start work on display capture

also fix a bug in the compositing shader

src/GPU.cpp

@@ -907,7 +907,7 @@ void StartScanline(u32 line)
     else
     {
         if (VCount == 192)
-        {//printf("- VBLANK -\n");vbltime=NDS::ARM9Timestamp;
+        {
             // VBlank
             DispStat[0] |= (1<<0);
             DispStat[1] |= (1<<0);

src/GPU2D.cpp

@@ -656,8 +656,16 @@ void GPU2D::DrawScanline(u32 line)
     u32 dispmode = DispCnt >> 16;
     dispmode &= (Num ? 0x1 : 0x3);
 
-    if (Num == 0 && !Accelerated)
+    if (Num == 0)
-        _3DLine = GPU3D::GetLine(n3dline);
+    {
+        if (!Accelerated)
+            _3DLine = GPU3D::GetLine(n3dline);
+        else if ((CaptureCnt & (1<<31)) && (((CaptureCnt >> 29) & 0x3) != 1))
+        {
+            _3DLine = GPU3D::GetLine(n3dline);
+            //GPU3D::GLRenderer43::PrepareCaptureFrame();
+        }
+    }
 
     // always render regular graphics
     DrawScanline_Mode1(line);
@@ -862,6 +870,12 @@ void GPU2D::VBlankEnd()
     BGMosaicYMax = BGMosaicSize[1];
     OBJMosaicY = 0;
     OBJMosaicYMax = OBJMosaicSize[1];
+
+    // TODO: make optional
+    if ((Num == 0) && (CaptureCnt & (1<<31)) && (((CaptureCnt >> 29) & 0x3) != 1))
+    {
+        GPU3D::GLRenderer43::PrepareCaptureFrame();
+    }
 }
 
 
@@ -884,7 +898,7 @@ void GPU2D::DoCapture(u32 line, u32 width)
         srcA = _3DLine;
     else
         srcA = BGOBJLine;
-
+srcA = _3DLine;
     u16* srcB = NULL;
     u32 srcBaddr = line * 256;
 

src/GPU3D.h

@@ -153,6 +153,7 @@ int GetScale();
 
 void VCount144();
 void RenderFrame();
+void PrepareCaptureFrame();
 u32* GetLine(int line);
 void SetupAccelFrame();
 

src/GPU3D_OpenGL43.cpp

@@ -100,9 +100,9 @@ int ScaleFactor;
 bool Accelerated;
 int ScreenW, ScreenH;
 
-GLuint FramebufferTex[6];
+GLuint FramebufferTex[8];
 int FrontBuffer;
-GLuint FramebufferID[2], PixelbufferID;
+GLuint FramebufferID[4], PixelbufferID;
 u32* Framebuffer = NULL;
 
 bool ChunkedRendering = false;
@@ -260,10 +260,10 @@ bool Init()
     glVertexAttribIPointer(3, 3, GL_UNSIGNED_INT, 7*4, (void*)(4*4));
 
 
-    glGenFramebuffers(2, &FramebufferID[0]);
+    glGenFramebuffers(4, &FramebufferID[0]);
     glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
 
-    glGenTextures(6, &FramebufferTex[0]);
+    glGenTextures(8, &FramebufferTex[0]);
     FrontBuffer = 0;
 
     // color buffers
@@ -286,7 +286,12 @@ bool Init()
     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);
+
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[6]);
+    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);
 
     // attribute buffer
     // R: opaque polyID (for edgemarking)
@@ -297,16 +302,40 @@ bool Init()
     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[5], 0);
 
-    // downscale framebuffer, for antialiased mode
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[7]);
-    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[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);
+
+    // downscale framebuffer for antialiased mode
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[2]);
     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[2], 0);
+
+    // downscale framebuffer for display capture (always 256x192)
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[3]);
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[3]);
+    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);
+    glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA, 256, 192, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[3], 0);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[0], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[4], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[5], 0);
+
+    glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[1]);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[1], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, FramebufferTex[6], 0);
+    glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT1, FramebufferTex[7], 0);
 
     glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
 
@@ -373,10 +402,13 @@ void SetDisplaySettings(int scale, bool accel)
     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);
+    glBindTexture(GL_TEXTURE_2D, FramebufferTex[6]);
+    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[7]);
+    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);
+    glBufferData(GL_PIXEL_PACK_BUFFER, 256*192*4, NULL, GL_DYNAMIC_READ);
-    else       glBufferData(GL_PIXEL_PACK_BUFFER, 256*192, NULL, GL_DYNAMIC_READ);
 
     if (Framebuffer) delete[] Framebuffer;
     if (accel) Framebuffer = new u32[256*192];
@@ -774,7 +806,8 @@ void RenderFrame()
 
     if (Accelerated)
     {
-        glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[FrontBuffer], 0);
+        //glFramebufferTexture(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, FramebufferTex[FrontBuffer], 0);
+        glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[FrontBuffer]);
         FrontBuffer = FrontBuffer ? 0 : 1;
     }
 
@@ -856,18 +889,36 @@ void RenderFrame()
         glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[0]);
     }
 
-    if (!Accelerated)
+    /*if (!Accelerated)
     {
         glReadBuffer(GL_COLOR_ATTACHMENT0);
         glBindBuffer(GL_PIXEL_PACK_BUFFER, PixelbufferID);
 
         glReadPixels(0, 0, 256<<ScaleFactor, 192<<ScaleFactor, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
-    }
+    }*/
+}
+
+void PrepareCaptureFrame()
+{
+    // TODO: make sure this picks the right buffer when doing antialiasing
+    int original_fb = FrontBuffer^1;
+
+    glBindFramebuffer(GL_READ_FRAMEBUFFER, FramebufferID[original_fb]);
+    glReadBuffer(GL_COLOR_ATTACHMENT0);
+    glBindFramebuffer(GL_DRAW_FRAMEBUFFER, FramebufferID[3]);
+    glDrawBuffer(GL_COLOR_ATTACHMENT0);
+    glBlitFramebuffer(0, 0, ScreenW, ScreenH, 0, 0, 256, 192, GL_COLOR_BUFFER_BIT, GL_NEAREST);
+
+    glBindFramebuffer(GL_READ_FRAMEBUFFER, FramebufferID[3]);
+    glReadPixels(0, 0, 256, 192, GL_BGRA, GL_UNSIGNED_BYTE, NULL);
+
+    //glBindFramebuffer(GL_FRAMEBUFFER, FramebufferID[original_fb]);
+    //glFlush();
 }
 
 u32* GetLine(int line)
 {
-    int stride = 256 << (ScaleFactor*2);
+    int stride = 256;// << (ScaleFactor*2);
 
     if (line == 0)
     {

src/libui_sdl/main_shaders.h

@@ -65,7 +65,7 @@ out vec4 oColor;
 void main()
 {
     ivec4 pixel = ivec4(texelFetch(ScreenTex, ivec2(fTexcoord), 0));
-
+ivec4 zog=pixel;
     // bit0-13: BLDCNT
     // bit14-15: DISPCNT display mode
     // bit16-20: EVA
@@ -86,7 +86,7 @@ void main()
 
         int winmask = top.b >> 7;
 
-        if ((top.a & 0x40) != 0)
+        if ((top.a & 0xC0) == 0x40)
         {
             float xpos = top.r + fract(fTexcoord.x);
             float ypos = mod(fTexcoord.y, 768);
@@ -96,7 +96,7 @@ void main()
             if (_3dpix.a > 0) { top = _3dpix; top.a |= 0x40; bot = mid; }
             else              top = mid;
         }
-        else if ((mid.a & 0x40) != 0)
+        else if ((mid.a & 0xC0) == 0x40)
         {
             float xpos = mid.r + fract(fTexcoord.x);
             float ypos = mod(fTexcoord.y, 768);