GS: Add slow path for odd width 4bpp host->local transfers

pcsx2/GS/GSLocalMemory.cpp

@@ -815,8 +815,8 @@ void GSLocalMemory::WriteImage(int& tx, int& ty, const u8* src, int len, GIFRegB
 	if (TRXREG.RRW == 0)
 		return;
 
-	int l = (int)TRXPOS.DSAX;
-	int r = l + (int)TRXREG.RRW;
+	const int l = (int)TRXPOS.DSAX;
+	const int r = l + (int)TRXREG.RRW;
 
 	// finish the incomplete row first
 
@@ -828,12 +828,32 @@ void GSLocalMemory::WriteImage(int& tx, int& ty, const u8* src, int len, GIFRegB
 		len -= n;
 	}
 
-	int la = (l + (bsx - 1)) & ~(bsx - 1);
-	int ra = r & ~(bsx - 1);
+	const int la = (l + (bsx - 1)) & ~(bsx - 1);
+	const int ra = r & ~(bsx - 1);
 	// Round up to the nearest byte (NFL 2K5 does r = 1, l = 0 bpp =4, causing divide by zero)
-	int srcpitch = (((r - l) * trbpp) + 7) >> 3;
+	const int srcpitch = (((r - l) * trbpp) + 7) >> 3;
 	int h = len / srcpitch;
 
+	// Slow path for odd width 4bpp, the fast path expects everything to be perfectly aligned and great,
+	// but things get hairy with 4bpp pixels and odd widths since the lowest size we can address is 8bits, it goes out of sync.
+	// Although I call this a slow path, it's probably faster than modifying the data alignment every other line.
+	// GT3 demo, Jak 2 Japanese subtitles, and the BG Dark Alliance minimap do this.
+	if (trbpp == 4 && (TRXREG.RRW & 0x1))
+	{
+		int count = 0;
+		const int t = TRXPOS.DSAY;
+		const int b = t + (int)TRXREG.RRH;
+		for (int y = t; y < b; y++)
+		{
+			for (int x = l; x < r; x++)
+			{
+				WritePixel4(x, y, src[count >> 1] >> ((count & 1) << 2), BITBLTBUF.DBP, BITBLTBUF.DBW);
+				count++;
+			}
+		}
+		return;
+	}
+
 	if (ra - la >= bsx && h > 0) // "transfer width" >= "block width" && there is at least one full row
 	{
 		const u8* s = &src[-l * trbpp >> 3];