Fix/implement cursor

Author: Stovent

src/CDI/CMakeLists.txt

@@ -41,7 +41,7 @@ if(NOT CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
 endif()
 
 if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
-    target_compile_options(CeDImu PRIVATE -Wnrvo)
+    # target_compile_options(CeDImu PRIVATE -Wnrvo)
     # target_link_options(CeDImu PUBLIC -fsanitize=address -static-libasan)
     target_link_libraries(CeDImu PRIVATE stdc++fs stdc++exp)
 endif()

src/CDI/Video/RendererSIMD.cpp

@@ -3,6 +3,7 @@
 #include "../common/panic.hpp"
 #include "../common/utils.hpp"
 #include "../common/VideoSIMD.hpp"
+#include "../common/VideoU32.hpp"
 
 #include <cstring>
 
@@ -65,7 +66,8 @@ const Plane& RendererSIMD::RenderFrame() noexcept
     if(m_cursorEnabled)
     {
         DrawCursor();
-        Video::paste(m_screen.data(), m_screen.m_width, m_screen.m_height, m_cursorPlane.data(), m_cursorPlane.m_width, m_cursorPlane.m_height, m_cursorX, m_cursorY);
+        // TODO: double resolution
+        Video::paste(m_screenARGB.data(), m_screen.m_width, m_screen.m_height, m_cursorPlaneARGB.data(), m_cursorPlaneARGB.m_width, m_cursorPlaneARGB.m_height, m_cursorX >> 1, m_cursorY);
     }
 
     // TODO: this should be on the GUI side.
@@ -161,7 +163,7 @@ void RendererSIMD::DrawCursor() noexcept
     {
         FixedPixelSIMD pixel{&*it, stdx::element_aligned};
 
-        for(int x = m_cursorPlaneARGB.m_width, pix = 0; --x >= 0; pix++)
+        for(int x = m_cursorPlaneARGB.m_width - 1, pix = 0; --x >= 0; pix++)
         {
             const uint16_t mask = (1 << x);
             if(m_cursorPatterns[pattern] & mask)

src/CDI/Video/RendererSoftware.cpp

@@ -62,7 +62,7 @@ const Plane& RendererSoftware::RenderFrame() noexcept
     if(m_cursorEnabled)
     {
         DrawCursor();
-        Video::paste(m_screen.data(), m_screen.m_width, m_screen.m_height, m_cursorPlane.data(), m_cursorPlane.m_width, m_cursorPlane.m_height, m_cursorX, m_cursorY);
+        Video::paste(m_screen.data(), m_screen.m_width, m_screen.m_height, m_cursorPlane.data(), m_cursorPlane.m_width, m_cursorPlane.m_height, m_cursorX >> 1, m_cursorY);
     }
 
     m_lineNumber = 0;

src/CDI/Video/RendererSoftwareU32.cpp

@@ -64,7 +64,8 @@ const Plane& RendererSoftwareU32::RenderFrame() noexcept
     if(m_cursorEnabled)
     {
         DrawCursor();
-        Video::paste(m_screen.data(), m_screen.m_width, m_screen.m_height, m_cursorPlane.data(), m_cursorPlane.m_width, m_cursorPlane.m_height, m_cursorX, m_cursorY);
+        // TODO: double resolution
+        Video::paste(m_screenARGB.data(), m_screen.m_width, m_screen.m_height, m_cursorPlaneARGB.data(), m_cursorPlaneARGB.m_width, m_cursorPlaneARGB.m_height, m_cursorX >> 1, m_cursorY);
     }
 
     // TODO: this should be on the GUI side.
@@ -137,26 +138,24 @@ void RendererSoftwareU32::DrawCursor() noexcept
     const Pixel color = backdropCursorColorToARGB(m_cursorColor);
     const Pixel black{0};
 
-    int pattern = 0;
-    for(PlaneU32::iterator it = m_cursorPlaneARGB.begin(); it < m_cursorPlaneARGB.end();)
+    PlaneU32::iterator it = m_cursorPlaneARGB.begin();
+    for(int y = 0; y < m_cursorPlaneARGB.m_height; ++y)
     {
-        for(int x = 0; x < m_cursorPlaneARGB.m_width; ++x)
+        for(int x = m_cursorPlaneARGB.m_width - 1; x >= 0; --x)
         {
             const uint16_t mask = (1 << x);
-            if(m_cursorPatterns[pattern] & mask)
+            if(m_cursorPatterns[y] & mask)
                 *it = color;
             else
                 *it = black;
             ++it;
         }
-
-        ++pattern;
     }
 }
 
-static constexpr uint8_t intByteMult(uint32_t color1, uint32_t color2) noexcept {
-    return static_cast<uint8_t>(((color1 * (color2 | color2 << 8)) + 0x8080) >> 16);
-}
+// static constexpr uint8_t intByteMult(uint32_t color1, uint32_t color2) noexcept {
+//     return static_cast<uint8_t>(((color1 * (color2 | color2 << 8)) + 0x8080) >> 16);
+// }
 
 /** \brief Apply the given Image Contribution Factor to the given color component (V.5.9). */
 static constexpr uint32_t applyICFComponent(const int color, const int icf) noexcept
@@ -179,8 +178,6 @@ static constexpr Pixel applyICF(const Pixel pixel, const int icf) noexcept
 /** \brief Apply mixing to the given color components after ICF (V.5.9.1). */
 static constexpr Pixel mix(const Pixel a, const Pixel b) noexcept
 {
-    // return limu8(a + b - 16);
-
     uint8_t ra = a >> 16;
     uint8_t ga = a >> 8;
     uint8_t ba = a;

src/CDI/common/Video.cpp

@@ -344,9 +344,12 @@ void paste(uint8_t* dst, const uint16_t dstWidth, const uint16_t dstHeight, cons
         const uint8_t* srcRow = src + srcWidth * 4 * sy;
         for(uint16_t dx = xOffset, sx = 0; dx < dstWidth && sx < srcWidth; dx++, sx++)
         {
-            dstRow[dx * 3]     = (srcRow[sx * 4 + 1] * srcRow[sx * 4] + dstRow[dx * 3]     * (255 - srcRow[sx * 4])) / 255;
-            dstRow[dx * 3 + 1] = (srcRow[sx * 4 + 2] * srcRow[sx * 4] + dstRow[dx * 3 + 1] * (255 - srcRow[sx * 4])) / 255;
-            dstRow[dx * 3 + 2] = (srcRow[sx * 4 + 3] * srcRow[sx * 4] + dstRow[dx * 3 + 2] * (255 - srcRow[sx * 4])) / 255;
+            if(srcRow[sx * 4] != 0) // Alpha is either 0 or 255.
+            {
+                dstRow[dx * 3]     = srcRow[sx * 4 + 1];
+                dstRow[dx * 3 + 1] = srcRow[sx * 4 + 2];
+                dstRow[dx * 3 + 2] = srcRow[sx * 4 + 3];
+            }
         }
     }
 }

src/CDI/common/VideoU32.cpp

@@ -298,4 +298,33 @@ void decodeDYUVU32(Pixel* dst, const uint16_t pixel, uint32_t& previous) noexcep
     matrixRGBU32(&dst[1], y2, u2, v2);
 }
 
+/** \brief Copy the ARGB pixels to an ARGB plane.
+ *
+ * \param dst The destination ARGB plane.
+ * \param dstWidth The width of the destination plane.
+ * \param dstHeight The height of the destination plane.
+ * \param src The source ARGB plane.
+ * \param srcWidth The width of the source plane.
+ * \param srcHeight The width of the source plane.
+ * \param xOffset The x offset in pixels where the paste will occur on dst.
+ * \param yOffset The y offset in pixels where the paste will occur on dst.
+ *
+ * If the source does not fit in the destination, only the pixels that fit in the destination are copied.
+ */
+void paste(Pixel* dst, const uint16_t dstWidth, const uint16_t dstHeight, const Pixel* src, const uint16_t srcWidth, const uint16_t srcHeight, const uint16_t xOffset, const uint16_t yOffset)
+{
+    for(uint16_t dy = yOffset, sy = 0; dy < dstHeight && sy < srcHeight; ++dy, ++sy)
+    {
+              Pixel* dstRow = dst + dstWidth * dy;
+        const Pixel* srcRow = src + srcWidth * sy;
+        for(uint16_t dx = xOffset, sx = 0; dx < dstWidth && sx < srcWidth; ++dx, ++sx)
+        {
+            if((srcRow[sx] & 0xFF'00'00'00) != 0) // Alpha is either 0 or 255.
+            {
+                dstRow[dx] = srcRow[sx];
+            }
+        }
+    }
+}
+
 } // namespace Video

src/CDI/common/VideoU32.hpp

@@ -67,6 +67,8 @@ constexpr void decodeCLUTU32(Pixel* dst, const uint8_t pixel, const uint32_t* CL
     *dst = CLUTTable[pixel]; // We don't care about transparency for CLUT.
 }
 
+void paste(Pixel* dst, const uint16_t dstWidth, const uint16_t dstHeight, const Pixel* src, const uint16_t srcWidth, const uint16_t srcHeight, const uint16_t xOffset = 0, const uint16_t yOffset = 0);
+
 } // namespace Video
 
 #endif // CDI_COMMON_VIDEOU32_HPP