GPU: Precompute texture window expression [SAVEVERSION+]

src/core/gpu.cpp

@@ -95,6 +95,7 @@ void GPU::SoftReset()
   m_fifo.Clear();
   m_blit_buffer.clear();
   m_blit_remaining_words = 0;
+  m_draw_mode.texture_window_value = 0xFFFFFFFFu;
   SetDrawMode(0);
   SetTexturePalette(0);
   SetTextureWindow(0);
@@ -121,10 +122,10 @@ bool GPU::DoState(StateWrapper& sw)
   sw.Do(&m_draw_mode.texture_page_y);
   sw.Do(&m_draw_mode.texture_palette_x);
   sw.Do(&m_draw_mode.texture_palette_y);
-  sw.Do(&m_draw_mode.texture_window_mask_x);
-  sw.Do(&m_draw_mode.texture_window_mask_y);
-  sw.Do(&m_draw_mode.texture_window_offset_x);
-  sw.Do(&m_draw_mode.texture_window_offset_y);
+  sw.Do(&m_draw_mode.texture_window_and_x);
+  sw.Do(&m_draw_mode.texture_window_and_y);
+  sw.Do(&m_draw_mode.texture_window_or_x);
+  sw.Do(&m_draw_mode.texture_window_or_y);
   sw.Do(&m_draw_mode.texture_x_flip);
   sw.Do(&m_draw_mode.texture_y_flip);
 
@@ -1316,10 +1317,16 @@ void GPU::SetTextureWindow(u32 value)
 
   FlushRender();
 
-  m_draw_mode.texture_window_mask_x = value & UINT32_C(0x1F);
-  m_draw_mode.texture_window_mask_y = (value >> 5) & UINT32_C(0x1F);
-  m_draw_mode.texture_window_offset_x = (value >> 10) & UINT32_C(0x1F);
-  m_draw_mode.texture_window_offset_y = (value >> 15) & UINT32_C(0x1F);
+  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));
+  const u8 offset_y = Truncate8((value >> 15) & UINT32_C(0x1F));
+  Log_DebugPrintf("Set texture window %02X %02X %02X %02X", mask_x, mask_y, offset_x, offset_y);
+
+  m_draw_mode.texture_window_and_x = ~(mask_x * 8);
+  m_draw_mode.texture_window_and_y = ~(mask_y * 8);
+  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;
 }

src/core/gpu.h

@@ -306,7 +306,7 @@ protected:
   };
 
   // Sprites/rectangles should be clipped to 12 bits before drawing.
-  static constexpr s32 TruncateVertexPosition(s32 x) { return SignExtendN<11, s32>(x); }
+  static constexpr s32 TruncateVertexPosition(s32 x) { return SignExtendN<12, s32>(x); }
 
   struct NativeVertex
   {
@@ -581,10 +581,10 @@ protected:
     u32 texture_page_y;
     u32 texture_palette_x;
     u32 texture_palette_y;
-    u8 texture_window_mask_x;   // in 8 pixel steps
-    u8 texture_window_mask_y;   // in 8 pixel steps
-    u8 texture_window_offset_x; // in 8 pixel steps
-    u8 texture_window_offset_y; // in 8 pixel steps
+    u8 texture_window_and_x;
+    u8 texture_window_and_y;
+    u8 texture_window_or_x;
+    u8 texture_window_or_y;
     bool texture_x_flip;
     bool texture_y_flip;
     bool texture_page_changed;

src/core/gpu_commands.cpp

@@ -228,10 +228,6 @@ bool GPU::HandleSetTextureWindowCommand()
 {
   const u32 param = FifoPop() & 0x00FFFFFFu;
   SetTextureWindow(param);
-  Log_DebugPrintf("Set texture window %02X %02X %02X %02X", m_draw_mode.texture_window_mask_x,
-                  m_draw_mode.texture_window_mask_y, m_draw_mode.texture_window_offset_x,
-                  m_draw_mode.texture_window_offset_y);
-
   AddCommandTicks(1);
   EndCommand();
   return true;

src/core/gpu_hw.cpp

@@ -948,10 +948,10 @@ void GPU_HW::DispatchRenderCommand()
   {
     m_draw_mode.ClearTextureWindowChangedFlag();
 
-    m_batch_ubo_data.u_texture_window_mask[0] = ZeroExtend32(m_draw_mode.texture_window_mask_x);
-    m_batch_ubo_data.u_texture_window_mask[1] = ZeroExtend32(m_draw_mode.texture_window_mask_y);
-    m_batch_ubo_data.u_texture_window_offset[0] = ZeroExtend32(m_draw_mode.texture_window_offset_x);
-    m_batch_ubo_data.u_texture_window_offset[1] = ZeroExtend32(m_draw_mode.texture_window_offset_y);
+    m_batch_ubo_data.u_texture_window_and[0] = ZeroExtend32(m_draw_mode.texture_window_and_x);
+    m_batch_ubo_data.u_texture_window_and[1] = ZeroExtend32(m_draw_mode.texture_window_and_y);
+    m_batch_ubo_data.u_texture_window_or[0] = ZeroExtend32(m_draw_mode.texture_window_or_x);
+    m_batch_ubo_data.u_texture_window_or[1] = ZeroExtend32(m_draw_mode.texture_window_or_y);
     m_batch_ubo_dirty = true;
   }
 

src/core/gpu_hw.h

@@ -119,8 +119,8 @@ protected:
 
   struct BatchUBOData
   {
-    u32 u_texture_window_mask[2];
-    u32 u_texture_window_offset[2];
+    u32 u_texture_window_and[2];
+    u32 u_texture_window_or[2];
     float u_src_alpha_factor;
     float u_dst_alpha_factor;
     u32 u_interlaced_displayed_field;

src/core/gpu_hw_shadergen.cpp

@@ -68,7 +68,7 @@ float4 RGBA5551ToRGBA8(uint v)
 void GPU_HW_ShaderGen::WriteBatchUniformBuffer(std::stringstream& ss)
 {
   DeclareUniformBuffer(ss,
-                       {"uint2 u_texture_window_mask", "uint2 u_texture_window_offset", "float u_src_alpha_factor",
+                       {"uint2 u_texture_window_and", "uint2 u_texture_window_or", "float u_src_alpha_factor",
                         "float u_dst_alpha_factor", "uint u_interlaced_displayed_field",
                         "bool u_set_mask_while_drawing"},
                        false);
@@ -700,16 +700,16 @@ CONSTANT float4 TRANSPARENT_PIXEL_COLOR = float4(0.0, 0.0, 0.0, 0.0);
 
 uint2 ApplyTextureWindow(uint2 coords)
 {
-  uint x = (uint(coords.x) & ~(u_texture_window_mask.x * 8u)) | ((u_texture_window_offset.x & u_texture_window_mask.x) * 8u);
-  uint y = (uint(coords.y) & ~(u_texture_window_mask.y * 8u)) | ((u_texture_window_offset.y & u_texture_window_mask.y) * 8u);
+  uint x = (uint(coords.x) & u_texture_window_and.x) | u_texture_window_or.x;
+  uint y = (uint(coords.y) & u_texture_window_and.y) | u_texture_window_or.y;
   return uint2(x, y);
 }
 
 uint2 ApplyUpscaledTextureWindow(uint2 coords)
 {
-  uint x = (uint(coords.x) & ~(u_texture_window_mask.x * 8u * RESOLUTION_SCALE)) | ((u_texture_window_offset.x & u_texture_window_mask.x) * 8u * RESOLUTION_SCALE);
-  uint y = (uint(coords.y) & ~(u_texture_window_mask.y * 8u * RESOLUTION_SCALE)) | ((u_texture_window_offset.y & u_texture_window_mask.y) * 8u * RESOLUTION_SCALE);
-  return uint2(x, y);
+  uint2 native_coords = coords / uint2(RESOLUTION_SCALE, RESOLUTION_SCALE);
+  uint2 coords_offset = coords % uint2(RESOLUTION_SCALE, RESOLUTION_SCALE);
+  return (ApplyTextureWindow(native_coords) * uint2(RESOLUTION_SCALE, RESOLUTION_SCALE)) + coords_offset;
 }
 
 uint2 FloatToIntegerCoords(float2 coords)

src/core/gpu_sw.cpp

@@ -602,10 +602,8 @@ void GPU_SW::ShadePixel(u32 x, u32 y, u8 color_r, u8 color_g, u8 color_b, u8 tex
   {
     // Apply texture window
     // TODO: Precompute the second half
-    texcoord_x = (texcoord_x & ~(m_draw_mode.texture_window_mask_x * 8u)) |
-                 ((m_draw_mode.texture_window_offset_x & m_draw_mode.texture_window_mask_x) * 8u);
-    texcoord_y = (texcoord_y & ~(m_draw_mode.texture_window_mask_y * 8u)) |
-                 ((m_draw_mode.texture_window_offset_y & m_draw_mode.texture_window_mask_y) * 8u);
+    texcoord_x = (texcoord_x & m_draw_mode.texture_window_and_x) | m_draw_mode.texture_window_or_x;
+    texcoord_y = (texcoord_y & m_draw_mode.texture_window_and_y) | m_draw_mode.texture_window_or_y;
 
     VRAMPixel texture_color;
     switch (m_draw_mode.GetTextureMode())

src/core/save_state_version.h

@@ -2,7 +2,7 @@
 #include "types.h"
 
 static constexpr u32 SAVE_STATE_MAGIC = 0x43435544;
-static constexpr u32 SAVE_STATE_VERSION = 40;
+static constexpr u32 SAVE_STATE_VERSION = 41;
 
 #pragma pack(push, 4)
 struct SAVE_STATE_HEADER