From f097cc3b4e93bcd9b5db73d6be68d21dcfb85955 Mon Sep 17 00:00:00 2001
From: FireNX70 <firenx70@gmail.com>
Date: Thu, 13 Jun 2024 00:30:33 +0200
Subject: [PATCH] Tie BinCombined's local_size_x to the coarse tiles' area

---
 src/GPU3D_Compute.cpp       | 5 ++++-
 src/GPU3D_Compute.h         | 1 +
 src/GPU3D_Compute_shaders.h | 2 +-
 3 files changed, 6 insertions(+), 2 deletions(-)

diff --git a/src/GPU3D_Compute.cpp b/src/GPU3D_Compute.cpp
index c22a66b9..ecf48544 100644
--- a/src/GPU3D_Compute.cpp
+++ b/src/GPU3D_Compute.cpp
@@ -57,6 +57,8 @@ bool ComputeRenderer::CompileShader(GLuint& shader, const std::string& source, c
     shaderSource += std::to_string(TileSize);
     shaderSource += "\nconst int CoarseTileCountY = ";
     shaderSource += std::to_string(CoarseTileCountY) + ";";
+    shaderSource += "\n#define CoarseTileArea ";
+    shaderSource += std::to_string(CoarseTileArea);
 
     shaderSource += ComputeRendererShaders::Common;
     shaderSource += source;
@@ -330,6 +332,7 @@ void ComputeRenderer::SetRenderSettings(int scale, bool highResolutionCoordinate
     
     TileSize = std::min(8 * TileScale, 32);
     CoarseTileCountY = TileSize < 32 ? 4 : 6;
+    CoarseTileArea = CoarseTileCountX * CoarseTileCountY;
     CoarseTileW = CoarseTileCountX * TileSize;
     CoarseTileH = CoarseTileCountY * TileSize;
 
@@ -959,7 +962,7 @@ void ComputeRenderer::RenderFrame(GPU& gpu)
 
         // bin polygons
         glUseProgram(ShaderBinCombined);
-        glDispatchCompute(((gpu.GPU3D.RenderNumPolygons + 31) / 32), ScreenWidth/CoarseTileW, ScreenHeight/CoarseTileH);
+        glDispatchCompute(((gpu.GPU3D.RenderNumPolygons + CoarseTileArea - 1) / CoarseTileArea), ScreenWidth/CoarseTileW, ScreenHeight/CoarseTileH);
         glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT);
 
         // calculate list offsets
diff --git a/src/GPU3D_Compute.h b/src/GPU3D_Compute.h
index 4f944e42..6a5fd499 100644
--- a/src/GPU3D_Compute.h
+++ b/src/GPU3D_Compute.h
@@ -166,6 +166,7 @@ private:
     int TileSize;
     static constexpr int CoarseTileCountX = 8;
     int CoarseTileCountY;
+    int CoarseTileArea;
     int CoarseTileW;
     int CoarseTileH;
 
diff --git a/src/GPU3D_Compute_shaders.h b/src/GPU3D_Compute_shaders.h
index 9fb4aae7..9b319029 100644
--- a/src/GPU3D_Compute_shaders.h
+++ b/src/GPU3D_Compute_shaders.h
@@ -862,7 +862,7 @@ const std::string BinCombined =
     XSpanSetupBuffer +
     WorkDescBuffer + R"(
 
-layout (local_size_x = 32) in;
+layout (local_size_x = CoarseTileArea) in;
 
 bool BinPolygon(Polygon polygon, ivec2 topLeft, ivec2 botRight)
 {