diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index f280a6cdb..cb07139bb 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -851,18 +851,18 @@ void RasterizerOpenGL::ReloadColorBuffer() {
     using VideoCore::CopyTextureAndUntile;
     switch (bytes_per_pixel) {
     case 4:
-        CopyTextureAndUntile<u32>(reinterpret_cast<u32*>(color_buffer),
-                reinterpret_cast<u32*>(temp_fb_color_buffer.get()),
+        CopyTextureAndUntile<4>(color_buffer,
+                temp_fb_color_buffer.get(),
                 fb_color_texture.width, fb_color_texture.height);
         break;
     case 3:
-        CopyTextureAndUntile<u24_be>(reinterpret_cast<u24_be*>(color_buffer),
-                reinterpret_cast<u24_be*>(temp_fb_color_buffer.get()),
+        CopyTextureAndUntile<3>(color_buffer,
+                temp_fb_color_buffer.get(),
                 fb_color_texture.width, fb_color_texture.height);
         break;
     case 2:
-        CopyTextureAndUntile<u16>(reinterpret_cast<u16*>(color_buffer),
-                reinterpret_cast<u16*>(temp_fb_color_buffer.get()),
+        CopyTextureAndUntile<2>(color_buffer,
+                temp_fb_color_buffer.get(),
                 fb_color_texture.width, fb_color_texture.height);
         break;
     default:
@@ -974,18 +974,18 @@ void RasterizerOpenGL::CommitColorBuffer() {
             using VideoCore::CopyTextureAndTile;
             switch (bytes_per_pixel) {
             case 4:
-                CopyTextureAndTile<u32>(reinterpret_cast<u32*>(temp_gl_color_buffer.get()),
-                        reinterpret_cast<u32*>(color_buffer),
+                CopyTextureAndTile<4>(temp_gl_color_buffer.get(),
+                        color_buffer,
                         fb_color_texture.width, fb_color_texture.height);
                 break;
             case 3:
-                CopyTextureAndTile<u24_be>(reinterpret_cast<u24_be*>(temp_gl_color_buffer.get()),
-                        reinterpret_cast<u24_be*>(color_buffer),
+                CopyTextureAndTile<3>(temp_gl_color_buffer.get(),
+                        color_buffer,
                         fb_color_texture.width, fb_color_texture.height);
                 break;
             case 2:
-                CopyTextureAndTile<u16>(reinterpret_cast<u16*>(temp_gl_color_buffer.get()),
-                        reinterpret_cast<u16*>(color_buffer),
+                CopyTextureAndTile<2>(temp_gl_color_buffer.get(),
+                        color_buffer,
                         fb_color_texture.width, fb_color_texture.height);
                 break;
             default:
diff --git a/src/video_core/utils.cpp b/src/video_core/utils.cpp
index a8c3afc7f..9c2929d10 100644
--- a/src/video_core/utils.cpp
+++ b/src/video_core/utils.cpp
@@ -34,15 +34,15 @@ void DumpTGA(std::string filename, short width, short height, u8* raw_data) {
     fclose(fout);
 }
 
-template<typename T>
-void CopyTextureAndTile(const T* src, T* dst, unsigned int width, unsigned int height) {
+template<int size>
+void CopyTextureAndTile(const u8* src, u8* dst, unsigned int width, unsigned int height) {
     for (unsigned int y = 0; y + 8 <= height; y += 8) {
         for (unsigned int x = 0; x + 8 <= width; x += 8) {
-            const T* line = &src[y * width + x];
+            const u8* line = &src[y * width + x];
 
             for (unsigned int yy = 0; yy < 8; ++yy) {
                 for (unsigned int xx = 0; xx < 8; ++xx) {
-                    dst[morton_lut[yy * 8 + xx]] = line[xx];
+                    memcpy(dst + size * morton_lut[yy * 8 + xx], line + size * xx, size);
                 }
                 line += width;
             }
@@ -52,19 +52,19 @@ void CopyTextureAndTile(const T* src, T* dst, unsigned int width, unsigned int h
     }
 }
 
-template void CopyTextureAndTile<u16>(const u16* src, u16* dst, unsigned int width, unsigned int height);
-template void CopyTextureAndTile<u24_be>(const u24_be* src, u24_be* dst, unsigned int width, unsigned int height);
-template void CopyTextureAndTile<u32>(const u32* src, u32* dst, unsigned int width, unsigned int height);
+template void CopyTextureAndTile<2>(const u8* src, u8* dst, unsigned int width, unsigned int height);
+template void CopyTextureAndTile<3>(const u8* src, u8* dst, unsigned int width, unsigned int height);
+template void CopyTextureAndTile<4>(const u8* src, u8* dst, unsigned int width, unsigned int height);
 
-template<typename T>
-void CopyTextureAndUntile(const T* src, T* dst, unsigned int width, unsigned int height) {
+template<int size>
+void CopyTextureAndUntile(const u8* src, u8* dst, unsigned int width, unsigned int height) {
     for (unsigned int y = 0; y + 8 <= height; y += 8) {
         for (unsigned int x = 0; x + 8 <= width; x += 8) {
-            T* line = &dst[y * width + x];
+            u8* line = &dst[y * width + x];
 
             for (unsigned int yy = 0; yy < 8; ++yy) {
                 for (unsigned int xx = 0; xx < 8; ++xx) {
-                    line[xx] = src[morton_lut[yy * 8 + xx]];
+                    memcpy(line + size * xx, src + size * morton_lut[yy * 8 + xx], size);
                 }
                 line += width;
             }
@@ -74,8 +74,8 @@ void CopyTextureAndUntile(const T* src, T* dst, unsigned int width, unsigned int
     }
 }
 
-template void CopyTextureAndUntile<u16>(const u16* src, u16* dst, unsigned int width, unsigned int height);
-template void CopyTextureAndUntile<u24_be>(const u24_be* src, u24_be* dst, unsigned int width, unsigned int height);
-template void CopyTextureAndUntile<u32>(const u32* src, u32* dst, unsigned int width, unsigned int height);
+template void CopyTextureAndUntile<2>(const u8* src, u8* dst, unsigned int width, unsigned int height);
+template void CopyTextureAndUntile<3>(const u8* src, u8* dst, unsigned int width, unsigned int height);
+template void CopyTextureAndUntile<4>(const u8* src, u8* dst, unsigned int width, unsigned int height);
 
 } // namespace
diff --git a/src/video_core/utils.h b/src/video_core/utils.h
index 812d3818f..2827e5b7a 100644
--- a/src/video_core/utils.h
+++ b/src/video_core/utils.h
@@ -60,28 +60,26 @@ static inline u32 MortonInterleave(u32 x, u32 y) {
  * Copies the texture data from the source address to the destination address, applying a
  * Morton-order transformation while copying.
  *
- * @param T Type of the source and destination pointers, the swizzling process depends on the size
- *        of this parameter.
+ * @param T Byte size of a pixel in the source and destination pointers.
  * @param src Pointer to the source texture data.
  * @param dst Pointer to which the texture will be copied.
  * @param width Width of the texture, should be a multiple of 8.
  * @param height Height of the texture, should be a multiple of 8.
  */
-template<typename T>
-void CopyTextureAndTile(const T* src, T* dst, unsigned int width, unsigned int height);
+template<int size>
+void CopyTextureAndTile(const u8* src, u8* dst, unsigned int width, unsigned int height);
 
 /**
  * Copies texture data while undoing the transformation applied by `CopyTextureAndTile`.
  *
- * @param T Type of the source and destination pointers, the swizzling process depends on the size
- *        of this parameter.
+ * @param T Byte size of a pixel in the source and destination pointers.
  * @param src Pointer to the source texture data.
  * @param dst Pointer to which the texture will be copied.
  * @param width Width of the texture, should be a multiple of 8.
  * @param height Height of the texture, should be a multiple of 8.
  */
-template<typename T>
-void CopyTextureAndUntile(const T* src, T* dst, unsigned int width, unsigned int height);
+template<int size>
+void CopyTextureAndUntile(const u8* src, u8* dst, unsigned int width, unsigned int height);
 
 /**
  * Calculates the offset of the position of the pixel in Morton order