BAHAHAHHHH

HARK HARK HARK

8 files changed, 219 insertions, 80 deletions

diff --git a/src/GPU.cpp b/src/GPU.cpp
index 1d073f5..8142e47 100644
--- a/src/GPU.cpp
+++ b/src/GPU.cpp
@@ -254,7 +254,7 @@ void SetDisplaySettings(int topscale, int bottomscale, bool accel)
         FBScale[0] = accel ? 0 : topscale;
 
         int fbsize;
-        if (accel) fbsize = 256*3 * 192;
+        if (accel) fbsize = (256*3 + 1) * 192;
         else       fbsize = (256 * 192) << (FBScale[0] * 2);
         if (Framebuffer[0][0]) delete[] Framebuffer[0][0];
         if (Framebuffer[1][0]) delete[] Framebuffer[1][0];
@@ -283,7 +283,7 @@ void SetDisplaySettings(int topscale, int bottomscale, bool accel)
         FBScale[1] = accel ? 0 : bottomscale;
 
         int fbsize;
-        if (accel) fbsize = 256*3 * 192;
+        if (accel) fbsize = (256*3 + 1) * 192;
         else       fbsize = (256 * 192) << (FBScale[1] * 2);
         if (Framebuffer[0][1]) delete[] Framebuffer[0][1];
         if (Framebuffer[1][1]) delete[] Framebuffer[1][1];
diff --git a/src/GPU2D.cpp b/src/GPU2D.cpp
index 34c17ff..f80eb85 100644
--- a/src/GPU2D.cpp
+++ b/src/GPU2D.cpp
@@ -626,7 +626,7 @@ u32 GPU2D::ColorBrightnessDown(u32 val, u32 factor)
 
 void GPU2D::DrawScanline(u32 line)
 {
-    int stride = Accelerated ? (256*3) : LineStride;
+    int stride = Accelerated ? (256*3 + 1) : LineStride;
     u32* dst = &Framebuffer[stride * line];
 
     int n3dline = line;
@@ -787,7 +787,17 @@ void GPU2D::DrawScanline(u32 line)
             DoCapture(line, capwidth);
     }
 
-    if (Accelerated) return;
+    if (Accelerated)
+    {
+        u32 ctl = (BlendCnt & 0x3FFF);
+        ctl    |= ((DispCnt & 0x30000) >> 2);
+        ctl    |= (EVA << 16);
+        ctl    |= (EVB << 21);
+        ctl    |= (EVY << 26);
+
+        dst[256*3] = ctl;
+        return;
+    }
 
     // master brightness
     if (dispmode != 0)
@@ -1367,6 +1377,13 @@ void GPU2D::DrawScanline_Mode1(u32 line)
             }
         }
     }
+    else
+    {
+        for (int i = 0; i < 256; i++)
+        {
+            BGOBJLine[i] |= ((WindowMask[i] & 0x20) << 18);
+        }
+    }
 
     if (BGMosaicY >= BGMosaicYMax)
     {
@@ -1471,11 +1488,13 @@ void GPU2D::DrawBG_3D()
     {
         for (; i < iend; i++)
         {
+            int pos = xoff++;
+
             if (!(WindowMask[i] & 0x01)) continue;
 
             BGOBJLine[i+512] = BGOBJLine[i+256];
             BGOBJLine[i+256] = BGOBJLine[i];
-            BGOBJLine[i] = 0x40000000; // 3D-layer placeholder
+            BGOBJLine[i] = 0x40000000 | pos; // 3D-layer placeholder
         }
     }
     else if (LineScale == 1)
diff --git a/src/GPU3D.cpp b/src/GPU3D.cpp
index 61629a7..ec395ad 100644
--- a/src/GPU3D.cpp
+++ b/src/GPU3D.cpp
@@ -612,6 +612,11 @@ void SetDisplaySettings(int scale, bool accel)
     GLRenderer43::SetDisplaySettings(scale, accel);
 }
 
+int GetScale()
+{
+    return GLRenderer43::GetScale();
+}
+
 
 
 void MatrixLoadIdentity(s32* m)
@@ -2441,6 +2446,11 @@ u32* GetLine(int line)
     return GLRenderer43::GetLine(line);
 }
 
+void SetupAccelFrame()
+{
+    GLRenderer43::SetupAccelFrame();
+}
+
 
 void WriteToGXFIFO(u32 val)
 {
diff --git a/src/GPU3D.h b/src/GPU3D.h
index 2fe5bee..10c8966 100644
--- a/src/GPU3D.h
+++ b/src/GPU3D.h
@@ -98,6 +98,7 @@ void DoSavestate(Savestate* file);
 
 void SetEnabled(bool geometry, bool rendering);
 void SetDisplaySettings(int scale, bool accel);
+int GetScale();
 
 void ExecuteCommand();
 
@@ -129,6 +130,7 @@ void DeInit();
 void Reset();
 
 void SetDisplaySettings(int scale, bool accel);
+int GetScale();
 
 void SetupRenderThread();
 
@@ -147,6 +149,7 @@ void DeInit();
 void Reset();
 
 void SetDisplaySettings(int scale, bool accel);
+int GetScale();
 
 void VCount144();
 void RenderFrame();
diff --git a/src/GPU3D_OpenGL43.cpp b/src/GPU3D_OpenGL43.cpp
index 6db9a18..0181192 100644
--- a/src/GPU3D_OpenGL43.cpp
+++ b/src/GPU3D_OpenGL43.cpp
@@ -100,7 +100,8 @@ int ScaleFactor;
 bool Accelerated;
 int ScreenW, ScreenH;
 
-GLuint FramebufferTex[4];
+GLuint FramebufferTex[6];
+int FrontBuffer;
 GLuint FramebufferID[2], PixelbufferID;
 u32* Framebuffer = NULL;
 
@@ -262,8 +263,10 @@ bool Init()
     glGenFramebuffers(2, &FramebufferID[0]);
     glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
 
-    // color buffer
-    glGenTextures(4, &FramebufferTex[0]);
+    glGenTextures(6, &FramebufferTex[0]);
+    FrontBuffer = 0;
+
+    // color buffers
     glBindTexture(GL_TEXTURE_2D, FramebufferTex[0]);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
@@ -271,33 +274,39 @@ bool Init()
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
     glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[0], 0);
 
-    // depth/stencil buffer
     glBindTexture(GL_TEXTURE_2D, FramebufferTex[1]);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[1], 0);
+
+    // depth/stencil buffer
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[4]);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
+    glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[4], 0);
 
     // attribute buffer
     // R: opaque polyID (for edgemarking)
     // G: opaque polyID (for shadows, suppressed when rendering translucent polygons)
     // B: stencil flag
-    glBindTexture(GL_TEXTURE_2D, FramebufferTex[2]);
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[5]);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[2], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[5], 0);
 
     // downscale framebuffer, for antialiased mode
     glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[1]);
-    glBindTexture(GL_TEXTURE_2D, FramebufferTex[3]);
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[2]);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_CLAMP_TO_EDGE);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
     glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
-    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[3], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[2], 0);
 
     glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
 
@@ -330,7 +339,7 @@ bool Init()
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB5_A1, 1024, 48, 0, GL_RGBA, GL_UNSIGNED_SHORT_1_5_5_5_REV, NULL);
 
     glActiveTexture(GL_TEXTURE2);
-    glBindTexture(GL_TEXTURE_2D, FramebufferTex[2]); // opaque polyID / shadow bits
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[5]); // opaque polyID / shadow bits
 
     return true;
 }
@@ -357,11 +366,13 @@ void SetDisplaySettings(int scale, bool accel)
     glBindTexture(GL_TEXTURE_2D, FramebufferTex[0]);
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, ScreenW, ScreenH, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
     glBindTexture(GL_TEXTURE_2D, FramebufferTex[1]);
-    glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, ScreenW, ScreenH, 0, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, NULL);
+    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, ScreenW, ScreenH, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
     glBindTexture(GL_TEXTURE_2D, FramebufferTex[2]);
-    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB8UI, ScreenW, ScreenH, 0, GL_RGB_INTEGER, GL_UNSIGNED_BYTE, NULL);
-    glBindTexture(GL_TEXTURE_2D, FramebufferTex[3]);
     glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, ScreenW/2, ScreenH/2, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[4]);
+    glTexImage2D(GL_TEXTURE_2D, 0, GL_DEPTH24_STENCIL8, ScreenW, ScreenH, 0, GL_DEPTH_STENCIL, GL_UNSIGNED_INT_24_8, NULL);
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[5]);
+    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB8UI, ScreenW, ScreenH, 0, GL_RGB_INTEGER, GL_UNSIGNED_BYTE, NULL);
 
     glBindBuffer(GL_PIXEL_PACK_BUFFER, PixelbufferID);
     if (accel) glBufferData(GL_PIXEL_PACK_BUFFER, ScreenW*ScreenH*4, NULL, GL_DYNAMIC_READ);
@@ -372,6 +383,11 @@ void SetDisplaySettings(int scale, bool accel)
     else       Framebuffer = new u32[ScreenW*ScreenH];
 }
 
+int GetScale()
+{
+    return ScaleFactor;
+}
+
 
 void SetupPolygon(RendererPolygon* rp, Polygon* polygon)
 {
@@ -756,6 +772,12 @@ void RenderFrame()
     glColorMaski(0, GL_TRUE, GL_TRUE, GL_TRUE, GL_TRUE);
     glColorMaski(1, GL_TRUE, GL_TRUE, GL_TRUE, GL_FALSE);
 
+    if (Accelerated)
+    {
+        int backbuf = FrontBuffer ? 0 : 1;
+        glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[backbuf], 0);
+    }
+
     // clear buffers
     // TODO: clear bitmap
     // TODO: check whether 'clear polygon ID' affects translucent polyID
@@ -818,15 +840,7 @@ void RenderFrame()
         glBindBuffer(GL_ARRAY_BUFFER, VertexBufferID);
         glBufferSubData(GL_ARRAY_BUFFER, 0, NumVertices*7*4, VertexBuffer);
 
-        if (!ChunkedRendering)
-        {
-            RenderSceneChunk(0, 192);
-        }
-        else
-        {
-            glEnable(GL_SCISSOR_TEST);
-            RenderSceneChunk(0, 48);
-        }
+        RenderSceneChunk(0, 192);
     }
 
     if (false)
@@ -842,63 +856,24 @@ void RenderFrame()
         glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
     }
 
-    glReadBuffer(GL_COLOR_ATTACHMENT0);
-    glBindBuffer(GL_PIXEL_PACK_BUFFER, PixelbufferID);
+    if (!Accelerated)
+    {
+        glReadBuffer(GL_COLOR_ATTACHMENT0);
+        glBindBuffer(GL_PIXEL_PACK_BUFFER, PixelbufferID);
 
-    if (!ChunkedRendering)
         glReadPixels(0, 0, 256<<ScaleFactor, 192<<ScaleFactor, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
-    else
-        glReadPixels(0, 0, 256<<ScaleFactor, 48<<ScaleFactor, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
+    }
 }
 
 u32* GetLine(int line)
 {
     int stride = 256 << (ScaleFactor*2);
 
-    if (!ChunkedRendering)
+    if (line == 0)
     {
-        if (line == 0)
-        {
-            u8* data = (u8*)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
-            if (data) memcpy(&Framebuffer[stride*0], data, 4*stride*192);
-            glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
-        }
-    }
-    else
-    {
-        if (line == 0)
-        {
-            u8* data = (u8*)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
-            if (data) memcpy(&Framebuffer[stride*0], data, 4*stride*48);
-            glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
-
-            if (RenderNumPolygons) RenderSceneChunk(48, 48);
-            glReadPixels(0, 48<<ScaleFactor, 256<<ScaleFactor, 48<<ScaleFactor, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
-        }
-        else if (line == 48)
-        {
-            u8* data = (u8*)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
-            if (data) memcpy(&Framebuffer[stride*48], data, 4*stride*48);
-            glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
-
-            if (RenderNumPolygons) RenderSceneChunk(96, 48);
-            glReadPixels(0, 96<<ScaleFactor, 256<<ScaleFactor, 48<<ScaleFactor, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
-        }
-        else if (line == 96)
-        {
-            u8* data = (u8*)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
-            if (data) memcpy(&Framebuffer[stride*96], data, 4*stride*48);
-            glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
-
-            if (RenderNumPolygons) RenderSceneChunk(144, 48);
-            glReadPixels(0, 144<<ScaleFactor, 256<<ScaleFactor, 48<<ScaleFactor, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
-        }
-        else if (line == 144)
-        {
-            u8* data = (u8*)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
-            if (data) memcpy(&Framebuffer[stride*144], data, 4*stride*48);
-            glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
-        }
+        u8* data = (u8*)glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
+        if (data) memcpy(&Framebuffer[stride*0], data, 4*stride*192);
+        glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
     }
 
     u64* ptr = (u64*)&Framebuffer[stride * line];
@@ -915,7 +890,8 @@ u32* GetLine(int line)
 
 void SetupAccelFrame()
 {
-    //
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[FrontBuffer]);
+    FrontBuffer = FrontBuffer ? 0 : 1;
 }
 
 }
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index 20283d3..23e45c2 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -143,6 +143,11 @@ void SetDisplaySettings(int scale, bool accel)
     printf("SOFT RENDERER SCALE FACTOR: TODO!!!\n");
 }
 
+int GetScale()
+{
+    return 0;
+}
+
 
 
 // Notes on the interpolator:
diff --git a/src/libui_sdl/main.cpp b/src/libui_sdl/main.cpp
index b77d67f..9cdb610 100644
--- a/src/libui_sdl/main.cpp
+++ b/src/libui_sdl/main.cpp
@@ -104,6 +104,7 @@ GLuint GL_ScreenShader[3];
 struct
 {
     float uScreenSize[2];
+    u32 u3DScale;
     u32 uFilterMode;
 
 } GL_ShaderConfig;
@@ -217,6 +218,7 @@ void GLDrawing_DrawScreen()
 
         GL_ShaderConfig.uScreenSize[0] = WindowWidth;
         GL_ShaderConfig.uScreenSize[1] = WindowHeight;
+        GL_ShaderConfig.u3DScale = 1 << GPU3D::GetScale();
 
         glBindBuffer(GL_UNIFORM_BUFFER, GL_ShaderConfigUBO);
         void* unibuf = glMapBuffer(GL_UNIFORM_BUFFER, GL_WRITE_ONLY);
@@ -342,7 +344,7 @@ void GLDrawing_DrawScreen()
 
     OpenGL_UseShaderProgram(GL_ScreenShader);
 
-    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
+    glBindFramebuffer(GL_FRAMEBUFFER, 0);
     glClearColor(0, 0, 0, 1);
     glClear(GL_COLOR_BUFFER_BIT);
 
@@ -353,11 +355,14 @@ void GLDrawing_DrawScreen()
                     GL_UNSIGNED_BYTE, GPU::Framebuffer[frontbuf][0]);
     glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 768, 256<<ScreenScale[1], 192<<ScreenScale[1], GL_RGBA_INTEGER,
                     GL_UNSIGNED_BYTE, GPU::Framebuffer[frontbuf][1]);*/
-    glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 256*3, 192, GL_RGBA_INTEGER,
+    glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, 256*3 + 1, 192, GL_RGBA_INTEGER,
                     GL_UNSIGNED_BYTE, GPU::Framebuffer[frontbuf][0]);
-    glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 768, 256*3, 192, GL_RGBA_INTEGER,
+    glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 768, 256*3 + 1, 192, GL_RGBA_INTEGER,
                     GL_UNSIGNED_BYTE, GPU::Framebuffer[frontbuf][1]);
 
+    glActiveTexture(GL_TEXTURE1);
+    GPU3D::SetupAccelFrame();
+
     glBindBuffer(GL_ARRAY_BUFFER, GL_ScreenVertexBufferID);
     glBindVertexArray(GL_ScreenVertexArrayID);
     glDrawArrays(GL_TRIANGLES, 0, 4*3);
diff --git a/src/libui_sdl/main_shaders.h b/src/libui_sdl/main_shaders.h
index dcf79d9..b602013 100644
--- a/src/libui_sdl/main_shaders.h
+++ b/src/libui_sdl/main_shaders.h
@@ -24,6 +24,7 @@ const char* kScreenVS = R"(#version 420
 layout(std140, binding=0) uniform uConfig
 {
     vec2 uScreenSize;
+    uint u3DScale;
     uint uFilterMode;
 };
 
@@ -50,10 +51,12 @@ const char* kScreenFS = R"(#version 420
 layout(std140, binding=0) uniform uConfig
 {
     vec2 uScreenSize;
+    uint u3DScale;
     uint uFilterMode;
 };
 
 layout(binding=0) uniform usampler2D ScreenTex;
+layout(binding=1) uniform sampler2D _3DTex;
 
 smooth in vec2 fTexcoord;
 
@@ -63,9 +66,127 @@ void main()
 {
     uvec4 pixel = texelFetch(ScreenTex, ivec2(fTexcoord), 0);
 
+    // bit0-13: BLDCNT
+    // bit14-15: DISPCNT display mode
+    // bit16-20: EVA
+    // bit21-25: EVB
+    // bit26-30: EVY
+    uvec4 ctl = texelFetch(ScreenTex, ivec2(256*3, int(fTexcoord.y)), 0);
+    uint dispmode = (ctl.g >> 6) & 0x3;
+
+    if (dispmode == 1)
+    {
+        uint eva = ctl.b & 0x1F;
+        uint evb = (ctl.b >> 5) | ((ctl.a & 0x03) << 3);
+        uint evy = ctl.a >> 2;
+
+        uvec4 top = pixel;
+        uvec4 mid = texelFetch(ScreenTex, ivec2(fTexcoord) + ivec2(256,0), 0);
+        uvec4 bot = texelFetch(ScreenTex, ivec2(fTexcoord) + ivec2(512,0), 0);
+
+        uint winmask = top.b >> 7;
+
+        if ((top.a & 0x40) != 0)
+        {
+            float xpos = top.r + fract(fTexcoord.x);
+            uvec4 _3dpix = uvec4(texelFetch(_3DTex, ivec2(vec2(xpos, fTexcoord.y)*u3DScale), 0).bgra
+                         * vec4(63,63,63,31));
+
+            if (_3dpix.a > 0) { top = _3dpix; top.a |= 0x40; bot = mid; }
+            else              top = mid;
+        }
+        else if ((mid.a & 0x40) != 0)
+        {
+            float xpos = mid.r + fract(fTexcoord.x);
+            uvec4 _3dpix = uvec4(texelFetch(_3DTex, ivec2(vec2(xpos, fTexcoord.y)*u3DScale), 0).bgra
+                         * vec4(63,63,63,31));
+
+            if (_3dpix.a > 0) { bot = _3dpix; bot.a |= 0x40; }
+        }
+        else
+        {
+            // conditional texture fetch no good for performance, apparently
+            //texelFetch(_3DTex, ivec2(0, fTexcoord.y*2), 0);
+            bot = mid;
+        }
+
+        top.b &= 0x3F;
+        bot.b &= 0x3F;
+
+        uint target2;
+        if      ((bot.a & 0x80) != 0) target2 = 0x10;
+        else if ((bot.a & 0x40) != 0) target2 = 0x01;
+        else                          target2 = bot.a;
+        bool t2pass = ((ctl.g & target2) != 0);
+
+        uint coloreffect = 0;
+
+        if ((top.a & 0x80) != 0 && t2pass)
+        {
+            // sprite blending
+
+            coloreffect = 1;
+
+            if ((top.a & 0x40) != 0)
+            {
+                eva = top.a & 0x1F;
+                evb = 16 - eva;
+            }
+        }
+        else if ((top.a & 0x40) != 0 && t2pass)
+        {
+            // 3D layer blending
+
+            coloreffect = 4;
+            eva = (top.a & 0x1F) + 1;
+            evb = 32 - eva;
+        }
+        else
+        {
+            if      ((top.a & 0x80) != 0) top.a = 0x10;
+            else if ((top.a & 0x40) != 0) top.a = 0x01;
+
+            if ((ctl.r & top.a) != 0 && winmask != 0)
+            {
+                uint effect = ctl.r >> 6;
+                if ((effect != 1) || t2pass) coloreffect = effect;
+            }
+        }
+
+        if (coloreffect == 0)
+        {
+            pixel = top;
+        }
+        else if (coloreffect == 1)
+        {
+            pixel = ((top * eva) + (bot * evb)) >> 4;
+            pixel = min(pixel, 0x3F);
+        }
+        else if (coloreffect == 2)
+        {
+            pixel = top;
+            pixel += ((uvec4(0x3F,0x3F,0x3F,0) - pixel) * evy) >> 4;
+        }
+        else if (coloreffect == 3)
+        {
+            pixel = top;
+            pixel -= (pixel * evy) >> 4;
+        }
+        else
+        {
+            pixel = ((top * eva) + (bot * evb)) >> 5;
+            if (eva <= 16) pixel += uvec4(1,1,1,0);
+            pixel = min(pixel, 0x3F);
+        }
+    }
+
+    pixel.rgb <<= 2;
+    pixel.rgb |= (pixel.rgb >> 6);
+
     // TODO: filters
 
-    oColor = vec4(vec3(pixel.bgr) / 255.0, 1.0);
+    oColor = vec4(vec3(pixel.rgb) / 255.0, 1.0);
+    //oColor = texelFetch(_3DTex, ivec2(fTexcoord*4), 0).bgra;
 }
 )";