diff --git a/draw.h b/draw.h
index 03a2582..27744ef 100644
--- a/draw.h
+++ b/draw.h
@@ -5,6 +5,28 @@
 #include "display.h"
 #include <inline_c.h>
 
+#define gte_ld_intpol_cv0( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+    "lbu	$14, 2( %0 );"	\
+	"ctc2	$12, $21;"		\
+	"ctc2	$13, $22;"		\
+    "ctc2	$14, $23;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
+#define gte_ld_intpol_cv1( r0 ) __asm__ volatile ( \
+	"lbu	$12, 0( %0 );"	\
+	"lbu	$13, 1( %0 );"	\
+    "lbu	$14, 2( %0 );"	\
+	"mtc2	$12, $9;"		\
+	"mtc2	$13, $10;"		\
+    "mtc2	$13, $11;"		\
+	:						\
+	: "r"( r0 )				\
+	: "$12", "$13", "$14" )
+
 // Macros for quickly blitting RGB and UV values with a single copy
 // This is faster than copying each value individually
 #define setColorFast(pr, r)         *((u_int*)(pr)) = *((u_int*)(r))
@@ -354,11 +376,14 @@ typedef struct _TMPVERT
     gte_ld_intpol_sv0(&(src1).vx); \
     gte_intpl(); \
     gte_stsv(&(dst).vx); \
-    (dst).r = (uint8_t)(((uint16_t)(src0).r + (uint16_t)(src1).r) >> 1); \
-    (dst).g = (uint8_t)(((uint16_t)(src0).g + (uint16_t)(src1).g) >> 1); \
-    (dst).b = (uint8_t)(((uint16_t)(src0).b + (uint16_t)(src1).b) >> 1); \
-    (dst).u = (uint8_t)(((uint16_t)(src0).u + (uint16_t)(src1).u) >> 1); \
-    (dst).v = (uint8_t)(((uint16_t)(src0).v + (uint16_t)(src1).v) >> 1);
+    gte_ld_intpol_cv1(&(src0).r); \
+    gte_ld_intpol_cv0(&(src1).r); \
+    gte_intpl(); \
+    gte_stcv(&(dst).r); \
+    gte_ld_intpol_uv1(&(src0).u); \
+    gte_ld_intpol_uv0(&(src1).u); \
+    gte_intpl(); \
+    gte_stbv(&(dst).u);
 
 #define blitVert(dst, i, src)  \
     blit32(&(dst)->x ## i, &(src).vx); \