Implement hardware resolve path for both color and depth targets.

UnleashedRecomp/gpu/rhi/plume_d3d12.cpp

@@ -27,6 +27,8 @@
 //#   define D3D12_DEBUG_LAYER_GPU_BASED_VALIDATION_ENABLED
 #endif
 
+//#define D3D12_DEBUG_SET_STABLE_POWER_STATE
+
 // Old Windows SDK versions don't provide this macro, so we workaround it by making sure it is defined.
 #ifndef D3D12_RESOURCE_STATE_ALL_SHADER_RESOURCE
 #define D3D12_RESOURCE_STATE_ALL_SHADER_RESOURCE (D3D12_RESOURCE_STATE_PIXEL_SHADER_RESOURCE | D3D12_RESOURCE_STATE_NON_PIXEL_SHADER_RESOURCE)
@@ -692,6 +694,20 @@ namespace plume {
         );
     }
 
+    static D3D12_RESOLVE_MODE toD3D12(RenderResolveMode resolveMode) {
+        switch (resolveMode) {
+        case RenderResolveMode::MIN:
+            return D3D12_RESOLVE_MODE_MIN;
+        case RenderResolveMode::MAX:
+            return D3D12_RESOLVE_MODE_MAX;
+        case RenderResolveMode::AVERAGE:
+            return D3D12_RESOLVE_MODE_AVERAGE;
+        default:
+            assert(false && "Unknown resolve mode.");
+            return D3D12_RESOLVE_MODE_AVERAGE;
+        }
+    }
+
     static void setObjectName(ID3D12Object *object, const std::string &name) {
         const std::wstring wideCharName = Utf8ToUtf16(name);
         object->SetName(wideCharName.c_str());
@@ -1916,7 +1932,7 @@ namespace plume {
         resetSamplePositions();
     }
 
-    void D3D12CommandList::resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect) {
+    void D3D12CommandList::resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect, RenderResolveMode resolveMode) {
         assert(dstTexture != nullptr);
         assert(srcTexture != nullptr);
 
@@ -1931,7 +1947,7 @@ namespace plume {
         }
 
         setSamplePositions(interfaceDstTexture);
-        d3d->ResolveSubresourceRegion(interfaceDstTexture->d3d, 0, dstX, dstY, interfaceSrcTexture->d3d, 0, (srcRect != nullptr) ? &rect : nullptr, toDXGI(interfaceDstTexture->desc.format), D3D12_RESOLVE_MODE_AVERAGE);
+        d3d->ResolveSubresourceRegion(interfaceDstTexture->d3d, 0, dstX, dstY, interfaceSrcTexture->d3d, 0, (srcRect != nullptr) ? &rect : nullptr, toDXGI(interfaceDstTexture->desc.format), toD3D12(resolveMode));
         resetSamplePositions();
     }
 
@@ -3373,6 +3389,10 @@ namespace plume {
             return;
         }
 
+        #ifdef D3D12_DEBUG_SET_STABLE_POWER_STATE
+            d3d->SetStablePowerState(TRUE);
+        #endif
+
         D3D12MA::ALLOCATOR_DESC allocatorDesc = {};
         allocatorDesc.pDevice = d3d;
         allocatorDesc.pAdapter = adapter;

UnleashedRecomp/gpu/rhi/plume_d3d12.h

@@ -192,7 +192,7 @@ namespace plume {
         void copyBuffer(const RenderBuffer *dstBuffer, const RenderBuffer *srcBuffer) override;
         void copyTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) override;
         void resolveTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) override;
-        void resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect) override;
+        void resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect, RenderResolveMode resolveMode) override;
         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;

UnleashedRecomp/gpu/rhi/plume_render_interface.h

@@ -143,7 +143,7 @@ namespace plume {
         virtual void copyBuffer(const RenderBuffer *dstBuffer, const RenderBuffer *srcBuffer) = 0;
         virtual void copyTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) = 0;
         virtual void resolveTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) = 0;
-        virtual void resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect = nullptr) = 0;
+        virtual void resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect = nullptr, RenderResolveMode resolveMode = RenderResolveMode::AVERAGE) = 0;
         virtual void buildBottomLevelAS(const RenderAccelerationStructure *dstAccelerationStructure, RenderBufferReference scratchBuffer, const RenderBottomLevelASBuildInfo &buildInfo) = 0;
         virtual void buildTopLevelAS(const RenderAccelerationStructure *dstAccelerationStructure, RenderBufferReference scratchBuffer, RenderBufferReference instancesBuffer, const RenderTopLevelASBuildInfo &buildInfo) = 0;
         virtual void discardTexture(const RenderTexture* texture) = 0; // D3D12 only.

UnleashedRecomp/gpu/rhi/plume_render_interface_types.h

@@ -483,6 +483,12 @@ namespace plume {
         CPU
     };
 
+    enum class RenderResolveMode {
+        MIN,
+        MAX,
+        AVERAGE
+    };
+
     // Global functions.
 
     constexpr uint32_t RenderFormatSize(RenderFormat format) {

UnleashedRecomp/gpu/rhi/plume_vulkan.cpp

@@ -3074,12 +3074,13 @@ namespace plume {
     }
 
     void VulkanCommandList::resolveTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) {
-        resolveTextureRegion(dstTexture, 0, 0, srcTexture, nullptr);
+        resolveTextureRegion(dstTexture, 0, 0, srcTexture, nullptr, RenderResolveMode::AVERAGE);
     }
 
-    void VulkanCommandList::resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect) {
+    void VulkanCommandList::resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect, RenderResolveMode resolveMode) {
         assert(dstTexture != nullptr);
         assert(srcTexture != nullptr);
+        assert(resolveMode == RenderResolveMode::AVERAGE && "Vulkan only supports AVERAGE resolve mode.");
 
         thread_local std::vector<VkImageResolve> imageResolves;
         imageResolves.clear();

UnleashedRecomp/gpu/rhi/plume_vulkan.h

@@ -315,7 +315,7 @@ namespace plume {
         void copyBuffer(const RenderBuffer *dstBuffer, const RenderBuffer *srcBuffer) override;
         void copyTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) override;
         void resolveTexture(const RenderTexture *dstTexture, const RenderTexture *srcTexture) override;
-        void resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect) override;
+        void resolveTextureRegion(const RenderTexture *dstTexture, uint32_t dstX, uint32_t dstY, const RenderTexture *srcTexture, const RenderRect *srcRect, RenderResolveMode resolveMode) override;
         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;

UnleashedRecomp/gpu/video.cpp

@@ -235,6 +235,9 @@ static bool g_vulkan = false;
 static constexpr bool g_vulkan = true;
 #endif
 
+static constexpr bool g_hardwareResolve = true;
+static constexpr bool g_hardwareDepthResolve = true;
+
 static std::unique_ptr<RenderInterface> g_interface;
 static std::unique_ptr<RenderDevice> g_device;
 
@@ -3024,10 +3027,35 @@ static bool PopulateBarriersForStretchRect(GuestSurface* renderTarget, GuestSurf
     {
         if (surface != nullptr && !surface->destinationTextures.empty())
         {
-            AddBarrier(surface, RenderTextureLayout::SHADER_READ);
+            const bool multiSampling = surface->sampleCount != RenderSampleCount::COUNT_1;
+
+            RenderTextureLayout srcLayout;
+            RenderTextureLayout dstLayout;
+            bool shaderResolve = true;
+
+            if (multiSampling && g_hardwareResolve)
+            {
+                // Hardware depth resolve is only supported on D3D12 when programmable sample positions are available.
+                bool hardwareDepthResolveAvailable = g_hardwareDepthResolve && !g_vulkan && g_capabilities.sampleLocations;
+
+                if (surface->format != RenderFormat::D32_FLOAT || hardwareDepthResolveAvailable)
+                {
+                    srcLayout = RenderTextureLayout::RESOLVE_SOURCE;
+                    dstLayout = RenderTextureLayout::RESOLVE_DEST;
+                    shaderResolve = false;
+                }
+            }
+            
+            if (shaderResolve)
+            {
+                srcLayout = RenderTextureLayout::SHADER_READ;
+                dstLayout = (surface->format == RenderFormat::D32_FLOAT ? RenderTextureLayout::DEPTH_WRITE : RenderTextureLayout::COLOR_WRITE);
+            }
+
+            AddBarrier(surface, srcLayout);
 
             for (const auto texture : surface->destinationTextures)
-                AddBarrier(texture, texture->format == RenderFormat::D32_FLOAT ? RenderTextureLayout::DEPTH_WRITE : RenderTextureLayout::COLOR_WRITE);
+                AddBarrier(texture, dstLayout);
 
             addedAny = true;
         }
@@ -3047,6 +3075,25 @@ static void ExecutePendingStretchRectCommands(GuestSurface* renderTarget, GuestS
             const bool multiSampling = surface->sampleCount != RenderSampleCount::COUNT_1;
 
             for (const auto texture : surface->destinationTextures)
+            {
+                bool shaderResolve = true;
+
+                if (multiSampling && g_hardwareResolve)
+                {
+                    bool hardwareDepthResolveAvailable = g_hardwareDepthResolve && !g_vulkan && g_capabilities.sampleLocations;
+
+                    if (surface->format != RenderFormat::D32_FLOAT || hardwareDepthResolveAvailable)
+                    {
+                        if (surface->format == RenderFormat::D32_FLOAT)
+                            commandList->resolveTextureRegion(texture->texture, 0, 0, surface->texture, nullptr, RenderResolveMode::MIN);
+                        else
+                            commandList->resolveTexture(texture->texture, surface->texture);
+
+                        shaderResolve = false;
+                    }
+                }
+
+                if (shaderResolve)
                 {
                     RenderPipeline* pipeline = nullptr;
 
@@ -3156,6 +3203,7 @@ static void ExecutePendingStretchRectCommands(GuestSurface* renderTarget, GuestS
                         g_dirtyStates.vertexShaderConstants = true; // The push constant call invalidates vertex shader constants.
                         g_dirtyStates.depthBias = true; // Static depth bias in copy pipeline invalidates dynamic depth bias.
                     }
+                }
 
                 texture->sourceSurface = nullptr;