improve command buffer sync

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

@@ -38,12 +38,12 @@ bool LatteTextureReadbackInfoMtl::IsFinished()
     if (m_mtlr->GetCurrentCommandBuffer() == m_commandBuffer)
         m_mtlr->CommitCommandBuffer();
 
-    return m_mtlr->CommandBufferCompleted(m_commandBuffer);
+    return CommandBufferCompleted(m_commandBuffer);
 }
 
 void LatteTextureReadbackInfoMtl::ForceFinish()
 {
-    m_mtlr->WaitForCommandBufferCompletion(m_commandBuffer);
+    m_commandBuffer->waitUntilCompleted();
 }
 
 uint8* LatteTextureReadbackInfoMtl::GetData()

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

@@ -255,7 +255,7 @@ public:
             auto& buffer = m_buffers[i];
             for (uint32_t j = 0; j < buffer.m_data.m_commandBuffers.size(); j++)
             {
-                if (m_mtlr->CommandBufferCompleted(buffer.m_data.m_commandBuffers[j]))
+                if (CommandBufferCompleted(buffer.m_data.m_commandBuffers[j]))
                 {
                     if (buffer.m_data.m_commandBuffers.size() == 1)
                     {

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

@@ -69,3 +69,9 @@ inline bool IsValidDepthTextureType(Latte::E_DIM dim)
 {
     return (dim == Latte::E_DIM::DIM_2D || dim == Latte::E_DIM::DIM_2D_MSAA || dim == Latte::E_DIM::DIM_2D_ARRAY || dim == Latte::E_DIM::DIM_2D_ARRAY_MSAA || dim == Latte::E_DIM::DIM_CUBEMAP);
 }
+
+inline bool CommandBufferCompleted(MTL::CommandBuffer* commandBuffer)
+{
+    auto status = commandBuffer->status();
+    return (status == MTL::CommandBufferStatusCompleted || status == MTL::CommandBufferStatusError);
+}

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

@@ -4,14 +4,21 @@
 
 bool LatteQueryObjectMtl::getResult(uint64& numSamplesPassed)
 {
-    if (!m_mtlr->CommandBufferCompleted(m_commandBuffer))
+    if (!CommandBufferCompleted(m_commandBuffer))
         return false;
 
     numSamplesPassed = m_mtlr->GetOcclusionQueryResultsPtr()[m_queryIndex];
+    printf("Num samples: %llu\n", numSamplesPassed);
 
     return true;
 }
 
+LatteQueryObjectMtl::~LatteQueryObjectMtl()
+{
+    if (m_queryIndex != INVALID_UINT32)
+        m_mtlr->ReleaseOcclusionQueryIndex(m_queryIndex);
+}
+
 void LatteQueryObjectMtl::begin()
 {
     m_queryIndex = m_mtlr->GetAvailableOcclusionQueryIndex();
@@ -21,5 +28,7 @@ void LatteQueryObjectMtl::begin()
 void LatteQueryObjectMtl::end()
 {
     m_mtlr->SetActiveOcclusionQueryIndex(INVALID_UINT32);
-    // TODO: request soon submit of the command buffer
+    m_commandBuffer = m_mtlr->GetCurrentCommandBuffer();
+    // TODO: request soon submit instead?
+    m_mtlr->CommitCommandBuffer();
 }

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

@@ -8,6 +8,7 @@ class LatteQueryObjectMtl : public LatteQueryObject
 {
 public:
 	LatteQueryObjectMtl(class MetalRenderer* mtlRenderer) : m_mtlr{mtlRenderer} {}
+	~LatteQueryObjectMtl();
 
 	bool getResult(uint64& numSamplesPassed) override;
 	void begin() override;
@@ -16,7 +17,7 @@ public:
 private:
 	class MetalRenderer* m_mtlr;
 
-	uint32 m_queryIndex;
+	uint32 m_queryIndex = INVALID_UINT32;
+	// TODO: make this a list of command buffers
 	MTL::CommandBuffer* m_commandBuffer;
-	uint64 m_acccumulatedSum;
 };

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

@@ -334,7 +334,7 @@ void MetalRenderer::Flush(bool waitIdle)
         {
             cemu_assert_debug(commandBuffer.m_commited);
 
-            WaitForCommandBufferCompletion(commandBuffer.m_commandBuffer);
+            commandBuffer.m_commandBuffer->waitUntilCompleted();
         }
     }
 }
@@ -1059,7 +1059,7 @@ void MetalRenderer::draw_execute(uint32 baseVertex, uint32 baseInstance, uint32
 	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);
+	    renderCommandEncoder->setVisibilityResultMode(mode, m_occlusionQuery.m_activeIndex * sizeof(uint64));
 		encoderState.m_visibilityResultOffset = m_occlusionQuery.m_activeIndex;
 	}
 
@@ -1309,16 +1309,16 @@ LatteQueryObject* MetalRenderer::occlusionQuery_create() {
 }
 
 void MetalRenderer::occlusionQuery_destroy(LatteQueryObject* queryObj) {
-    // TODO: do something?
+    auto queryObjMtl = static_cast<LatteQueryObjectMtl*>(queryObj);
+    delete queryObjMtl;
 }
 
 void MetalRenderer::occlusionQuery_flush() {
     // TODO: implement
-    debug_printf("Occlusion query flush is not implemented\n");
 }
 
 void MetalRenderer::occlusionQuery_updateState() {
-    // TODO
+    // TODO: implement
 }
 
 void MetalRenderer::SetBuffer(MTL::RenderCommandEncoder* renderCommandEncoder, MetalShaderType shaderType, MTL::Buffer* buffer, size_t offset, uint32 index)
@@ -1420,17 +1420,6 @@ MTL::CommandBuffer* MetalRenderer::GetCommandBuffer()
 	}
 }
 
-bool MetalRenderer::CommandBufferCompleted(MTL::CommandBuffer* commandBuffer)
-{
-    auto status = commandBuffer->status();
-    return (status == MTL::CommandBufferStatusCompleted || status == MTL::CommandBufferStatusError);
-}
-
-void MetalRenderer::WaitForCommandBufferCompletion(MTL::CommandBuffer* commandBuffer)
-{
-    commandBuffer->waitUntilCompleted();
-}
-
 MTL::RenderCommandEncoder* MetalRenderer::GetTemporaryRenderCommandEncoder(MTL::RenderPassDescriptor* renderPassDescriptor)
 {
     EndEncoding();
@@ -1594,8 +1583,6 @@ 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

@@ -323,8 +323,6 @@ public:
     void SetSamplerState(MTL::RenderCommandEncoder* renderCommandEncoder, MetalShaderType shaderType, MTL::SamplerState* samplerState, uint32 index);
 
 	MTL::CommandBuffer* GetCommandBuffer();
-	bool CommandBufferCompleted(MTL::CommandBuffer* commandBuffer);
-	void WaitForCommandBufferCompletion(MTL::CommandBuffer* commandBuffer);
 	MTL::RenderCommandEncoder* GetTemporaryRenderCommandEncoder(MTL::RenderPassDescriptor* renderPassDescriptor);
 	MTL::RenderCommandEncoder* GetRenderCommandEncoder(bool forceRecreate = false);
     MTL::ComputeCommandEncoder* GetComputeCommandEncoder();
@@ -397,11 +395,15 @@ public:
 
         uint32 queryIndex = m_occlusionQuery.m_availableIndices.back();
         m_occlusionQuery.m_availableIndices.pop_back();
-        m_occlusionQuery.m_crntCmdBuffIndices.push_back(queryIndex);
 
         return queryIndex;
     }
 
+    void ReleaseOcclusionQueryIndex(uint32 queryIndex)
+    {
+        m_occlusionQuery.m_availableIndices.push_back(queryIndex);
+    }
+
     void SetActiveOcclusionQueryIndex(uint32 queryIndex)
     {
         m_occlusionQuery.m_activeIndex = queryIndex;
@@ -460,7 +462,6 @@ private:
     	MTL::Buffer* m_resultBuffer;
     	uint64* m_resultsPtr;
     	std::vector<uint32> m_availableIndices;
-    	std::vector<uint32> m_crntCmdBuffIndices;
         uint32 m_activeIndex = INVALID_UINT32;
 	} m_occlusionQuery;