Made a custom vertex lerp macro using pure GTE assembly code

draw.h

@@ -398,6 +398,68 @@ typedef struct _TMPVERT
     gte_intpl(); \
     gte_stbv(&(dst).u);
 
+#define lerpVert2(dst, src0, src1) __asm__ volatile ( \
+	"lh	$12, 0( %1 );"		\
+	"lh	$13, 2( %1 );"		\
+	"lh	$14, 4( %1 );"		\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"mtc2	$14, $11;"		\
+	"lh	$12, 0( %2 );"		\
+	"lh	$13, 2( %2 );"		\
+	"lh	$14, 4( %2 );"		\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+	"ctc2	$14, $23;"		\
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0980011;"       \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"mfc2	$14, $11;"		\
+	"sh	$12, 0( %0 );"		\
+	"sh	$13, 2( %0 );"		\
+	"sh	$14, 4( %0 );"		\
+	"lbu	$12, 8( %1 );"	\
+	"lbu	$13, 9( %1 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+	"lbu	$12, 8( %2 );"	\
+	"lbu	$13, 9( %2 );"	\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0980011;"       \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+	"sb	$12, 8( %0 );"		\
+	"sb	$13, 9( %0 );"		\
+	"lbu	$12, 10( %1 );"	\
+	"lbu	$13, 11( %1 );"	\
+    "lbu	$14, 12( %1 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+    "mtc2	$14, $11;"		\
+	"lbu	$12, 10( %2 );"	\
+	"lbu	$13, 11( %2 );"	\
+    "lbu	$14, 12( %2 );"	\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+    "ctc2	$14, $23;"		\
+	"nop;"					\
+	"nop;"					\
+	"cop2 0x0980011;"       \
+	"mfc2	$12, $9;"		\
+	"mfc2	$13, $10;"		\
+    "mfc2	$14, $11;"		\
+	"sb	$12, 10( %0 );"		\
+	"sb	$13, 11( %0 );"		\
+    "sb	$14, 12( %0 );"		\
+	:						\
+	: "r"(dst), "r"(src0), "r"(src1)	\
+	: "$12", "$13", "$14", "memory" )
+
 #define blitVert(dst, i, src)  \
     blit32(&(dst)->x ## i, &(src).vx); \
     blit32(&(dst)->r ## i, &(src).r); \
@@ -447,11 +509,11 @@ static INLINE void draw_quadstrip_tess2(const ps1bsp_vertex_t *vertices, const p
         copyVertFast(tmp[8], pv3, v3);
 
         // TODO Optimization: start loading vertices into GTE as soon as we're done with them
-        lerpVert(tmp[1], tmp[0], tmp[2]);
-        lerpVert(tmp[5], tmp[2], tmp[8]);    // After this, 0 1 2 are ready for GTE
-        lerpVert(tmp[3], tmp[0], tmp[6]);
-        lerpVert(tmp[4], tmp[3], tmp[5]);    // After this, 3 4 5 are ready for GTE
-        lerpVert(tmp[7], tmp[6], tmp[8]);
+        lerpVert2(&tmp[1], &tmp[0], &tmp[2]);
+        lerpVert2(&tmp[5], &tmp[2], &tmp[8]);    // After this, 0 1 2 are ready for GTE
+        lerpVert2(&tmp[3], &tmp[0], &tmp[6]);
+        lerpVert2(&tmp[4], &tmp[3], &tmp[5]);    // After this, 3 4 5 are ready for GTE
+        lerpVert2(&tmp[7], &tmp[6], &tmp[8]);
 
         // Transform the vertices in groups of three
         gte_ldv3(&tmp[0], &tmp[1], &tmp[2]);