From 7888f44bb8361afd3c69a860af65019cf5ef1860 Mon Sep 17 00:00:00 2001
From: Nico de Poel <ndepoel@gmail.com>
Date: Fri, 10 Feb 2023 10:25:41 +0100
Subject: [PATCH] Revert "Intersperse GTE instructions with vertex lerping and
 quad building code to mask load/store delays."

This reverts commit 4f886bd40c6955ff0e4390285b05e8938c1c3b18.

Rolling back to simpler code to facilitate conversion to GTE interpolation.
---
 draw.h | 29 +++++++++++++++--------------
 1 file changed, 15 insertions(+), 14 deletions(-)

diff --git a/draw.h b/draw.h
index d224b86..adb2c2a 100644
--- a/draw.h
+++ b/draw.h
@@ -404,24 +404,34 @@ static INLINE void draw_quadstrip_tess2(const ps1bsp_vertex_t *vertices, const p
         copyVertFast(&tmp[6], pv2, v2);
         copyVertFast(&tmp[8], pv3, v3);
 
-        // Interpolate vertices and start loading them into GTE as soon as we're done with them
-        // This way we avoid waiting for the GTE load/store delays
+        // 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]);
+
+        // Transform the vertices in groups of three
         gte_ldv3(&tmp[0], &tmp[1], &tmp[2]);
         gte_rtpt();
-        lerpVert(&tmp[3], &tmp[0], &tmp[6]);
         gte_stsxy0(&tmp[0].vx);
         gte_stsxy1(&tmp[1].vx);
         gte_stsxy2(&tmp[2].vx);
-        lerpVert(&tmp[4], &tmp[3], &tmp[5]);    // After this, 3 4 5 are ready for GTE
+
         gte_ldv3(&tmp[3], &tmp[4], &tmp[5]);
         gte_rtpt();
-        lerpVert(&tmp[7], &tmp[6], &tmp[8]);
         gte_stsxy0(&tmp[3].vx);
         gte_stsxy1(&tmp[4].vx);
         gte_stsxy2(&tmp[5].vx);
 
+        gte_ldv3(&tmp[6], &tmp[7], &tmp[8]);
+        gte_rtpt();
+        gte_stsxy0(&tmp[6].vx);
+        gte_stsxy1(&tmp[7].vx);
+        gte_stsxy2(&tmp[8].vx);
+
+        // TODO Optimization: interperse quad building instructions while waiting on GTE store/load delays
+
         // Draw the first quad
         p0 = (POLY_GT4*)mem_prim(sizeof(POLY_GT4));
         blitVert(p0, 0, tmp[0]);
@@ -432,10 +442,6 @@ static INLINE void draw_quadstrip_tess2(const ps1bsp_vertex_t *vertices, const p
         blitVert(p0, 3, tmp[4]);
         setPolyGT4(p0);
         addPrim(ot, p0);
-        
-        // Transform the final three vertices, needed for the final two quads
-        gte_ldv3(&tmp[6], &tmp[7], &tmp[8]);
-        gte_rtpt();
 
         // Second quad
         p1 = (POLY_GT4*)mem_prim(sizeof(POLY_GT4));
@@ -448,11 +454,6 @@ static INLINE void draw_quadstrip_tess2(const ps1bsp_vertex_t *vertices, const p
         setPolyGT4(p1);
         addPrim(ot, p1);
 
-        // Make sure the transformed vertices are loaded from GTE before we need them
-        gte_stsxy0(&tmp[6].vx);
-        gte_stsxy1(&tmp[7].vx);
-        gte_stsxy2(&tmp[8].vx);
-        
         // Third quad
         p2 = (POLY_GT4*)mem_prim(sizeof(POLY_GT4));
         blitVert(p2, 0, tmp[3]);