64_t));
+    }
+
     void D3D12CommandList::checkDescriptorHeaps() {
         if (!descriptorHeapsSet) {
             ID3D12DescriptorHeap *descriptorHeaps[] = { device->viewHeapAllocator->heap, device->samplerHeapAllocator->heap };
@@ -3461,6 +3520,13 @@ namespace plume {
         samplerHeapAllocator = std::make_unique<D3D12DescriptorHeapAllocator>(this, SamplerDescriptorHeapSize, D3D12_DESCRIPTOR_HEAP_TYPE_SAMPLER);
         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 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);
+        if (FAILED(res)) {
+            fprintf(stderr, "GetTimestampFrequency failed with error code 0x%lX. Timestamps will be inaccurate.\n", res);
+        }
     }
 
     D3D12Device::~D3D12Device() {
@@ -3535,6 +3601,10 @@ namespace plume {
         return std::make_unique<D3D12Framebuffer>(this, desc);
     }
 
+    std::unique_ptr<RenderQueryPool> D3D12Device::createQueryPool(uint32_t queryCount) {
+        return std::make_unique<D3D12QueryPool>(this, queryCount);
+    }
+
     void D3D12Device::setBottomLevelASBuildInfo(RenderBottomLevelASBuildInfo &buildInfo, const RenderBottomLevelASMesh *meshes, uint32_t meshCount, bool preferFastBuild, bool preferFastTrace) {
         assert(meshes != nullptr);
         assert(meshCount > 0);

UnleashedRecomp/gpu/rhi/plume_d3d12.h

@@ -144,6 +144,19 @@ namespace plume {
         uint32_t getHeight() const override;
     };
 
+    struct D3D12QueryPool : RenderQueryPool {
+        D3D12Device *device = nullptr;
+        ID3D12QueryHeap *d3d = nullptr;
+        std::vector<uint64_t> results;
+        std::unique_ptr<RenderBuffer> readbackBuffer;
+
+        D3D12QueryPool(D3D12Device *device, uint32_t queryCount);
+        virtual ~D3D12QueryPool() override;
+        virtual void queryResults() override;
+        virtual const uint64_t *getResults() const override;
+        virtual uint32_t getCount() const override;
+    };
+
     struct D3D12CommandList : RenderCommandList {
         ID3D12GraphicsCommandList9 *d3d = nullptr;
         ID3D12CommandAllocator *commandAllocator = nullptr;
@@ -196,6 +209,8 @@ namespace plume {
         void buildBottomLevelAS(const RenderAccelerationStructure *dstAccelerationStructure, RenderBufferReference scratchBuffer, const RenderBottomLevelASBuildInfo &buildInfo) override;
         void buildTopLevelAS(const RenderAccelerationStructure *dstAccelerationStructure, RenderBufferReference scratchBuffer, RenderBufferReference instancesBuffer, const RenderTopLevelASBuildInfo &buildInfo) override;
         void discardTexture(const RenderTexture* texture) override;
+        void resetQueryPool(const RenderQueryPool *queryPool, uint32_t queryFirstIndex, uint32_t queryCount) override;
+        void writeTimestamp(const RenderQueryPool *queryPool, uint32_t queryIndex) override;
         void checkDescriptorHeaps();
         void notifyDescriptorHeapWasChangedExternally();
         void checkTopology();
@@ -417,6 +432,7 @@ namespace plume {
         std::unique_ptr<D3D12DescriptorHeapAllocator> depthTargetHeapAllocator;
         RenderDeviceCapabilities capabilities;
         RenderDeviceDescription description;
+        uint64_t timestampFrequency = 1;
 
         D3D12Device(D3D12Interface *renderInterface, const std::string &preferredDeviceName);
         ~D3D12Device() override;
@@ -436,6 +452,7 @@ namespace plume {
         std::unique_ptr<RenderCommandFence> createCommandFence() override;
         std::unique_ptr<RenderCommandSemaphore> createCommandSemaphore() override;
         std::unique_ptr<RenderFramebuffer> createFramebuffer(const RenderFramebufferDesc &desc) override;
+        std::unique_ptr<RenderQueryPool> createQueryPool(uint32_t queryCount) override;
         void setBottomLevelASBuildInfo(RenderBottomLevelASBuildInfo &buildInfo, const RenderBottomLevelASMesh *meshes, uint32_t meshCount, bool preferFastBuild, bool preferFastTrace) override;
         void setTopLevelASBuildInfo(RenderTopLevelASBuildInfo &buildInfo, const RenderTopLevelASInstance *instances, uint32_t instanceCount, bool preferFastBuild, bool preferFastTrace) override;
         void setShaderBindingTableInfo(RenderShaderBindingTableInfo &tableInfo, const RenderShaderBindingGroups &groups, const RenderPipeline *pipeline, RenderDescriptorSet **descriptorSets, uint32_t descriptorSetCount) override;

UnleashedRecomp/gpu/rhi/plume_vulkan.cpp

@@ -2580,7 +2580,7 @@ namespace plume {
         constexpr uint64_t shift_bits = 16;
         double timestampPeriod = double(device->physicalDeviceProperties.limits.timestampPeriod);
         uint64_t h = 0, l = 0;
-        for (size_t &result : results) {
+        for (uint64_t &result : results) {
             mult64to128(result, uint64_t(timestampPeriod * double(1 << shift_bits)), h, l);
             result = l;
             result >>= shift_bits;

UnleashedRecomp/gpu/video.cpp

@@ -2254,9 +2254,9 @@ static void DrawProfiler()
         ImGui::Text("Current GPU Frame: %g ms (%g FPS)", g_gpuFrameProfiler.value.load(), 1000.0 / g_gpuFrameProfiler.value.load());
         ImGui::Text("Current Present: %g ms (%g FPS)", g_presentProfiler.value.load(), 1000.0 / g_presentProfiler.value.load());
         ImGui::Text("Current Render Director: %g ms (%g FPS)", g_renderDirectorProfiler.value.load(), 1000.0 / g_renderDirectorProfiler.value.load());
-        ImGui::Text("Current Frame Fence: %g ms (%g FPS)", g_frameFenceProfiler.value.load(), 1000.0 / g_frameFenceProfiler.value.load());
+        ImGui::Text("Current Frame Fence: %g ms", g_frameFenceProfiler.value.load());
-        ImGui::Text("Current Present Wait: %g ms (%g FPS)", g_presentWaitProfiler.value.load(), 1000.0 / g_presentWaitProfiler.value.load());
+        ImGui::Text("Current Present Wait: %g ms", g_presentWaitProfiler.value.load());
-        ImGui::Text("Current Swap Chain Acquire: %g ms (%g FPS)", g_swapChainAcquireProfiler.value.load(), 1000.0 / g_swapChainAcquireProfiler.value.load());
+        ImGui::Text("Current Swap Chain Acquire: %g ms", g_swapChainAcquireProfiler.value.load());
 
         ImGui::NewLine();
 
@@ -2264,9 +2264,9 @@ static void DrawProfiler()
         ImGui::Text("Average GPU Frame: %g ms (%g FPS)", gpuFrameAvg, 1000.0 / gpuFrameAvg);
         ImGui::Text("Average Present: %g ms (%g FPS)", presentAvg, 1000.0 / presentAvg);
         ImGui::Text("Average Render Director: %g ms (%g FPS)", renderDirectorAvg, 1000.0 / renderDirectorAvg);
-        ImGui::Text("Average Frame Fence: %g ms (%g FPS)", frameFenceAvg, 1000.0 / frameFenceAvg);
+        ImGui::Text("Average Frame Fence: %g ms", frameFenceAvg);
-        ImGui::Text("Average Present Wait: %g ms (%g FPS)", presentWaitAvg, 1000.0 / presentWaitAvg);
+        ImGui::Text("Average Present Wait: %g ms", presentWaitAvg);
-        ImGui::Text("Average Swap Chain Acquire: %g ms (%g FPS)", swapChainAcquireAvg, 1000.0 / swapChainAcquireAvg);
+        ImGui::Text("Average Swap Chain Acquire: %g ms", swapChainAcquireAvg);
 
         ImGui::NewLine();