make occlusion queries accumulate with draws

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

@@ -22,6 +22,13 @@ struct MetalPixelFormatSupport
 	}
 };
 
+// TODO: don't define a new struct for this
+struct MetalQueryRange
+{
+    uint32 begin;
+	uint32 end;
+};
+
 #define MAX_MTL_BUFFERS 31
 // Buffer indices 28-30 are reserved for the helper shaders
 #define GET_MTL_VERTEX_BUFFER_INDEX(index) (MAX_MTL_BUFFERS - index - 4)

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

@@ -1,6 +1,5 @@
 #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)
 {
@@ -13,29 +12,31 @@ bool LatteQueryObjectMtl::getResult(uint64& numSamplesPassed)
     if (!CommandBufferCompleted(m_commandBuffer))
         return false;
 
-    numSamplesPassed = m_mtlr->GetOcclusionQueryResultsPtr()[m_queryIndex];
+    uint64* resultPtr = m_mtlr->GetOcclusionQueryResultsPtr();
+
+    numSamplesPassed = 0;
+    for (uint32 i = m_range.begin; i != m_range.end; i = (i + 1) % MetalRenderer::OCCLUSION_QUERY_POOL_SIZE)
+        numSamplesPassed += resultPtr[i];
 
     return true;
 }
 
 LatteQueryObjectMtl::~LatteQueryObjectMtl()
 {
-    if (m_queryIndex != INVALID_UINT32)
-        m_mtlr->ReleaseOcclusionQueryIndex(m_queryIndex);
-
     if (m_commandBuffer)
         m_commandBuffer->release();
 }
 
 void LatteQueryObjectMtl::begin()
 {
-    m_queryIndex = m_mtlr->GetAvailableOcclusionQueryIndex();
+    m_range.begin = m_mtlr->GetOcclusionQueryIndex();
-    m_mtlr->SetActiveOcclusionQueryIndex(m_queryIndex);
+    m_mtlr->BeginOcclusionQuery();
 }
 
 void LatteQueryObjectMtl::end()
 {
-    m_mtlr->SetActiveOcclusionQueryIndex(INVALID_UINT32);
+    m_range.end = m_mtlr->GetOcclusionQueryIndex();
+    m_mtlr->EndOcclusionQuery();
     if (m_mtlr->IsCommandBufferActive())
     {
         m_commandBuffer = m_mtlr->GetCurrentCommandBuffer()->retain();

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

@@ -14,10 +14,15 @@ public:
 	void begin() override;
 	void end() override;
 
+	void GrowRange()
+	{
+	    m_range.end++;
+	}
+
 private:
 	class MetalRenderer* m_mtlr;
 
-	uint32 m_queryIndex = INVALID_UINT32;
+	MetalQueryRange m_range = {INVALID_UINT32, INVALID_UINT32};
 	// TODO: make this a list of command buffers
 	MTL::CommandBuffer* m_commandBuffer = nullptr;
 };

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

@@ -31,7 +31,6 @@
 #include "imgui/imgui_impl_metal.h"
 
 #define DEFAULT_COMMIT_TRESHOLD 196
-#define OCCLUSION_QUERY_POOL_SIZE 1024
 
 extern bool hasValidFramebufferAttached;
 
@@ -98,10 +97,6 @@ MetalRenderer::MetalRenderer()
 #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++)
     {
@@ -1115,11 +1110,10 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 	}
 
 	// Visibility result mode
-	if (m_occlusionQuery.m_activeIndex != encoderState.m_visibilityResultOffset)
+	if (m_occlusionQuery.m_active)
 	{
-	    auto mode = (m_occlusionQuery.m_activeIndex == INVALID_UINT32 ? MTL::VisibilityResultModeDisabled : MTL::VisibilityResultModeCounting);
+	    auto mode = (m_occlusionQuery.m_currentIndex == INVALID_UINT32 ? MTL::VisibilityResultModeDisabled : MTL::VisibilityResultModeCounting);
-	    renderCommandEncoder->setVisibilityResultMode(mode, m_occlusionQuery.m_activeIndex * sizeof(uint64));
+	    renderCommandEncoder->setVisibilityResultMode(mode, m_occlusionQuery.m_currentIndex * sizeof(uint64));
-		encoderState.m_visibilityResultOffset = m_occlusionQuery.m_activeIndex;
 	}
 
 	// todo - how does culling behave with rects?
@@ -1302,6 +1296,11 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 
 	m_state.m_isFirstDrawInRenderPass = false;
 
+	// Occlusion queries
+	if (m_occlusionQuery.m_active)
+	    m_occlusionQuery.m_currentIndex = (m_occlusionQuery.m_currentIndex + 1) % OCCLUSION_QUERY_POOL_SIZE;
+
+	// Streamout
 	LatteStreamout_FinishDrawcall(false);
 
 	// Debug

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

@@ -104,7 +104,6 @@ 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;
@@ -170,7 +169,8 @@ enum class MetalEncoderType
 class MetalRenderer : public Renderer
 {
 public:
-    static const inline int TEXTURE_READBACK_SIZE = 32 * 1024 * 1024; // 32 MB
+    static constexpr uint32 OCCLUSION_QUERY_POOL_SIZE = 1024;
+    static constexpr uint32 TEXTURE_READBACK_SIZE = 32 * 1024 * 1024; // 32 MB
 
     MetalRenderer();
 	~MetalRenderer() override;
@@ -428,28 +428,19 @@ public:
         return m_occlusionQuery.m_resultsPtr;
     }
 
-    uint32 GetAvailableOcclusionQueryIndex()
+    uint32 GetOcclusionQueryIndex()
     {
-        if (m_occlusionQuery.m_availableIndices.empty())
+        return m_occlusionQuery.m_currentIndex;
-        {
-            cemuLog_log(LogType::Force, "No occlusion query index available");
-            return 0;
     }
 
-        uint32 queryIndex = m_occlusionQuery.m_availableIndices.back();
+    void BeginOcclusionQuery()
-        m_occlusionQuery.m_availableIndices.pop_back();
+    {
-
+        m_occlusionQuery.m_active = true;
-        return queryIndex;
     }
 
-    void ReleaseOcclusionQueryIndex(uint32 queryIndex)
+    void EndOcclusionQuery()
     {
-        m_occlusionQuery.m_availableIndices.push_back(queryIndex);
+        m_occlusionQuery.m_active = false;
-    }
-
-    void SetActiveOcclusionQueryIndex(uint32 queryIndex)
-    {
-        m_occlusionQuery.m_activeIndex = queryIndex;
     }
 
 private:
@@ -505,8 +496,8 @@ private:
 	{
     	MTL::Buffer* m_resultBuffer;
     	uint64* m_resultsPtr;
-    	std::vector<uint32> m_availableIndices;
+    	uint32 m_currentIndex = 0;
-        uint32 m_activeIndex = INVALID_UINT32;
+        bool m_active = false;
 	} m_occlusionQuery;
 
 	// Active objects