GPU/HW: Handle redundant texture window updates

Significantly reduces draw count in THPS2.

~150 draw calls per frame to ~70.

src/core/gpu.cpp

@@ -356,7 +356,6 @@ bool GPU::DoState(StateWrapper& sw, GPUTexture** host_texture, bool update_displ
   if (sw.IsReading())
   {
     m_draw_mode.texture_page_changed = true;
-    m_draw_mode.texture_window_changed = true;
     m_drawing_area_changed = true;
     SetClampedDrawingArea();
     UpdateDMARequest();
@@ -1584,8 +1583,6 @@ void GPU::SetTextureWindow(u32 value)
   if (m_draw_mode.texture_window_value == value)
     return;
 
-  FlushRender();
-
   const u8 mask_x = Truncate8(value & UINT32_C(0x1F));
   const u8 mask_y = Truncate8((value >> 5) & UINT32_C(0x1F));
   const u8 offset_x = Truncate8((value >> 10) & UINT32_C(0x1F));
@@ -1597,7 +1594,6 @@ void GPU::SetTextureWindow(u32 value)
   m_draw_mode.texture_window.or_x = (offset_x & mask_x) * 8u;
   m_draw_mode.texture_window.or_y = (offset_y & mask_y) * 8u;
   m_draw_mode.texture_window_value = value;
-  m_draw_mode.texture_window_changed = true;
 }
 
 void GPU::ReadCLUT(u16* dest, GPUTexturePaletteReg reg, bool clut_is_8bit)

src/core/gpu.h

@@ -478,15 +478,10 @@ protected:
     bool texture_x_flip;
     bool texture_y_flip;
     bool texture_page_changed;
-    bool texture_window_changed;
 
     ALWAYS_INLINE bool IsTexturePageChanged() const { return texture_page_changed; }
     ALWAYS_INLINE void SetTexturePageChanged() { texture_page_changed = true; }
     ALWAYS_INLINE void ClearTexturePageChangedFlag() { texture_page_changed = false; }
-
-    ALWAYS_INLINE bool IsTextureWindowChanged() const { return texture_window_changed; }
-    ALWAYS_INLINE void SetTextureWindowChanged() { texture_window_changed = true; }
-    ALWAYS_INLINE void ClearTextureWindowChangedFlag() { texture_window_changed = false; }
   } m_draw_mode = {};
 
   GPUDrawingArea m_drawing_area = {};

src/core/gpu_hw.cpp

@@ -3634,7 +3634,7 @@ void GPU_HW::DispatchRenderCommand()
   {
     if (texture_mode != m_batch.texture_mode || transparency_mode != m_batch.transparency_mode ||
         (transparency_mode == GPUTransparencyMode::BackgroundMinusForeground && !m_allow_shader_blend) ||
-        dithering_enable != m_batch.dithering ||
+        dithering_enable != m_batch.dithering || m_batch_ubo_data.u_texture_window_bits != m_draw_mode.texture_window ||
         (texture_mode == BatchTextureMode::PageTexture && m_batch.texture_cache_key != texture_cache_key))
     {
       FlushRender();
@@ -3684,17 +3684,12 @@ void GPU_HW::DispatchRenderCommand()
     m_batch.dithering = dithering_enable;
     m_batch.texture_cache_key = texture_cache_key;
 
-    if (m_draw_mode.IsTextureWindowChanged())
+    if (m_batch_ubo_data.u_texture_window_bits != m_draw_mode.texture_window)
     {
-      m_draw_mode.ClearTextureWindowChangedFlag();
+      m_batch_ubo_data.u_texture_window_bits = m_draw_mode.texture_window;
-
+      m_texture_window_active = (m_draw_mode.texture_window != GPUTextureWindow{0xFF, 0xFF, 0x00, 0x00});
-      m_batch_ubo_data.u_texture_window[0] = ZeroExtend32(m_draw_mode.texture_window.and_x);
+      GSVector4i::store<true>(&m_batch_ubo_data.u_texture_window[0],
-      m_batch_ubo_data.u_texture_window[1] = ZeroExtend32(m_draw_mode.texture_window.and_y);
+                              GSVector4i::load32(&m_draw_mode.texture_window).u8to32());
-      m_batch_ubo_data.u_texture_window[2] = ZeroExtend32(m_draw_mode.texture_window.or_x);
-      m_batch_ubo_data.u_texture_window[3] = ZeroExtend32(m_draw_mode.texture_window.or_y);
-
-      m_texture_window_active = ((m_draw_mode.texture_window.and_x & m_draw_mode.texture_window.and_y) != 0xFF ||
-                                 ((m_draw_mode.texture_window.or_x | m_draw_mode.texture_window.or_y) != 0));
       m_batch_ubo_dirty = true;
     }
 

src/core/gpu_hw.h

@@ -113,7 +113,7 @@ private:
     void SetUVLimits(u32 min_u, u32 max_u, u32 min_v, u32 max_v);
   };
 
-  struct BatchConfig
+  struct alignas(4) BatchConfig
   {
     BatchTextureMode texture_mode = BatchTextureMode::Disabled;
     GPUTransparencyMode transparency_mode = GPUTransparencyMode::Disabled;
@@ -130,7 +130,7 @@ private:
     BatchRenderMode GetRenderMode() const;
   };
 
-  struct BatchUBOData
+  struct alignas(VECTOR_ALIGNMENT) BatchUBOData
   {
     u32 u_texture_window[4]; // and_x, and_y, or_x, or_y
     float u_src_alpha_factor;
@@ -140,7 +140,7 @@ private:
     float u_resolution_scale;
     float u_rcp_resolution_scale;
     float u_resolution_scale_minus_one;
-    u32 pad;
+    GPUTextureWindow u_texture_window_bits; // not actually used on GPU
   };
 
   struct RendererStats
@@ -302,10 +302,10 @@ private:
 
   u8 m_texpage_dirty = 0;
 
+  bool m_batch_ubo_dirty = true;
   BatchConfig m_batch;
 
   // Changed state
-  bool m_batch_ubo_dirty = true;
   BatchUBOData m_batch_ubo_data = {};
 
   // Bounding box of VRAM area that the GPU has drawn into.

src/core/gpu_types.h

@@ -10,6 +10,7 @@
 #include "common/gsvector.h"
 
 #include <array>
+#include <string>
 
 enum : u32
 {
@@ -229,6 +230,16 @@ struct GPUTextureWindow
   u8 and_y;
   u8 or_x;
   u8 or_y;
+
+  ALWAYS_INLINE bool operator==(const GPUTextureWindow& rhs) const
+  {
+    return (std::memcmp(this, &rhs, sizeof(*this)) == 0);
+  }
+
+  ALWAYS_INLINE bool operator!=(const GPUTextureWindow& rhs) const
+  {
+    return (std::memcmp(this, &rhs, sizeof(*this)) != 0);
+  }
 };
 
 ALWAYS_INLINE static constexpr u32 VRAMPageIndex(u32 px, u32 py)