add: basic debug overlay

src/Cafe/CMakeLists.txt

@@ -567,6 +567,7 @@ if(ENABLE_METAL)
             HW/Latte/Renderer/Metal/MetalSamplerCache.h
             HW/Latte/Renderer/Metal/MetalHybridComputePipeline.cpp
             HW/Latte/Renderer/Metal/MetalHybridComputePipeline.h
+            HW/Latte/Renderer/Metal/MetalPerformanceMonitor.h
             HW/Latte/Renderer/Metal/UtilityShaderSource.h
         )
 

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

@@ -93,6 +93,9 @@ public:
             m_freeBufferRanges.push_back(range);
         }
 
+        // Debug
+        m_mtlr->GetPerformanceMonitor().m_bufferAllocatorMemory += m_allocationSize;
+
         // Increase the allocation size for the next buffer
         if (m_allocationSize < 128 * 1024 * 1024)
             m_allocationSize *= 2;

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

@@ -0,0 +1,18 @@
+#pragma once
+
+class MetalPerformanceMonitor
+{
+public:
+    size_t m_bufferAllocatorMemory = 0;
+
+    // Per frame data
+    uint32 m_renderPasses = 0;
+
+    MetalPerformanceMonitor() = default;
+    ~MetalPerformanceMonitor() = default;
+
+    void ResetPerFrameData()
+    {
+        m_renderPasses = 0;
+    }
+};

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

@@ -20,6 +20,9 @@ public:
 
     MTL::RenderPipelineState* GetMeshPipelineState(const LatteFetchShader* fetchShader, const LatteDecompilerShader* vertexShader, const LatteDecompilerShader* geometryShader, const LatteDecompilerShader* pixelShader, class CachedFBOMtl* activeFBO, const LatteContextRegister& lcr, Renderer::INDEX_TYPE hostIndexType);
 
+    // Debug
+    size_t GetPipelineCacheSize() const { return m_pipelineCache.size(); }
+
 private:
     class MetalRenderer* m_mtlr;
 

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

@@ -395,7 +395,13 @@ void MetalRenderer::DeleteFontTextures()
 
 void MetalRenderer::AppendOverlayDebugInfo()
 {
-    // TODO: implement
+    ImGui::Text("--- Metal info ---");
+    ImGui::Text("Render pipeline states  %zu", m_pipelineCache->GetPipelineCacheSize());
+    ImGui::Text("Buffer allocator memory %zuMB", m_performanceMonitor.m_bufferAllocatorMemory / 1024 / 1024);
+
+    ImGui::Text("--- Metal info (per frame) ---");
+    ImGui::Text("Command buffers         %zu", m_commandBuffers.size());
+    ImGui::Text("Render passes           %u", m_performanceMonitor.m_renderPasses);
 }
 
 // TODO: halfZ
@@ -1725,6 +1731,9 @@ void MetalRenderer::SwapBuffer(bool mainWindow)
     {
         debug_printf("skipped present!\n");
     }
+
+    // Debug
+    m_performanceMonitor.ResetPerFrameData();
 }
 
 void MetalRenderer::EnsureImGuiBackend()

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

@@ -3,6 +3,7 @@
 #include "Cafe/HW/Latte/Renderer/Renderer.h"
 
 #include "Cafe/HW/Latte/Renderer/Metal/MetalLayerHandle.h"
+#include "Cafe/HW/Latte/Renderer/Metal/MetalPerformanceMonitor.h"
 
 struct MetalBufferAllocation
 {
@@ -312,6 +313,8 @@ public:
 	}
 
 	// Helpers
+	MetalPerformanceMonitor& GetPerformanceMonitor() { return m_performanceMonitor; }
+
 	MTL::CommandBuffer* GetCurrentCommandBuffer()
     {
         cemu_assert_debug(m_commandBuffers.size() != 0);
@@ -407,6 +410,8 @@ private:
 	MetalLayerHandle m_mainLayer;
 	MetalLayerHandle m_padLayer;
 
+	MetalPerformanceMonitor m_performanceMonitor;
+
 	// Metal objects
 	MTL::Device* m_device;
 	MTL::CommandQueue* m_commandQueue;