gistfile1.diff

diff --git a/externals/boost b/externals/boost
--- a/externals/boost
+++ b/externals/boost
@@ -1 +1 @@
-Subproject commit 728a4d7d1c8b28355544ae829df9c4b5f28373c5
+Subproject commit 728a4d7d1c8b28355544ae829df9c4b5f28373c5-dirty
diff --git a/externals/nihstro b/externals/nihstro
--- a/externals/nihstro
+++ b/externals/nihstro
@@ -1 +1 @@
-Subproject commit 0a8b4d221425f13e24a3cef9b02edc3221bab211
+Subproject commit 0a8b4d221425f13e24a3cef9b02edc3221bab211-dirty
diff --git a/src/core/hw/gpu.cpp b/src/core/hw/gpu.cpp
index 8800235..31a2045 100644
--- a/src/core/hw/gpu.cpp
+++ b/src/core/hw/gpu.cpp
@@ -52,6 +52,37 @@ inline void Read(T &var, const u32 raw_addr) {
     var = g_regs[addr / 4];
 }
 
+u32 Compact1By1(u32 x)
+{
+    x &= 0x55555555;                  // x = -f-e -d-c -b-a -9-8 -7-6 -5-4 -3-2 -1-0
+    x = (x ^ (x >> 1)) & 0x33333333; // x = --fe --dc --ba --98 --76 --54 --32 --10
+    x = (x ^ (x >> 2)) & 0x0f0f0f0f; // x = ---- fedc ---- ba98 ---- 7654 ---- 3210
+    x = (x ^ (x >> 4)) & 0x00ff00ff; // x = ---- ---- fedc ba98 ---- ---- 7654 3210
+    x = (x ^ (x >> 8)) & 0x0000ffff; // x = ---- ---- ---- ---- fedc ba98 7654 3210
+    return x;
+}
+
+// Inverse of Part1By2 - "delete" all bits not at positions divisible by 3
+u32 Compact1By2(u32 x)
+{
+    x &= 0x09249249;                  // x = ---- 9--8 --7- -6-- 5--4 --3- -2-- 1--0
+    x = (x ^ (x >> 2)) & 0x030c30c3; // x = ---- --98 ---- 76-- --54 ---- 32-- --10
+    x = (x ^ (x >> 4)) & 0x0300f00f; // x = ---- --98 ---- ---- 7654 ---- ---- 3210
+    x = (x ^ (x >> 8)) & 0xff0000ff; // x = ---- --98 ---- ---- ---- ---- 7654 3210
+    x = (x ^ (x >> 16)) & 0x000003ff; // x = ---- ---- ---- ---- ---- --98 7654 3210
+    return x;
+}
+
+u32 DecodeMorton2X(u32 code)
+{
+    return Compact1By1(code >> 0);
+}
+
+u32 DecodeMorton2Y(u32 code)
+{
+    return Compact1By1(code >> 1);
+}
+
 template <typename T>
 inline void Write(u32 addr, const T data) {
     addr -= 0x1EF00000;
@@ -130,6 +161,10 @@ inline void Write(u32 addr, const T data) {
                         int r, g, b, a;
                     } source_color = { 0, 0, 0, 0 };
 
+                    u32 morton_code = y * config.input_width + x;
+                    u32 curve_x = DecodeMorton2X(morton_code);
+                    u32 curve_y = DecodeMorton2Y(morton_code);
+
                     switch (config.input_format) {
                     case Regs::PixelFormat::RGBA8:
                     {
@@ -180,7 +215,7 @@ inline void Write(u32 addr, const T data) {
 
                     case Regs::PixelFormat::RGB8:
                     {
-                        u8* dstptr = dest_pointer + (x + y * output_width) * 3;
+                        u8* dstptr = dest_pointer + (curve_x + curve_y * output_width) * 3;
                         dstptr[2] = source_color.r; // red
                         dstptr[1] = source_color.g; // green
                         dstptr[0] = source_color.b; // blue
@@ -189,7 +224,7 @@ inline void Write(u32 addr, const T data) {
 
                     case Regs::PixelFormat::RGB5A1:
                     {
-                        u16* dstptr = (u16*)(dest_pointer + x * 2 + y * config.output_width * 2);
+                        u16* dstptr = (u16*)(dest_pointer + curve_x * 2 + curve_y * config.output_width * 2);
                         *dstptr = ((source_color.r >> 3) << 11) | ((source_color.g >> 3) << 6)
                                 | ((source_color.b >> 3) <<  1) | ( source_color.a >> 7);
                         break;
@@ -197,7 +232,7 @@ inline void Write(u32 addr, const T data) {
 
                     case Regs::PixelFormat::RGBA4:
                     {
-                        u16* dstptr = (u16*)(dest_pointer + x * 2 + y * config.output_width * 2);
+                        u16* dstptr = (u16*)(dest_pointer + curve_x * 2 + curve_y * config.output_width * 2);
                         *dstptr = ((source_color.r >> 4) << 12) | ((source_color.g >> 4) << 8)
                                 | ((source_color.b >> 4) <<  4) | ( source_color.a >> 4);
                         break;
diff --git a/src/video_core/rasterizer.cpp b/src/video_core/rasterizer.cpp
index 17f8f70..b932873 100644
--- a/src/video_core/rasterizer.cpp
+++ b/src/video_core/rasterizer.cpp
@@ -18,6 +18,19 @@ namespace Pica {
 
 namespace Rasterizer {
 
+    unsigned int SeparateBy1(unsigned int x) {
+        x &= 0x0000ffff;                  // x = ---- ---- ---- ---- fedc ba98 7654 3210
+        x = (x ^ (x << 8)) & 0x00ff00ff; // x = ---- ---- fedc ba98 ---- ---- 7654 3210
+        x = (x ^ (x << 4)) & 0x0f0f0f0f; // x = ---- fedc ---- ba98 ---- 7654 ---- 3210
+        x = (x ^ (x << 2)) & 0x33333333; // x = --fe --dc --ba --98 --76 --54 --32 --10
+        x = (x ^ (x << 1)) & 0x55555555; // x = -f-e -d-c -b-a -9-8 -7-6 -5-4 -3-2 -1-0
+        return x;
+    }
+
+    u32 MortonCode2(unsigned int x, unsigned int y) {
+        return SeparateBy1(x) | (SeparateBy1(y) << 1);
+    }
+
 static void DrawPixel(int x, int y, const Math::Vec4<u8>& color) {
     const PAddr addr = registers.framebuffer.GetColorBufferPhysicalAddress();
     u8* color_buffer = Memory::GetPointer(PAddrToVAddr(addr));
@@ -26,10 +39,12 @@ static void DrawPixel(int x, int y, const Math::Vec4<u8>& color) {
     // NOTE: The framebuffer height register contains the actual FB height minus one.
     y = (registers.framebuffer.height - y);
 
+    u32 code = MortonCode2(x, y);
+
     switch (registers.framebuffer.color_format) {
     case registers.framebuffer.RGBA8:
     {
-        u8* pixel = color_buffer + (x + y * registers.framebuffer.GetWidth()) * 4;
+        u8* pixel = color_buffer + code * 4;
         pixel[3] = color.r();
         pixel[2] = color.g();
         pixel[1] = color.b();
@@ -48,12 +63,13 @@ static const Math::Vec4<u8> GetPixel(int x, int y) {
     u8* color_buffer = Memory::GetPointer(PAddrToVAddr(addr));
 
     y = (registers.framebuffer.height - y);
+    u32 code = MortonCode2(x, y);
 
     switch (registers.framebuffer.color_format) {
     case registers.framebuffer.RGBA8:
     {
         Math::Vec4<u8> ret;
-        u8* pixel = color_buffer + (x + y * registers.framebuffer.GetWidth()) * 4;
+        u8* pixel = color_buffer + code * 4;
         ret.r() = pixel[3];
         ret.g() = pixel[2];
         ret.b() = pixel[1];
@@ -73,9 +89,10 @@ static u32 GetDepth(int x, int y) {
     u16* depth_buffer = reinterpret_cast<u16*>(Memory::GetPointer(PAddrToVAddr(addr)));
 
     y = (registers.framebuffer.height - y);
+    u32 code = MortonCode2(x, y);
 
     // Assuming 16-bit depth buffer format until actual format handling is implemented
-    return *(depth_buffer + x + y * registers.framebuffer.GetWidth());
+    return depth_buffer[code];
 }
 
 static void SetDepth(int x, int y, u16 value) {
@@ -83,9 +100,10 @@ static void SetDepth(int x, int y, u16 value) {
     u16* depth_buffer = reinterpret_cast<u16*>(Memory::GetPointer(PAddrToVAddr(addr)));
 
     y = (registers.framebuffer.height - y);
+    u32 code = MortonCode2(x, y);
 
     // Assuming 16-bit depth buffer format until actual format handling is implemented
-    *(depth_buffer + x + y * registers.framebuffer.GetWidth()) = value;
+    depth_buffer[code] = value;
 }
 
 // NOTE: Assuming that rasterizer coordinates are 12.4 fixed-point values
diff --git a/src/video_core/renderer_opengl/renderer_opengl.cpp b/src/video_core/renderer_opengl/renderer_opengl.cpp
index 2726951..f357b5c 100644
--- a/src/video_core/renderer_opengl/renderer_opengl.cpp
+++ b/src/video_core/renderer_opengl/renderer_opengl.cpp
@@ -262,6 +262,7 @@ void RendererOpenGL::DrawScreens() {
 
     DrawSingleScreenRotated(textures[0], top_x, 0,
         (float)VideoCore::kScreenTopWidth, (float)VideoCore::kScreenTopHeight);
+    glFlush();
     DrawSingleScreenRotated(textures[1], bottom_x, (float)VideoCore::kScreenTopHeight,
         (float)VideoCore::kScreenBottomWidth, (float)VideoCore::kScreenBottomHeight);