From cdc205f59afa557e4d7f1fbf5f246106589bd348 Mon Sep 17 00:00:00 2001
From: Nico de Poel <ndepoel@gmail.com>
Date: Fri, 10 Feb 2023 12:25:13 +0100
Subject: [PATCH] First successful experiment with vertex interpolation using
 GTE intpl instructions.

---
 draw.h | 31 ++++++++++++++++++-------------
 1 file changed, 18 insertions(+), 13 deletions(-)

diff --git a/draw.h b/draw.h
index adb2c2a..03a2582 100644
--- a/draw.h
+++ b/draw.h
@@ -350,14 +350,15 @@ typedef struct _TMPVERT
     blit16(&(dst)->u, &(pv)->u);
 
 #define lerpVert(dst, src0, src1) \
-    (dst)->vx = (int16_t)(((int32_t)(src0)->vx + (int32_t)(src1)->vx) >> 1); \
-    (dst)->vy = (int16_t)(((int32_t)(src0)->vy + (int32_t)(src1)->vy) >> 1); \
-    (dst)->vz = (int16_t)(((int32_t)(src0)->vz + (int32_t)(src1)->vz) >> 1); \
-    (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_sv1(&(src0).vx); \
+    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);
 
 #define blitVert(dst, i, src)  \
     blit32(&(dst)->x ## i, &(src).vx); \
@@ -378,6 +379,7 @@ static INLINE void draw_quadstrip_tess2(const ps1bsp_vertex_t *vertices, const p
     i3 = head++;
 
     TMPVERT *tmp = (TMPVERT*)(scratchpad);
+    const short half = ONE >> 1;
    
     // Normally a quad strip would have (N-2)/2 quads, but we might end up with a sole triangle at the end which will be drawn as a collapsed quad
     // NOTE: testing has shown that the PS1 is faster just rendering quads and accepting the odd collapsed quad, rather than being clever with pointer comparisons and drawing a single triangle at the end.
@@ -404,12 +406,15 @@ static INLINE void draw_quadstrip_tess2(const ps1bsp_vertex_t *vertices, const p
         copyVertFast(&tmp[6], pv2, v2);
         copyVertFast(&tmp[8], pv3, v3);
 
+        // Interpolation factor will always be 0.5 so we need to set this register only once
+        gte_lddp(half);
+
         // 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]);
+        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]);
 
         // Transform the vertices in groups of three
         gte_ldv3(&tmp[0], &tmp[1], &tmp[2]);