GPU: forward BG0HOFS to internal rendering engine register for 3D layer scroll (only when the rendering engine is enabled).

fixes #840

thank you RSDuck and Hydr8gon for your insight into this.

6 files changed, 81 insertions, 37 deletions

diff --git a/src/GPU2D.cpp b/src/GPU2D.cpp
index d2a8b34..eb160d8 100644
--- a/src/GPU2D.cpp
+++ b/src/GPU2D.cpp
@@ -303,8 +303,14 @@ void GPU2D::Write8(u32 addr, u8 val)
     case 0x00E: BGCnt[3] = (BGCnt[3] & 0xFF00) | val; return;
     case 0x00F: BGCnt[3] = (BGCnt[3] & 0x00FF) | (val << 8); return;
 
-    case 0x010: BGXPos[0] = (BGXPos[0] & 0xFF00) | val; return;
-    case 0x011: BGXPos[0] = (BGXPos[0] & 0x00FF) | (val << 8); return;
+    case 0x010:
+        BGXPos[0] = (BGXPos[0] & 0xFF00) | val;
+        if (Num == 0) GPU3D::SetRenderXPos(BGXPos[0]);
+        return;
+    case 0x011:
+        BGXPos[0] = (BGXPos[0] & 0x00FF) | (val << 8);
+        if (Num == 0) GPU3D::SetRenderXPos(BGXPos[0]);
+        return;
     case 0x012: BGYPos[0] = (BGYPos[0] & 0xFF00) | val; return;
     case 0x013: BGYPos[0] = (BGYPos[0] & 0x00FF) | (val << 8); return;
     case 0x014: BGXPos[1] = (BGXPos[1] & 0xFF00) | val; return;
@@ -401,7 +407,10 @@ void GPU2D::Write16(u32 addr, u16 val)
     case 0x00C: BGCnt[2] = val; return;
     case 0x00E: BGCnt[3] = val; return;
 
-    case 0x010: BGXPos[0] = val; return;
+    case 0x010:
+        BGXPos[0] = val;
+        if (Num == 0) GPU3D::SetRenderXPos(BGXPos[0]);
+        return;
     case 0x012: BGYPos[0] = val; return;
     case 0x014: BGXPos[1] = val; return;
     case 0x016: BGYPos[1] = val; return;
@@ -716,4 +725,4 @@ void GPU2D::GetOBJVRAM(u8*& data, u32& mask)
         data = GPU::VRAMFlat_BOBJ;
         mask = 0x1FFFF;
     }
-}
\ No newline at end of file
+}
diff --git a/src/GPU2D_Soft.cpp b/src/GPU2D_Soft.cpp
index 7345af9..c686bad 100644
--- a/src/GPU2D_Soft.cpp
+++ b/src/GPU2D_Soft.cpp
@@ -403,7 +403,7 @@ void GPU2D_Soft::DoCapture(u32 line, u32 width)
                 {
                     // 3D on top, blending
 
-                    u32 _3dval = _3DLine[val3 & 0xFF];
+                    u32 _3dval = _3DLine[i];
                     if ((_3dval >> 24) > 0)
                         val1 = ColorBlend5(_3dval, val1);
                     else
@@ -413,7 +413,7 @@ void GPU2D_Soft::DoCapture(u32 line, u32 width)
                 {
                     // 3D on bottom, blending
 
-                    u32 _3dval = _3DLine[val3 & 0xFF];
+                    u32 _3dval = _3DLine[i];
                     if ((_3dval >> 24) > 0)
                     {
                         u32 eva = (val3 >> 8) & 0x1F;
@@ -428,7 +428,7 @@ void GPU2D_Soft::DoCapture(u32 line, u32 width)
                 {
                     // 3D on top, normal/fade
 
-                    u32 _3dval = _3DLine[val3 & 0xFF];
+                    u32 _3dval = _3DLine[i];
                     if ((_3dval >> 24) > 0)
                     {
                         u32 evy = (val3 >> 8) & 0x1F;
@@ -807,7 +807,7 @@ void GPU2D_Soft::DrawScanline_BGOBJ(u32 line)
 
                     BGOBJLine[i]     = val2;
                     BGOBJLine[256+i] = ColorComposite(i, val2, val3);
-                    BGOBJLine[512+i] = 0x04000000 | (val1 & 0xFF);
+                    BGOBJLine[512+i] = 0x04000000;
                 }
                 else if ((flag1 & 0xC0) == 0x40)
                 {
@@ -819,7 +819,7 @@ void GPU2D_Soft::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);
                 }
                 else if (((flag2 & 0xC0) == 0x40) && ((BlendCnt & 0x01C0) == 0x0140))
                 {
@@ -842,7 +842,7 @@ void GPU2D_Soft::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);
                 }
                 else
                 {
@@ -910,39 +910,24 @@ void GPU2D_Soft::DrawPixel_Accel(u32* dst, u16 color, u32 flag)
 
 void GPU2D_Soft::DrawBG_3D()
 {
-    u16 xoff = BGXPos[0];
     int i = 0;
-    int iend = 256;
-
-    if (xoff & 0x100)
-    {
-        i = (0x100 - (xoff & 0xFF));
-        xoff += i;
-    }
-    if ((xoff - i + iend - 1) & 0x100)
-    {
-        iend -= (xoff & 0xFF);
-    }
 
     if (Accelerated)
     {
-        for (; i < iend; i++)
+        for (i = 0; i < 256; i++)
         {
-            int pos = xoff++;
-
             if (!(WindowMask[i] & 0x01)) continue;
 
             BGOBJLine[i+512] = BGOBJLine[i+256];
             BGOBJLine[i+256] = BGOBJLine[i];
-            BGOBJLine[i] = 0x40000000 | pos; // 3D-layer placeholder
+            BGOBJLine[i] = 0x40000000; // 3D-layer placeholder
         }
     }
     else
     {
-        for (; i < iend; i++)
+        for (i = 0; i < 256; i++)
         {
-            u32 c = _3DLine[xoff];
-            xoff++;
+            u32 c = _3DLine[i];
 
             if ((c >> 24) == 0) continue;
             if (!(WindowMask[i] & 0x01)) continue;
@@ -2227,4 +2212,4 @@ void GPU2D_Soft::MosaicXSizeChanged()
 {
     CurBGXMosaicTable = MosaicTable[BGMosaicSize[0]];
     CurOBJXMosaicTable = MosaicTable[OBJMosaicSize[1]];
-}
\ No newline at end of file
+}
diff --git a/src/GPU3D.cpp b/src/GPU3D.cpp
index fd8d320..a02e286 100644
--- a/src/GPU3D.cpp
+++ b/src/GPU3D.cpp
@@ -181,6 +181,8 @@ u32 RenderClearAttr1, RenderClearAttr2;
 
 bool RenderFrameIdentical;
 
+u16 RenderXPos;
+
 u32 ZeroDotWLimit;
 
 u32 GXStat;
@@ -385,6 +387,8 @@ void Reset()
     FlushAttributes = 0;
 
     ResetRenderingState();
+
+    RenderXPos = 0;
 }
 
 void DoSavestate(Savestate* file)
@@ -430,6 +434,8 @@ void DoSavestate(Savestate* file)
     file->Var32(&RenderClearAttr1);
     file->Var32(&RenderClearAttr2);
 
+    file->Var16(&RenderXPos);
+
     file->Var32(&ZeroDotWLimit);
 
     file->Var32(&GXStat);
@@ -587,8 +593,6 @@ void DoSavestate(Savestate* file)
         }
     }
 
-    // probably not worth storing the vblank-latched Renderxxxxxx variables
-
     CmdStallQueue->DoSavestate(file);
     file->Var32((u32*)&VertexPipeline);
     file->Var32((u32*)&NormalPipeline);
@@ -2564,14 +2568,48 @@ void VCount215()
 #endif
 }
 
+void SetRenderXPos(u16 xpos)
+{
+    if (!RenderingEnabled) return;
+
+    RenderXPos = xpos & 0x01FF;
+}
+
+u32 ScrolledLine[256];
+
 u32* GetLine(int line)
 {
-    if (GPU::Renderer == 0) return SoftRenderer::GetLine(line);
+    u32* rawline;
+
+    if (GPU::Renderer == 0) rawline = SoftRenderer::GetLine(line);
 #ifdef OGLRENDERER_ENABLED
-    else                    return GLRenderer::GetLine(line);
+    else                    rawline = GLRenderer::GetLine(line);
 #else
     return NULL;
 #endif
+
+    if (RenderXPos == 0) return rawline;
+
+    // apply X scroll
+
+    if (RenderXPos & 0x100)
+    {
+        int i = 0, j = RenderXPos;
+        for (; j < 512; i++, j++)
+            ScrolledLine[i] = 0;
+        for (j = 0; i < 256; i++, j++)
+            ScrolledLine[i] = rawline[j];
+    }
+    else
+    {
+        int i = 0, j = RenderXPos;
+        for (; j < 256; i++, j++)
+            ScrolledLine[i] = rawline[j];
+        for (; i < 256; i++)
+            ScrolledLine[i] = 0;
+    }
+
+    return ScrolledLine;
 }
 
 
diff --git a/src/GPU3D.h b/src/GPU3D.h
index 0477c4f..69b67fa 100644
--- a/src/GPU3D.h
+++ b/src/GPU3D.h
@@ -89,6 +89,8 @@ extern u32 RenderClearAttr1, RenderClearAttr2;
 
 extern bool RenderFrameIdentical;
 
+extern u16 RenderXPos;
+
 extern std::array<Polygon*,2048> RenderPolygonRAM;
 extern u32 RenderNumPolygons;
 
@@ -114,6 +116,8 @@ void CheckFIFODMA();
 void VCount144();
 void VBlank();
 void VCount215();
+
+void SetRenderXPos(u16 xpos);
 u32* GetLine(int line);
 
 void WriteToGXFIFO(u32 val);
diff --git a/src/GPU_OpenGL.cpp b/src/GPU_OpenGL.cpp
index 359e9cd..0c6cf00 100644
--- a/src/GPU_OpenGL.cpp
+++ b/src/GPU_OpenGL.cpp
@@ -36,6 +36,7 @@ int ScreenH, ScreenW;
 
 GLuint CompShader[1][3];
 GLuint CompScaleLoc[1];
+GLuint Comp3DXPosLoc[1];
 
 GLuint CompVertexBufferID;
 GLuint CompVertexArrayID;
@@ -64,6 +65,7 @@ bool Init()
             return false;
 
         CompScaleLoc[i] = glGetUniformLocation(CompShader[i][2], "u3DScale");
+        Comp3DXPosLoc[i] = glGetUniformLocation(CompShader[i][2], "u3DXPos");
 
         glUseProgram(CompShader[i][2]);
         uni_id = glGetUniformLocation(CompShader[i][2], "ScreenTex");
@@ -180,6 +182,9 @@ void RenderFrame()
     OpenGL::UseShaderProgram(CompShader[0]);
     glUniform1ui(CompScaleLoc[0], Scale);
 
+    // TODO: support setting this midframe, if ever needed
+    glUniform1i(Comp3DXPosLoc[0], ((int)GPU3D::RenderXPos << 23) >> 23);
+
     int frontbuf = GPU::FrontBuffer;
     glActiveTexture(GL_TEXTURE0);
     glBindTexture(GL_TEXTURE_2D, CompScreenInputTex);
diff --git a/src/GPU_OpenGL_shaders.h b/src/GPU_OpenGL_shaders.h
index 20ac767..03ddb7a 100644
--- a/src/GPU_OpenGL_shaders.h
+++ b/src/GPU_OpenGL_shaders.h
@@ -40,6 +40,7 @@ void main()
 const char* kCompositorFS_Nearest = R"(#version 140
 
 uniform uint u3DScale;
+uniform int u3DXPos;
 
 uniform usampler2D ScreenTex;
 uniform sampler2D _3DTex;
@@ -52,6 +53,8 @@ void main()
 {
     ivec4 pixel = ivec4(texelFetch(ScreenTex, ivec2(fTexcoord), 0));
 
+    float _3dxpos = float(u3DXPos);
+
     ivec4 mbright = ivec4(texelFetch(ScreenTex, ivec2(256*3, int(fTexcoord.y)), 0));
     int dispmode = mbright.b & 0x3;
 
@@ -68,7 +71,7 @@ void main()
         {
             // 3D on top, blending
 
-            float xpos = val3.r + fract(fTexcoord.x);
+            float xpos = fTexcoord.x + _3dxpos;
             float ypos = mod(fTexcoord.y, 192);
             ivec4 _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
                          * vec4(63,63,63,31));
@@ -89,7 +92,7 @@ void main()
         {
             // 3D on bottom, blending
 
-            float xpos = val3.r + fract(fTexcoord.x);
+            float xpos = fTexcoord.x + _3dxpos;
             float ypos = mod(fTexcoord.y, 192);
             ivec4 _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
                          * vec4(63,63,63,31));
@@ -109,7 +112,7 @@ void main()
         {
             // 3D on top, normal/fade
 
-            float xpos = val3.r + fract(fTexcoord.x);
+            float xpos = fTexcoord.x + _3dxpos;
             float ypos = mod(fTexcoord.y, 192);
             ivec4 _3dpix = ivec4(texelFetch(_3DTex, ivec2(vec2(xpos, ypos)*u3DScale), 0).bgra
                          * vec4(63,63,63,31));