Improve text rendering performance again by treating font sheets as combined

The way TGLP_s is set up, all of a font's texture sheets are adjacent in
memory and have the same size. We can reinterpet the memory to describe
a smaller set of much taller textures if we'd like. If we choose the
right size, we can get all of the ASCII glyphs under a single texture,
which will massively improve performance by reducing texture swaps
within a piece of all-English text down to 0! We don't need any extra
linear allocating to do this!

The coalescing will be applied to all characters / glyph sheets up until
the last `glyphInfo->nSheets % 32` sheets. This means that there are
more operations per glyph being done in `C2D_FontCalcGlyphPos`, but this
is probably offset by the savings from not switching textures as often.
And, this won't matter for English text, which has these results cached.

Author: TheGag96

source/font.c

@@ -5,6 +5,7 @@
 #include <c2d/font.h>
 
 fontGlyphPos_s g_systemFontASCIICache[128];
+u32 g_numFontSheetsCombined;
 
 C2D_Font C2D_FontLoad(const char* filename)
 {
@@ -245,9 +246,27 @@ charWidthInfo_s* C2D_FontGetCharWidthInfo(C2D_Font font, int glyphIndex)
 void C2D_FontCalcGlyphPos(C2D_Font font, fontGlyphPos_s* out, int glyphIndex, u32 flags, float scaleX, float scaleY)
 {
 	if (!font)
+	{
 		fontCalcGlyphPos(out, fontGetSystemFont(), glyphIndex, flags, scaleX, scaleY);
+
+		if (out->sheetIndex < g_numFontSheetsCombined)
+		{
+			u32 indexWithinBigSheet = out->sheetIndex % SHEETS_PER_BIG_SHEET;
+			out->sheetIndex /= SHEETS_PER_BIG_SHEET;
+
+			// Readjust glyph UVs to account for being a part of the combined texture.
+			out->texcoord.top    = (out->texcoord.top    + (SHEETS_PER_BIG_SHEET - indexWithinBigSheet - 1)) / (float) SHEETS_PER_BIG_SHEET;
+			out->texcoord.bottom = (out->texcoord.bottom + (SHEETS_PER_BIG_SHEET - indexWithinBigSheet - 1)) / (float) SHEETS_PER_BIG_SHEET;
+		}
+		else
+		{
+			out->sheetIndex = out->sheetIndex - g_numFontSheetsCombined + g_numFontSheetsCombined / SHEETS_PER_BIG_SHEET;
+		}
+	}
 	else
+	{
 		fontCalcGlyphPos(out, font->cfnt, glyphIndex, flags, scaleX, scaleY);
+	}
 }
 
 void C2D_FontCalcGlyphPosFromCodePoint(C2D_Font font, fontGlyphPos_s* out, u32 codepoint, u32 flags, float scaleX, float scaleY)

source/internal.h

@@ -122,3 +122,6 @@ void C2Di_Update(void);
 
 #define NUM_ASCII_CHARACTERS 128
 extern fontGlyphPos_s g_systemFontASCIICache[NUM_ASCII_CHARACTERS];
+
+#define SHEETS_PER_BIG_SHEET 32
+extern u32 g_numFontSheetsCombined;

source/text.c

@@ -57,6 +57,19 @@ static int C2Di_GlyphComp(const void* _g1, const void* _g2)
 	return ret;
 }
 
+static void fillSheet(C3D_Tex *tex, void *data, TGLP_s *glyphInfo)
+{
+	tex->data     = data;
+	tex->fmt      = glyphInfo->sheetFmt;
+	tex->size     = glyphInfo->sheetSize;
+	tex->width    = glyphInfo->sheetWidth;
+	tex->height   = glyphInfo->sheetHeight;
+	tex->param    = GPU_TEXTURE_MAG_FILTER(GPU_LINEAR) | GPU_TEXTURE_MIN_FILTER(GPU_LINEAR)
+		| GPU_TEXTURE_WRAP_S(GPU_CLAMP_TO_EDGE) | GPU_TEXTURE_WRAP_T(GPU_CLAMP_TO_EDGE);
+	tex->border   = 0;
+	tex->lodParam = 0;
+}
+
 static void C2Di_TextEnsureLoad(void)
 {
 	// Skip if already loaded
@@ -70,24 +83,32 @@ static void C2Di_TextEnsureLoad(void)
 	// Load the glyph texture sheets
 	CFNT_s* font = fontGetSystemFont();
 	TGLP_s* glyphInfo = fontGetGlyphInfo(font);
-	s_glyphSheets = malloc(sizeof(C3D_Tex)*glyphInfo->nSheets);
-	s_textScale = 30.0f / glyphInfo->cellHeight;
+
+	// The way TGLP_s is set up, all of a font's texture sheets are adjacent in memory and have the same size. We can
+	// reinterpet the memory to describe a smaller set of much taller textures if we'd like. If we choose the right size,
+	// we can get all of the ASCII glyphs under a single texture, which will massively improve performance by reducing
+	// texture swaps within a piece of all-English text down to 0! We don't need any extra linear allocating to do this!
+	u32 numSheetsBig   = glyphInfo->nSheets / SHEETS_PER_BIG_SHEET;
+	u32 numSheetsSmall = glyphInfo->nSheets % SHEETS_PER_BIG_SHEET;
+	u32 numSheetsTotal = numSheetsBig + numSheetsSmall;
+	g_numFontSheetsCombined = glyphInfo->nSheets - numSheetsSmall;
+
+	s_glyphSheets = malloc(sizeof(C3D_Tex)*numSheetsTotal);
 	if (!s_glyphSheets)
 		svcBreak(USERBREAK_PANIC);
-
-	int i;
-	for (i = 0; i < glyphInfo->nSheets; i ++)
+	memset(s_glyphSheets, 0, sizeof(sizeof(C3D_Tex)*numSheetsTotal));
+	s_textScale = 30.0f / glyphInfo->cellHeight;
+	for (u32 i = 0; i < numSheetsBig; i++)
 	{
 		C3D_Tex* tex = &s_glyphSheets[i];
-		tex->data = fontGetGlyphSheetTex(font, i);
-		tex->fmt = glyphInfo->sheetFmt;
-		tex->size = glyphInfo->sheetSize;
-		tex->width = glyphInfo->sheetWidth;
-		tex->height = glyphInfo->sheetHeight;
-		tex->param = GPU_TEXTURE_MAG_FILTER(GPU_LINEAR) | GPU_TEXTURE_MIN_FILTER(GPU_LINEAR)
-			| GPU_TEXTURE_WRAP_S(GPU_CLAMP_TO_BORDER) | GPU_TEXTURE_WRAP_T(GPU_CLAMP_TO_BORDER);
-		tex->border = 0;
-		tex->lodParam = 0;
+		fillSheet(tex, fontGetGlyphSheetTex(font, i * SHEETS_PER_BIG_SHEET), glyphInfo);
+		tex->height = (uint16_t) (tex->height * SHEETS_PER_BIG_SHEET);
+		tex->size   = tex->size * SHEETS_PER_BIG_SHEET;
+	}
+
+	for (u32 i = 0; i < numSheetsSmall; i++)
+	{
+		fillSheet(&s_glyphSheets[numSheetsBig + i], fontGetGlyphSheetTex(font, numSheetsBig * SHEETS_PER_BIG_SHEET + i), glyphInfo);
 	}
 
 	// Initialize system font ASCII cache for C2D_FontCalcGlyphPosFromCodePoint