From 82b494203de76816eb5f9d2e32ddc46e71b9efe2 Mon Sep 17 00:00:00 2001
From: UnknownShadow200 <unknownshadow200@gmail.com>
Date: Thu, 8 May 2025 21:36:38 +1000
Subject: [PATCH] PS2: Slightly optimise drawing by using aligned data where
 possible

---
 misc/ps2/DrawColoured.S | 28 ++++------------------------
 misc/ps2/DrawTextured.S | 17 +++++------------
 src/Graphics_PS2.c      |  5 ++++-
 3 files changed, 13 insertions(+), 37 deletions(-)

diff --git a/misc/ps2/DrawColoured.S b/misc/ps2/DrawColoured.S
index e716a495b..d5235bec6 100644
--- a/misc/ps2/DrawColoured.S
+++ b/misc/ps2/DrawColoured.S
@@ -95,12 +95,7 @@
 FUNC DrawColouredQuad
 
 ### VERTEX 1 ###
-	# LOAD VERTEX 1
-	ld			$t0,0x00(SRC) # t0 = src[0].x,y
-	sd			$t0,0x00(TMP) # tmp.x,y = t0
-	lw			$t0,0x08(SRC) # t0 = src[0].z
-	sw			$t0,0x08(TMP) # tmp.z = t0
-	lqc2		POS_1, 0x00(TMP) # V1 = tmp
+	lqc2		POS_1, 0x00(SRC)
 
 	TransformVertex POS_1
 	lwc1		COL1, 0x0C(SRC)
@@ -108,12 +103,7 @@ FUNC DrawColouredQuad
 	BeginClip POS_1
 
 ### VERTEX 2 ###
-	# LOAD VERTEX 2
-	ld			$t0,0x10(SRC) # t0 = src[1].x,y
-	sd			$t0,0x00(TMP) # tmp.x,y = t0
-	lw			$t0,0x18(SRC) # t0 = src[1].z
-	sw			$t0,0x08(TMP) # tmp.z = t0
-	lqc2		POS_2, 0x00(TMP) # V2 = tmp
+	lqc2		POS_2, 0x10(SRC)
 
 	TransformVertex POS_2
 	lwc1		COL2,  0x1C(SRC)
@@ -122,12 +112,7 @@ FUNC DrawColouredQuad
 	BeginClip POS_2
 
 ### VERTEX 3 ###
-	# LOAD VERTEX 3
-	ld			$t0,0x20(SRC) # t0 = src[2].x,y
-	sd			$t0,0x00(TMP) # tmp.x,y = t0
-	lw			$t0,0x28(SRC) # t0 = src[2].z
-	sw			$t0,0x08(TMP) # tmp.z = t0
-	lqc2		POS_3, 0x00(TMP) # V3 = tmp
+	lqc2		POS_3, 0x20(SRC)
 
 	TransformVertex POS_3
 	lwc1		COL3,  0x2C(SRC)
@@ -136,12 +121,7 @@ FUNC DrawColouredQuad
 	BeginClip POS_3
 
 ### VERTEX 4 ###
-	# LOAD VERTEX 4
-	ld			$t0,0x30(SRC) # t0 = src[3].x,y
-	sd			$t0,0x00(TMP) # tmp.x,y = t0
-	lw			$t0,0x38(SRC) # t0 = src[3].z
-	sw			$t0,0x08(TMP) # tmp.z = t0
-	lqc2		POS_4, 0x00(TMP) # V4 = tmp
+	lqc2		POS_4, 0x30(SRC)
 
 	TransformVertex POS_4
 	lwc1		COL4,  0x3C(SRC)
diff --git a/misc/ps2/DrawTextured.S b/misc/ps2/DrawTextured.S
index ee214a415..926184c1d 100644
--- a/misc/ps2/DrawTextured.S
+++ b/misc/ps2/DrawTextured.S
@@ -105,12 +105,7 @@
 FUNC DrawTexturedQuad
 
 ### VERTEX 1 ###
-	# LOAD VERTEX 1
-	ld			$t0,0x00(SRC) # t0 = src[0].x,y
-	sd			$t0,0x00(TMP) # tmp.x,y = t0
-	lw			$t0,0x08(SRC) # t0 = src[0].z
-	sw			$t0,0x08(TMP) # tmp.z = t0
-	lqc2		POS_1, 0x00(TMP) # V1 = tmp
+	lqc2		POS_1, 0x00(SRC)
 
 	TransformVertex POS_1
 	lwc1		COL1, 0x0C(SRC)
@@ -131,22 +126,19 @@ FUNC DrawTexturedQuad
 	lwc1		COL2,  0x24(SRC)
 	lwc1		U_2,   0x28(SRC)
 	lwc1		V_2,   0x2C(SRC)
+
 	vmulq.w 	POS_1, V0001, $Q
 	vdiv	 	$Q,    _one, POS2w
 	BeginClip POS_2
 
 ### VERTEX 3 ###
-	# LOAD VERTEX 3
-	ld			$t0,0x30(SRC) # t0 = src[2].x,y
-	sd			$t0,0x00(TMP) # tmp.x,y = t0
-	lw			$t0,0x38(SRC) # t0 = src[2].z
-	sw			$t0,0x08(TMP) # tmp.z = t0
-	lqc2		POS_3, 0x00(TMP) # V3 = tmp
+	lqc2		POS_3, 0x30(SRC)
 
 	TransformVertex POS_3
 	lwc1		COL3,  0x3C(SRC)
 	lwc1		U_3,   0x40(SRC)
 	lwc1		V_3,   0x44(SRC)
+
 	vmulq.w 	POS_2, V0001, $Q
 	vdiv 		$Q,    _one, POS3w
 	BeginClip POS_3
@@ -163,6 +155,7 @@ FUNC DrawTexturedQuad
 	lwc1		COL4,  0x54(SRC)
 	lwc1		U_4,   0x58(SRC)
 	lwc1		V_4,   0x5C(SRC)
+
 	vmulq.w 	POS_3, V0001, $Q
 	vdiv 		$Q,    _one, POS4w
 	BeginClip POS_4
diff --git a/src/Graphics_PS2.c b/src/Graphics_PS2.c
index 0f49f5779..c56870e98 100644
--- a/src/Graphics_PS2.c
+++ b/src/Graphics_PS2.c
@@ -397,7 +397,10 @@ static void PreprocessColouredVertices(void) {
 }
 
 static GfxResourceID Gfx_AllocStaticVb(VertexFormat fmt, int count) {
-	return Mem_TryAlloc(count, strideSizes[fmt]);
+	//return Mem_TryAlloc(count, strideSizes[fmt]);
+	return memalign(16,count * strideSizes[fmt]);
+	// align to 16 bytes, so DrawTexturedQuad/DrawColouredQuad can
+	//  load vertices using the "load quad (16 bytes)" instruction
 }
 
 void Gfx_BindVb(GfxResourceID vb) { gfx_vertices = vb; }