diff --git a/lib/gfx/common.cpp b/lib/gfx/common.cpp
index f35d929..de7ce7b 100644
--- a/lib/gfx/common.cpp
+++ b/lib/gfx/common.cpp
@@ -123,6 +123,8 @@ wgpu::BindGroup g_uniformBindGroup;
 
 // for imgui debug
 AuroraStats g_stats{};
+uint32_t g_drawCallCount = 0;
+uint32_t g_mergedDrawCallCount = 0;
 
 using CommandList = std::vector<Command>;
 struct RenderPass {
@@ -214,7 +216,7 @@ gx::DrawData* get_last_draw_command() {
 
 static void push_draw_command(ShaderDrawCommand data) {
   push_command(CommandType::Draw, Command::Data{.draw = data});
-  ++g_stats.drawCallCount;
+  ++g_drawCallCount;
 }
 
 Vec2<uint32_t> get_render_target_size() noexcept {
@@ -669,8 +671,8 @@ bool begin_frame() {
     mapBuffer(g_textureUpload, TextureUploadSize);
   }
 
-  g_stats.drawCallCount = 0;
-  g_stats.mergedDrawCallCount = 0;
+  g_drawCallCount = 0;
+  g_mergedDrawCallCount = 0;
   g_suspendedEfbPass.reset();
 
   g_renderPasses.emplace_back();
@@ -703,6 +705,8 @@ void end_frame(const wgpu::CommandEncoder& cmd) {
   };
   g_stagingBuffers[currentStagingBuffer].Unmap();
   s_mappingState.store(BufferMapState::Unmapped, std::memory_order_release);
+  g_stats.drawCallCount = g_drawCallCount;
+  g_stats.mergedDrawCallCount = g_mergedDrawCallCount;
   g_stats.lastVertSize = writeBuffer(g_verts, g_vertexBuffer, VertexBufferSize, "Vertex");
   g_stats.lastUniformSize = writeBuffer(g_uniforms, g_uniformBuffer, UniformBufferSize, "Uniform");
   g_stats.lastIndexSize = writeBuffer(g_indices, g_indexBuffer, IndexBufferSize, "Index");
@@ -744,6 +748,7 @@ static void expire_cached_bind_groups() {
     return;
   }
 
+  ZoneScoped;
   for (auto it = g_cachedBindGroups.begin(); it != g_cachedBindGroups.end();) {
     if (g_frameIndex - it->second.lastUsedFrame > BindGroupCacheRetainFrames) {
       g_cachedBindGroups.erase(it++);
diff --git a/lib/gfx/common.hpp b/lib/gfx/common.hpp
index 9b3bf19..45608a8 100644
--- a/lib/gfx/common.hpp
+++ b/lib/gfx/common.hpp
@@ -178,6 +178,8 @@ inline constexpr uint64_t StorageBufferSize = 8388608;   // 8mb
 inline constexpr uint64_t TextureUploadSize = 25165824;  // 24mb
 
 extern AuroraStats g_stats;
+extern uint32_t g_drawCallCount;
+extern uint32_t g_mergedDrawCallCount;
 extern wgpu::Buffer g_vertexBuffer;
 extern wgpu::Buffer g_uniformBuffer;
 extern wgpu::Buffer g_indexBuffer;
diff --git a/lib/gx/command_processor.cpp b/lib/gx/command_processor.cpp
index f2a8463..8d051b7 100644
--- a/lib/gx/command_processor.cpp
+++ b/lib/gx/command_processor.cpp
@@ -491,7 +491,6 @@ inline static u32 bp_get(u32 reg, u32 size, u32 shift) { return reg >> shift & (
 
 // BP register handler - decodes BP (RAS/pixel engine) register writes and updates g_gxState
 static void handle_bp(u32 value, bool bigEndian) {
-  ZoneScoped;
   u32 regId = (value >> 24) & 0xFF;
   // Mask off the register ID from the value for field extraction
   // (the regId is stored in bits 24-31, data is in bits 0-23)
@@ -1123,7 +1122,6 @@ static void handle_bp(u32 value, bool bigEndian) {
 
 // CP register handler - decodes CP register writes and updates g_gxState
 static void handle_cp(u8 addr, u32 value, bool bigEndian) {
-  ZoneScoped;
   switch (addr) {
   // VCD low (0x50)
   case 0x50: {
@@ -1165,6 +1163,7 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
   // Matrix index A (0x30)
   case 0x30: {
     g_gxState.currentPnMtx = bp_get(value, 6, 0) / 3;
+    g_gxState.stateDirty = true;
     break;
   }
 
@@ -1260,7 +1259,6 @@ static void handle_cp(u8 addr, u32 value, bool bigEndian) {
 
 // XF register handler - decodes XF (transform unit) register writes and updates g_gxState
 static void handle_xf(const u8* data, u32& pos, u32 size, bool bigEndian) {
-  ZoneScoped;
   CHECK(pos + 4 <= size, "XF header read overrun");
   u32 header = read_u32(data + pos, bigEndian);
   pos += 4;
@@ -1583,7 +1581,7 @@ static void handle_draw(u8 cmd, const u8* data, u32& pos, u32 size, bool bigEndi
       lastDraw->idxRange.size += idxRange.size;
       lastDraw->vtxCount += vtxCount;
       lastDraw->indexCount += numIndices;
-      ++gfx::g_stats.mergedDrawCallCount;
+      ++gfx::g_mergedDrawCallCount;
       return;
     }
   }
diff --git a/lib/rmlui.cpp b/lib/rmlui.cpp
index 44f4904..01c69d0 100644
--- a/lib/rmlui.cpp
+++ b/lib/rmlui.cpp
@@ -5,6 +5,7 @@
 #include <RmlUi/Core.h>
 #include <RmlUi_Backend.h>
 #include <RmlUi_Platform_SDL.h>
+#include <tracy/Tracy.hpp>
 
 #include "window.hpp"
 #include "internal.hpp"
@@ -361,6 +362,7 @@ RenderOutput render(const wgpu::CommandEncoder& encoder, const webgpu::Viewport&
     return {};
   }
 
+  ZoneScoped;
   const Rml::Vector2i dim = dimensions_from_viewport(presentViewport);
   ensure_render_target(dim);
   if (!s_renderTarget.view) {