From 990df72ced66c4fa4a1f51684517bf4204c06a63 Mon Sep 17 00:00:00 2001
From: Jeffrey Pfau <jeffrey@endrift.com>
Date: Thu, 11 Aug 2016 03:03:49 -0700
Subject: [PATCH] GBA Video: Optimize sprite drawing

---
 CHANGES                          |  1 +
 src/gba/renderers/software-bg.c  |  7 +++---
 src/gba/renderers/software-obj.c | 37 ++++++++++++++++++++++++++++----
 3 files changed, 37 insertions(+), 8 deletions(-)

diff --git a/CHANGES b/CHANGES
index b6173d637..bbdeb9513 100644
--- a/CHANGES
+++ b/CHANGES
@@ -49,6 +49,7 @@ Misc:
  - GUI: Add back logging
  - PSP2: Add rumble for PS TV
  - PSP2: Add fit-to-height screen mode
+ - GBA Video: Optimize sprite drawing
 
 0.4.1: (2016-07-11)
 Bugfixes:
diff --git a/src/gba/renderers/software-bg.c b/src/gba/renderers/software-bg.c
index fec27af04..67d4dc26c 100644
--- a/src/gba/renderers/software-bg.c
+++ b/src/gba/renderers/software-bg.c
@@ -14,10 +14,9 @@
 #define MODE_2_COORD_NO_OVERFLOW \
 	if ((x | y) & ~(sizeAdjusted - 1)) { \
 		continue; \
-	} else { \
-		localX = x; \
-		localY = y; \
-	}
+	} \
+	localX = x; \
+	localY = y;
 
 #define MODE_2_MOSAIC(COORD) \
 		if (!mosaicWait) { \
diff --git a/src/gba/renderers/software-obj.c b/src/gba/renderers/software-obj.c
index 9179de85e..91b626499 100644
--- a/src/gba/renderers/software-obj.c
+++ b/src/gba/renderers/software-obj.c
@@ -56,7 +56,7 @@
 		int localY = (yAccum >> 8) + (height >> 1); \
 		\
 		if (localX & widthMask || localY & heightMask) { \
-			continue; \
+			break; \
 		} \
 		\
 		SPRITE_YBASE_ ## DEPTH(localY); \
@@ -204,6 +204,35 @@ int GBAVideoSoftwareRendererPreprocessSprite(struct GBAVideoSoftwareRenderer* re
 			condition = end;
 		}
 
+		// Clip off early pixels
+		if ((xAccum >> 8) < -(width >> 1) || (yAccum >> 8) < -(height >> 1)) {
+			int32_t diffX = -(width << 7) - xAccum;
+			int32_t diffY = -(height << 7) - yAccum;
+			int32_t x = mat.a ? diffX / mat.a : 0;
+			int32_t y = mat.c ? diffY / mat.c : 0;
+			if (y > x) {
+				x = y;
+			}
+			xAccum += mat.a * x;
+			yAccum += mat.c * x;
+			outX += x;
+			inX += x;
+		}
+		// Clip off late pixels
+		if ((xAccum >> 8) >= (width >> 1) || (yAccum >> 8) >= (height >> 1)) {
+			int32_t diffX = (width << 7) - xAccum;
+			int32_t diffY = (height << 7) - yAccum;
+			int32_t x = mat.a ? diffX / mat.a : 0;
+			int32_t y = mat.c ? diffY / mat.c : 0;
+			if (y > x) {
+				x = y;
+			}
+			xAccum += mat.a * x;
+			yAccum += mat.c * x;
+			outX += x;
+			inX += x;
+		}
+
 		if (!GBAObjAttributesAIs256Color(sprite->a)) {
 			palette = &palette[GBAObjAttributesCGetPalette(sprite->c) << 4];
 			if (flags & FLAG_OBJWIN) {
@@ -230,7 +259,7 @@ int GBAVideoSoftwareRendererPreprocessSprite(struct GBAVideoSoftwareRenderer* re
 		int outX = x >= start ? x : start;
 		int condition = x + width;
 		int mosaicH = 1;
-		if (GBAObjAttributesAIsMosaic(sprite->a)) {
+		if (mosaicH > 1) {
 			mosaicH = GBAMosaicControlGetObjH(renderer->mosaic) + 1;
 			if (condition % mosaicH) {
 				condition += mosaicH - (condition % mosaicH);
@@ -255,7 +284,7 @@ int GBAVideoSoftwareRendererPreprocessSprite(struct GBAVideoSoftwareRenderer* re
 			palette = &palette[GBAObjAttributesCGetPalette(sprite->c) << 4];
 			if (flags & FLAG_OBJWIN) {
 				SPRITE_NORMAL_LOOP(16, OBJWIN);
-			} else if (GBAObjAttributesAIsMosaic(sprite->a)) {
+			} else if (mosaicH > 1) {
 				if (objwinSlowPath) {
 					objwinPalette = &objwinPalette[GBAObjAttributesCGetPalette(sprite->c) << 4];
 					SPRITE_MOSAIC_LOOP(16, NORMAL_OBJWIN);
@@ -271,7 +300,7 @@ int GBAVideoSoftwareRendererPreprocessSprite(struct GBAVideoSoftwareRenderer* re
 		} else {
 			if (flags & FLAG_OBJWIN) {
 				SPRITE_NORMAL_LOOP(256, OBJWIN);
-			} else if (GBAObjAttributesAIsMosaic(sprite->a)) {
+			} else if (mosaicH > 1) {
 				if (objwinSlowPath) {
 					objwinPalette = &objwinPalette[GBAObjAttributesCGetPalette(sprite->c) << 4];
 					SPRITE_MOSAIC_LOOP(256, NORMAL_OBJWIN);