implement occlusion queries

src/Cafe/HW/Latte/Renderer/Metal/CachedFBOMtl.cpp

@@ -1,9 +1,9 @@
 #include "Cafe/HW/Latte/Renderer/Metal/CachedFBOMtl.h"
 #include "Cafe/HW/Latte/Renderer/Metal/LatteTextureViewMtl.h"
-#include "HW/Latte/Renderer/Metal/LatteToMtl.h"
+#include "Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h"
-#include "Metal/MTLRenderPass.hpp"
+#include "Cafe/HW/Latte/Renderer/Metal/LatteToMtl.h"
 
-void CachedFBOMtl::CreateRenderPass()
+CachedFBOMtl::CachedFBOMtl(class MetalRenderer* metalRenderer, uint64 key) : LatteCachedFBO(key)
 {
 	m_renderPassDescriptor = MTL::RenderPassDescriptor::alloc()->init();
 
@@ -39,6 +39,9 @@ void CachedFBOMtl::CreateRenderPass()
             stencilAttachment->setStoreAction(MTL::StoreActionStore);
 		}
 	}
+
+	// Visibility buffer
+	m_renderPassDescriptor->setVisibilityResultBuffer(metalRenderer->GetOcclusionQueryResultBuffer());
 }
 
 CachedFBOMtl::~CachedFBOMtl()

src/Cafe/HW/Latte/Renderer/Metal/CachedFBOMtl.h

@@ -8,10 +8,7 @@
 class CachedFBOMtl : public LatteCachedFBO
 {
 public:
-	CachedFBOMtl(uint64 key) : LatteCachedFBO(key)
+	CachedFBOMtl(class MetalRenderer* metalRenderer, uint64 key);
-	{
-		CreateRenderPass();
-	}
 
 	~CachedFBOMtl();
 
@@ -22,6 +19,4 @@ public:
 
 private:
     MTL::RenderPassDescriptor* m_renderPassDescriptor = nullptr;
-
-    void CreateRenderPass();
 };

src/Cafe/HW/Latte/Renderer/Metal/MetalQuery.cpp

@@ -1,17 +1,25 @@
 #include "Cafe/HW/Latte/Renderer/Metal/MetalQuery.h"
+#include "Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h"
+#include "HW/Latte/Renderer/Metal/MetalCommon.h"
 
 bool LatteQueryObjectMtl::getResult(uint64& numSamplesPassed)
 {
-    cemuLog_log(LogType::MetalLogging, "LatteQueryObjectMtl::getResult: occlusion queries are not yet supported on Metal");
+    if (!m_mtlr->CommandBufferCompleted(m_commandBuffer))
+        return false;
+
+    numSamplesPassed = m_mtlr->GetOcclusionQueryResultsPtr()[m_queryIndex];
+
     return true;
 }
 
 void LatteQueryObjectMtl::begin()
 {
-    cemuLog_log(LogType::MetalLogging, "LatteQueryObjectMtl::begin: occlusion queries are not yet supported on Metal");
+    m_queryIndex = m_mtlr->GetAvailableOcclusionQueryIndex();
+    m_mtlr->SetActiveOcclusionQueryIndex(m_queryIndex);
 }
 
 void LatteQueryObjectMtl::end()
 {
-    cemuLog_log(LogType::MetalLogging, "LatteQueryObjectMtl::end: occlusion queries are not yet supported on Metal");
+    m_mtlr->SetActiveOcclusionQueryIndex(INVALID_UINT32);
+    // TODO: request soon submit of the command buffer
 }

src/Cafe/HW/Latte/Renderer/Metal/MetalQuery.h

@@ -4,7 +4,6 @@
 
 #include "Cafe/HW/Latte/Renderer/Metal/MetalCommon.h"
 
-// HACK: Dummy occlusion query object
 class LatteQueryObjectMtl : public LatteQueryObject
 {
 public:
@@ -16,4 +15,8 @@ public:
 
 private:
 	class MetalRenderer* m_mtlr;
+
+	uint32 m_queryIndex;
+	MTL::CommandBuffer* m_commandBuffer;
+	uint64 m_acccumulatedSum;
 };

src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.cpp

@@ -21,12 +21,14 @@
 #include "HW/Latte/Renderer/Metal/MetalCommon.h"
 #include "HW/Latte/Renderer/Metal/MetalLayerHandle.h"
 #include "HW/Latte/Renderer/Renderer.h"
+#include "Metal/MTLRenderCommandEncoder.hpp"
 
 #define IMGUI_IMPL_METAL_CPP
 #include "imgui/imgui_extension.h"
 #include "imgui/imgui_impl_metal.h"
 
 #define COMMIT_TRESHOLD 256
+#define OCCLUSION_QUERY_POOL_SIZE 1024
 
 extern bool hasValidFramebufferAttached;
 
@@ -94,6 +96,17 @@ MetalRenderer::MetalRenderer()
     m_xfbRingBuffer->setLabel(GetLabel("Transform feedback buffer", m_xfbRingBuffer));
 #endif
 
+    // Occlusion queries
+    m_occlusionQuery.m_resultBuffer = m_device->newBuffer(OCCLUSION_QUERY_POOL_SIZE * sizeof(uint64), MTL::ResourceStorageModeShared);
+#ifdef CEMU_DEBUG_ASSERT
+    m_occlusionQuery.m_resultBuffer->setLabel(GetLabel("Occlusion query result buffer", m_occlusionQuery.m_resultBuffer));
+#endif
+    m_occlusionQuery.m_resultsPtr = (uint64*)m_occlusionQuery.m_resultBuffer->contents();
+
+    m_occlusionQuery.m_availableIndices.reserve(OCCLUSION_QUERY_POOL_SIZE);
+    for (uint32 i = 0; i < OCCLUSION_QUERY_POOL_SIZE; i++)
+        m_occlusionQuery.m_availableIndices.push_back(i);
+
     // Initialize state
     for (uint32 i = 0; i < METAL_SHADER_TYPE_TOTAL; i++)
     {
@@ -467,7 +480,7 @@ void MetalRenderer::renderTarget_setScissor(sint32 scissorX, sint32 scissorY, si
 
 LatteCachedFBO* MetalRenderer::rendertarget_createCachedFBO(uint64 key)
 {
-	return new CachedFBOMtl(key);
+	return new CachedFBOMtl(this, key);
 }
 
 void MetalRenderer::rendertarget_deleteCachedFBO(LatteCachedFBO* cfbo)
@@ -1042,6 +1055,14 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
         encoderState.m_depthClipEnable = zClipEnable;
 	}
 
+	// Visibility result mode
+	if (m_occlusionQuery.m_activeIndex != encoderState.m_visibilityResultOffset)
+	{
+	    auto mode = (m_occlusionQuery.m_activeIndex == INVALID_UINT32 ? MTL::VisibilityResultModeDisabled : MTL::VisibilityResultModeCounting);
+	    renderCommandEncoder->setVisibilityResultMode(mode, m_occlusionQuery.m_activeIndex);
+		encoderState.m_visibilityResultOffset = m_occlusionQuery.m_activeIndex;
+	}
+
 	// todo - how does culling behave with rects?
 	// right now we just assume that their winding is always CW
 	if (isPrimitiveRect)
@@ -1284,21 +1305,20 @@ void MetalRenderer::indexData_uploadIndexMemory(uint32 bufferIndex, uint32 offse
 }
 
 LatteQueryObject* MetalRenderer::occlusionQuery_create() {
-    cemuLog_log(LogType::MetalLogging, "MetalRenderer::occlusionQuery_create: Occlusion queries are not yet supported on Metal");
-
 	return new LatteQueryObjectMtl(this);
 }
 
 void MetalRenderer::occlusionQuery_destroy(LatteQueryObject* queryObj) {
-    cemuLog_log(LogType::MetalLogging, "MetalRenderer::occlusionQuery_destroy: occlusion queries are not yet supported on Metal");
+    // TODO: do something?
 }
 
 void MetalRenderer::occlusionQuery_flush() {
-    cemuLog_log(LogType::MetalLogging, "MetalRenderer::occlusionQuery_flush: occlusion queries are not yet supported on Metal");
+    // TODO: implement
+    debug_printf("Occlusion query flush is not implemented\n");
 }
 
 void MetalRenderer::occlusionQuery_updateState() {
-    cemuLog_log(LogType::MetalLogging, "MetalRenderer::occlusionQuery_updateState: occlusion queries are not yet supported on Metal");
+    // TODO
 }
 
 void MetalRenderer::SetBuffer(MTL::RenderCommandEncoder* renderCommandEncoder, MetalShaderType shaderType, MTL::Buffer* buffer, size_t offset, uint32 index)
@@ -1574,6 +1594,8 @@ void MetalRenderer::CommitCommandBuffer()
 
             m_memoryManager->GetTemporaryBufferAllocator().SetActiveCommandBuffer(nullptr);
 
+            m_occlusionQuery.m_availableIndices.insert(m_occlusionQuery.m_availableIndices.end(), m_occlusionQuery.m_crntCmdBuffIndices.begin(), m_occlusionQuery.m_crntCmdBuffIndices.end());
+
             // Debug
             //m_commandQueue->insertDebugCaptureBoundary();
         }

src/Cafe/HW/Latte/Renderer/Metal/MetalRenderer.h

@@ -99,6 +99,7 @@ struct MetalEncoderState
    	uint32 m_depthSlope = 0;
    	uint32 m_depthClamp = 0;
     bool m_depthClipEnable = true;
+    uint32 m_visibilityResultOffset = INVALID_UINT32;
     struct {
         MTL::Buffer* m_buffer;
         size_t m_offset;
@@ -376,6 +377,36 @@ public:
         return m_readbackBuffer;
     }
 
+    MTL::Buffer* GetOcclusionQueryResultBuffer() const
+    {
+        return m_occlusionQuery.m_resultBuffer;
+    }
+
+    uint64* GetOcclusionQueryResultsPtr()
+    {
+        return m_occlusionQuery.m_resultsPtr;
+    }
+
+    uint32 GetAvailableOcclusionQueryIndex()
+    {
+        if (m_occlusionQuery.m_availableIndices.empty())
+        {
+            cemuLog_log(LogType::Force, "No occlusion query index available");
+            return 0;
+        }
+
+        uint32 queryIndex = m_occlusionQuery.m_availableIndices.back();
+        m_occlusionQuery.m_availableIndices.pop_back();
+        m_occlusionQuery.m_crntCmdBuffIndices.push_back(queryIndex);
+
+        return queryIndex;
+    }
+
+    void SetActiveOcclusionQueryIndex(uint32 queryIndex)
+    {
+        m_occlusionQuery.m_activeIndex = queryIndex;
+    }
+
 private:
 	MetalLayerHandle m_mainLayer;
 	MetalLayerHandle m_padLayer;
@@ -423,6 +454,16 @@ private:
 	// Transform feedback
 	MTL::Buffer* m_xfbRingBuffer;
 
+	// Occlusion queries
+	struct
+	{
+    	MTL::Buffer* m_resultBuffer;
+    	uint64* m_resultsPtr;
+    	std::vector<uint32> m_availableIndices;
+    	std::vector<uint32> m_crntCmdBuffIndices;
+        uint32 m_activeIndex = INVALID_UINT32;
+	} m_occlusionQuery;
+
 	// Active objects
 	std::vector<MetalCommandBuffer> m_commandBuffers;
 	MetalEncoderType m_encoderType = MetalEncoderType::None;