From 2217a34d3912d71a87e91cf90b04fb49cd329808 Mon Sep 17 00:00:00 2001
From: Jaklyy <102590697+Jaklyy@users.noreply.github.com>
Date: Thu, 14 Dec 2023 23:00:12 -0500
Subject: [PATCH] misc improvements

---
 src/GPU3D.h        | 19 ++++++++++++-------
 src/GPU3D_Soft.cpp | 30 +++++++++++++++++-------------
 2 files changed, 29 insertions(+), 20 deletions(-)

diff --git a/src/GPU3D.h b/src/GPU3D.h
index 57553782..2dfacdc0 100644
--- a/src/GPU3D.h
+++ b/src/GPU3D.h
@@ -341,15 +341,20 @@ public:
 
     // GPU 3D rasterization timings, for emulating the timeout
     static constexpr int ScanlinePairLength = 2130 * TimingFrac;
-    static constexpr int ScanlineTimeout = 1686 * TimingFrac; // 2126? 1686?
-    static constexpr int ScanlineBreak = 4 * TimingFrac;
-    static constexpr int ScanlineBreak2 = 40 * TimingFrac;
-    static constexpr int IncrementStrange = 1618 * TimingFrac; // 1882? 1442? 1618??
-    static constexpr int FreeTiming = 440 * TimingFrac;
+    //static constexpr int ScanlineTimeout = 1686 * TimingFrac; // 2126? 1686?
+    //static constexpr int ScanlineBreak = 4 * TimingFrac;
+    //static constexpr int ScanlineBreak2 = 40 * TimingFrac;
+    static constexpr int ScanlineIncrement = 1618 * TimingFrac; // how much to increment per scanline pair
+    static constexpr int AbortIncrement = 12 * TimingFrac; // how much extra to increment after an aborted scanline (total 1630)
+    static constexpr int FreeTiming = 496 * TimingFrac; // every scanline has a free 496 pixels worth of timing for some reason.
+    static constexpr int InitialTiming = 48688 * TimingFrac; // add 1618*2 to get the timeout of the second scanline pair
+    static constexpr int Post50Max = 51116 * TimingFrac; // for some reason it doesn't care about how full it actually is, it just cares about if its the first 50 scanlines to speedrun rendering?
 
     // GPU 3D rasterization timings II, for counting each element with timing characteristics
-    static constexpr int PerPolyTiming = 12 * TimingFrac; // should be correct for *most* line polygons and polygons with vertical slopes
-    static constexpr int PerPixelTiming = 1 * TimingFrac; // does not apply to the first 4 pixels in a polygon (per scanline?)
+    static constexpr int FirstPolyScanline = 0 * TimingFrac; 
+    static constexpr int PerPolyScanline = 12 * TimingFrac; // should be correct for *most* line polygons and polygons with vertical slopes
+    static constexpr int PerPixelTiming = 1 * TimingFrac; // 1 pixel = 1 pixel
+    static constexpr int NumFreePixels = 4; // First 4 pixels in a polygon scanline are free (for some reason)
 
    // static constexpr int RasterTimingCap = 51116 * TimingFrac;
     static constexpr int PerScanlineTiming = 1064 * TimingFrac; // approximate currently, used to calc RDLines. TEMPORARY UNTIL ACCURATE "FRAMEBUFFER" CAN BE IMPLEMENTED
diff --git a/src/GPU3D_Soft.cpp b/src/GPU3D_Soft.cpp
index 7cb8002a..af23132d 100644
--- a/src/GPU3D_Soft.cpp
+++ b/src/GPU3D_Soft.cpp
@@ -993,7 +993,7 @@ bool SoftRenderer::RenderPolygonScanline(RendererPolygon* rp, s32 y, bool odd)
 
     CheckSlope(rp, y);
     
-    if (DoTimings(PerPolyTiming, odd))
+    if (DoTimings(PerPolyScanline, odd))
     {
         Step(rp);
         return true;
@@ -1439,18 +1439,19 @@ bool SoftRenderer::RenderScanline(s32 y, int npolys, bool odd)
         RendererPolygon* rp = &PolygonList[i];
         Polygon* polygon = rp->PolyData;
 
-        if (abort)
-        {
-            CheckSlope(rp, y);
-            Step(rp);
-        }
-        else if (y == polygon->YBottom && y != polygon->YTop)
+        if (y == polygon->YBottom && y != polygon->YTop)
         {
             if (DoTimings(EmptyPolyScanline, odd)) abort = true;
         }
         else if (y >= polygon->YTop && (y < polygon->YBottom || (y == polygon->YTop && polygon->YBottom == polygon->YTop)))
         {
-            if (polygon->IsShadowMask)
+            if (y == polygon->YTop) if(DoTimings(FirstPolyScanline, odd)) abort = true;
+            if (abort)
+            {
+                CheckSlope(rp, y);
+                Step(rp);
+            }
+            else if (polygon->IsShadowMask)
                 ;//RenderShadowMaskScanline(rp, y);
             else
                 if (RenderPolygonScanline(rp, y, odd)) abort = true;
@@ -1768,17 +1769,18 @@ void SoftRenderer::RenderPolygons(bool threaded, Polygon** polygons, int npolys)
             RasterTimingOdd = 0;
             RasterTimingEven = 0;
             
-            if (buffersize > 48)
+            RasterTiming += ScanlineIncrement;
+            if (abort) RasterTiming += AbortIncrement; // if previous scanline was aborted, allow an extra 12 pixels worth of timing
+            
+            if (y >= 50)
             {
-                RasterTiming = ScanlinePairLength * 23;
+                if (RasterTiming > Post50Max) RasterTiming = Post50Max;
                 timingadvance = 0;
                 buffersize = 48;
             }
-            if (!abort) RasterTiming += IncrementStrange;
-            else       RasterTiming += ScanlineTimeout;
 
             abort = RenderScanline(y, j, true);
-            abort = RenderScanline(y+1, j, false);
+            abort |= RenderScanline(y+1, j, false);
 
             buffersize += 2;
             //RasterTiming += ScanlineBreak;
@@ -1795,6 +1797,8 @@ void SoftRenderer::RenderPolygons(bool threaded, Polygon** polygons, int npolys)
                 timespent -= FreeTiming;
             }*/
             //if (!abort) 
+            //if (buffersize > 48) timespent -= PerScanlineRecup;
+            /*else*/
             timespent -= FreeTiming;
 
             if (timespent > 0)