proof of concept that only really works in one specific case

src/GPU2D.cpp

@@ -1509,7 +1509,8 @@ void GPU2D::DrawScanline_BGOBJ(u32 line)
     }
     else
     {
-        if (Num == 0)
+        //if (Num == 0)
+        if (true)
         {
             for (int i = 0; i < 256; i++)
             {
@@ -1538,7 +1539,7 @@ void GPU2D::DrawScanline_BGOBJ(u32 line)
 
                     BGOBJLine[i]     = val2;
                     BGOBJLine[256+i] = ColorComposite(i, val2, val3);
-                    BGOBJLine[512+i] = 0x04000000 | (val1 & 0xFF);
+                    BGOBJLine[512+i] = 0x04000000 | (val1 & 0xFF) | ((val1 & 0xFF0000) << 8);
                 }
                 else if ((flag1 & 0xC0) == 0x40)
                 {
@@ -1550,7 +1551,7 @@ void GPU2D::DrawScanline_BGOBJ(u32 line)
 
                     BGOBJLine[i]     = val2;
                     BGOBJLine[256+i] = ColorComposite(i, val2, val3);
-                    BGOBJLine[512+i] = (bldcnteffect << 24) | (EVY << 8) | (val1 & 0xFF);
+                    BGOBJLine[512+i] = (bldcnteffect << 24) | (EVY << 8) | (val1 & 0xFF) | ((val1 & 0xFF0000) << 8);
                 }
                 else if (((flag2 & 0xC0) == 0x40) && ((BlendCnt & 0x01C0) == 0x0140))
                 {
@@ -1573,7 +1574,7 @@ void GPU2D::DrawScanline_BGOBJ(u32 line)
 
                     BGOBJLine[i]     = val1;
                     BGOBJLine[256+i] = ColorComposite(i, val1, val3);
-                    BGOBJLine[512+i] = (bldcnteffect << 24) | (EVB << 16) | (EVA << 8) | (val2 & 0xFF);
+                    BGOBJLine[512+i] = (bldcnteffect << 24) | (EVB << 16) | (EVA << 8) | (val2 & 0xFF) | ((val2 & 0xFF0000) << 8);
                 }
                 else
                 {
@@ -2001,6 +2002,44 @@ void GPU2D::DrawBG_Extended(u32 line, u32 bgnum)
         {
             // direct color bitmap
 
+            if (Accelerated && yshift<=8)
+            {
+                if (true) // TODO insert condition here!!
+                {
+                    for (int i = 0; i < 256; i++)
+                    {
+                        if (WindowMask[i] & (1<<bgnum))
+                        {
+                            /*s32 finalX, finalY;
+                            if (mosaic)
+                            {
+                                int im = CurBGXMosaicTable[i];
+                                finalX = rotX - (im * rotA);
+                                finalY = rotY - (im * rotC);
+                            }
+                            else
+                            {
+                                finalX = rotX;
+                                finalY = rotY;
+                            }
+
+                            if (!(finalX & ofxmask) && !(finalY & ofymask))*/
+                            {
+                                BGOBJLine[i+512] = BGOBJLine[i+256];
+                                BGOBJLine[i+256] = BGOBJLine[i];
+                                BGOBJLine[i] = 0x40800000 | i; // placeholder
+                                //BGOBJLine[i] = 0x40100000 | ((finalX >> 8) & 0xFF) | (finalY & 0xFF00); // placeholder
+                            }
+                        }
+                    }
+
+                    BGXRefInternal[bgnum-2] += rotB;
+                    BGYRefInternal[bgnum-2] += rotD;
+
+                    return;
+                }
+            }
+
             u16 color;
 
             for (int i = 0; i < 256; i++)

src/GPU_OpenGL.cpp

@@ -45,8 +45,10 @@ GLuint CompScreenInputTex;
 GLuint CompScreenOutputTex;
 GLuint CompScreenOutputFB;
 
-GLuint CaptureTex[2][4];
+GLuint CaptureTex[4][2];
-GLuint CaptureFB[2][4];
+GLuint CaptureFB[4][2];
+int CaptureSet[4];
+u32 CaptureCnt;
 
 
 bool Init()
@@ -73,6 +75,8 @@ bool Init()
         glUniform1i(uni_id, 0);
         uni_id = glGetUniformLocation(CompShader[i][2], "_3DTex");
         glUniform1i(uni_id, 1);
+        uni_id = glGetUniformLocation(CompShader[i][2], "CaptureTex");
+        glUniform1i(uni_id, 2);
     }
 
 #define SETVERTEX(i, x, y) \
@@ -128,7 +132,7 @@ bool Init()
     glGenTextures(8, &CaptureTex[0][0]);
     for (int i = 0; i < 8; i++)
     {
-        glBindTexture(GL_TEXTURE_2D, CaptureTex[i>>2][i&0x3]);
+        glBindTexture(GL_TEXTURE_2D, CaptureTex[i>>1][i&0x1]);
         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);
@@ -158,6 +162,10 @@ void DeInit()
 
 void Reset()
 {
+    for (int i = 0; i < 4; i++)
+        CaptureSet[i] = 0;
+
+    CaptureCnt = 0;
 }
 
 
@@ -179,11 +187,11 @@ void SetRenderSettings(RenderSettings& settings)
 
     for (int i = 0; i < 8; i++)
     {
-        glBindTexture(GL_TEXTURE_2D, CaptureTex[i>>2][i&0x3]);
+        glBindTexture(GL_TEXTURE_2D, CaptureTex[i>>1][i&0x1]);
         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, ScreenW, ScreenW, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
 
-        glBindFramebuffer(GL_FRAMEBUFFER, CaptureFB[i>>2][i&0x3]);
+        glBindFramebuffer(GL_FRAMEBUFFER, CaptureFB[i>>1][i&0x1]);
-        glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, CaptureTex[i>>2][i&0x3], 0);
+        glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, CaptureTex[i>>1][i&0x1], 0);
     }
 
     glBindFramebuffer(GL_FRAMEBUFFER, 0);
@@ -204,7 +212,7 @@ void DoCapture(u32 capcnt)
         return;
     }
 
-    // COMPLETE ME!
+    CaptureCnt = capcnt;
 }
 
 void RenderFrame()
@@ -238,10 +246,48 @@ void RenderFrame()
     glActiveTexture(GL_TEXTURE1);
     GPU3D::GLRenderer::SetupAccelFrame();
 
+    glActiveTexture(GL_TEXTURE2);
+    glBindTexture(GL_TEXTURE_2D, CaptureTex[3][CaptureSet[3]]);
+
     glBindBuffer(GL_ARRAY_BUFFER, CompVertexBufferID);
     glBindVertexArray(CompVertexArrayID);
     glDrawArrays(GL_TRIANGLES, 0, 4*3);
 
+    // take care of capture as needed
+    if (CaptureCnt & (1<<31))
+    {
+        u32 dstvram = (CaptureCnt >> 16) & 0x3;
+
+        CaptureSet[dstvram] ^= 1;
+        int cset = CaptureSet[dstvram];
+
+        u32 capwidth, capheight;
+        switch ((CaptureCnt >> 20) & 0x3)
+        {
+        case 0: capwidth = 128; capheight = 128; break;
+        case 1: capwidth = 256; capheight = 64;  break;
+        case 2: capwidth = 256; capheight = 128; break;
+        case 3: capwidth = 256; capheight = 192; break;
+        }
+
+        u32 yoffset = ((CaptureCnt >> 18) & 0x3) * 64;
+
+        capwidth *= Scale;
+        capheight *= Scale;
+        yoffset *= Scale;
+
+        // TODO: other sources!!
+        glBindFramebuffer(GL_READ_FRAMEBUFFER, CompScreenOutputFB);
+        glBindFramebuffer(GL_DRAW_FRAMEBUFFER, CaptureFB[dstvram][cset]);
+        glReadBuffer(GL_COLOR_ATTACHMENT0);
+        glDrawBuffer(GL_COLOR_ATTACHMENT0);
+
+        // TODO: offset and shit!
+        glBlitFramebuffer(0, 192*Scale, capwidth, 192*Scale+capheight, 0, 0, capwidth, capheight, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+
+        CaptureCnt = 0;
+    }
+
     glFlush();
 }
 

src/GPU_OpenGL_shaders.h

@@ -43,6 +43,7 @@ uniform uint u3DScale;
 
 uniform usampler2D ScreenTex;
 uniform sampler2D _3DTex;
+uniform sampler2D CaptureTex;
 
 smooth in vec2 fTexcoord;
 
@@ -70,7 +71,12 @@ void main()
 
             float xpos = val3.r + fract(fTexcoord.x);
             float ypos = mod(fTexcoord.y, 192);
-            ivec4 _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
+            ivec4 _3dpix;
+            if ((val3.a & 0x80) == 0x80)
+                _3dpix = ivec4(texelFetch(CaptureTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
+                         * vec4(63,63,63,31));
+            else
+                _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
                          * vec4(63,63,63,31));
 
             if (_3dpix.a > 0)
@@ -91,7 +97,12 @@ void main()
 
             float xpos = val3.r + fract(fTexcoord.x);
             float ypos = mod(fTexcoord.y, 192);
-            ivec4 _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
+            ivec4 _3dpix;
+            if ((val3.a & 0x80) == 0x80)
+                _3dpix = ivec4(texelFetch(CaptureTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
+                         * vec4(63,63,63,31));
+            else
+                _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
                          * vec4(63,63,63,31));
 
             if (_3dpix.a > 0)
@@ -111,7 +122,12 @@ void main()
 
             float xpos = val3.r + fract(fTexcoord.x);
             float ypos = mod(fTexcoord.y, 192);
-            ivec4 _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
+            ivec4 _3dpix;
+            if ((val3.a & 0x80) == 0x80)
+                _3dpix = ivec4(texelFetch(CaptureTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
+                         * vec4(63,63,63,31));
+            else
+                _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
                          * vec4(63,63,63,31));
 
             if (_3dpix.a > 0)