From ae3ac8cdb12b72fc2df32ab360d8e4746a74f587 Mon Sep 17 00:00:00 2001
From: UnknownShadow200 <unknownshadow200@gmail.com>
Date: Sat, 27 Apr 2024 23:24:31 +1000
Subject: [PATCH] NDS: Optimise rendering to be a little bit faster

---
 src/Graphics_NDS.c | 25 +++++++++++++++++--------
 src/Window_NDS.c   | 20 ++++++++++----------
 2 files changed, 27 insertions(+), 18 deletions(-)

diff --git a/src/Graphics_NDS.c b/src/Graphics_NDS.c
index c88b1e370..ba081eaa5 100644
--- a/src/Graphics_NDS.c
+++ b/src/Graphics_NDS.c
@@ -468,19 +468,28 @@ static void Draw_ColouredTriangles(int verticesCount, int startVertex) {
 
 static void Draw_TexturedTriangles(int verticesCount, int startVertex) {
 	glBegin(GL_QUADS);
-    int width = 0, height = 0;
-    glGetInt(GL_GET_TEXTURE_WIDTH,  &width);  width  = inttof32(width);
-    glGetInt(GL_GET_TEXTURE_HEIGHT, &height); height = inttof32(height);
-    
+	int width = 0, height = 0;
+	glGetInt(GL_GET_TEXTURE_WIDTH,  &width);
+	glGetInt(GL_GET_TEXTURE_HEIGHT, &height);
+
+	// Original code used was
+	//   U = mulf32(v->u, inttof32(width))
+	// which behind the scenes expands to
+	//   W = width << 12
+	//   U = ((int64)v->u * W) >> 12;
+	// and in this case, the bit shifts can be cancelled out
+	//  to avoid calling __aeabi_lmul to perform the 64 bit multiplication
+	// therefore the code can be simplified to
+	//   U = v->u * width
 
 	for (int i = 0; i < verticesCount; i++)
 	{
 		struct DSTexturedVertex* v = (struct DSTexturedVertex*)gfx_vertices + startVertex + i;
 		
-		GFX_COLOR    = v->rgb;
-        glTexCoord2t16(f32tot16(mulf32(v->u, width)), f32tot16(mulf32(v->v, height)));
-		GFX_VERTEX16 = v->xy;
-        GFX_VERTEX16 = v->z;
+		GFX_COLOR     = v->rgb;
+		GFX_TEX_COORD = TEXTURE_PACK(f32tot16(v->u * width), f32tot16(v->v * height));
+		GFX_VERTEX16  = v->xy;
+		GFX_VERTEX16  = v->z;
 	}
 	glEnd();
 }
diff --git a/src/Window_NDS.c b/src/Window_NDS.c
index 208443392..0ea5a6379 100644
--- a/src/Window_NDS.c
+++ b/src/Window_NDS.c
@@ -30,7 +30,7 @@ extern u8 default_fontTiles[];
 #define FONT_ASCII_OFFSET 32
 static u16* conFontBgMap;
 static int  conFontCurPal;
-static int  conCursorX, conCursorY;
+static int  conCursorX, conCurrentRow;
 
 static void consoleClear(void) {
     for (int i = 0; i < CON_WIDTH * CON_HEIGHT; i++)
@@ -38,17 +38,17 @@ static void consoleClear(void) {
         conFontBgMap[i] = ' ' - FONT_ASCII_OFFSET;
     }
 
-    conCursorX = 0;
-    conCursorY = 0;
+    conCursorX    = 0;
+    conCurrentRow = 0;
 }
 
 static void consoleNewLine(void) {
     conCursorX = 0;
-    conCursorY++;
-    if (conCursorY < CON_HEIGHT) return;
+    conCurrentRow++;
+    if (conCurrentRow < CON_HEIGHT) return;
 
     // Shift entire screen upwards by one row
-    conCursorY--;
+    conCurrentRow--;
 
     for (int y = 0; y < CON_HEIGHT - 1; y++)
     {
@@ -72,13 +72,12 @@ static void consolePrintChar(char c) {
         consoleNewLine();
 
     u16 value = conFontCurPal | (c - FONT_ASCII_OFFSET);
-    conFontBgMap[conCursorX + conCursorY * CON_WIDTH] = value;
+    conFontBgMap[conCursorX + conCurrentRow * CON_WIDTH] = value;
     conCursorX++;
 }
 
 void consolePrintString(const char* ptr, int len) {
 	if (!conFontBgMap) return;
-	consoleClear();
 
     for (int i = 0; i < len; i++)
     {
@@ -103,7 +102,7 @@ static void consoleLoadFont(u16* fontBgGfx) {
         if (row & 0x20) gfx |= 0x00F00000;
         if (row & 0x40) gfx |= 0x0F000000;
         if (row & 0x80) gfx |= 0xF0000000;
-        ((u32 *)conFontBgGfx)[i] = gfx;
+        ((u32 *)fontBgGfx)[i] = gfx;
     }
 
     palette[16 * 16 - 1] = RGB15(31, 31, 31);
@@ -328,9 +327,10 @@ void OnscreenKeyboard_Draw3D(void) { }
 
 void OnscreenKeyboard_Close(void) {
     keyboardHide();
+	if (!keyboardOpen) return;
     keyboardOpen = false;
-    ResetHBank(); // reset shared VRAM
 
+    ResetHBank(); // reset shared VRAM
     videoBgEnableSub(0); // show console
     consoleInit();
 }