Day 17-ish: Final culling tweaks

- I talked about this a little on stream in the lost 30 minutes, but due
  to top-left rasterization convention it is probably safe to slightly
  reduce the tested bbox. This reduces triangle count by an extra ~1.5%
  without visible holes

- Meshlet frustum culling was accidentally committed without the far
  plane check; it's conservative to do this but it feels reasonable to
  keep it.

src/niagara.cpp

@@ -1226,7 +1226,7 @@ int main(int argc, const char** argv)
 		double drawsPerSec = double(drawCount) / double(frameGpuAvg * 1e-3);
 
 		char title[256];
-		sprintf(title, "cpu: %.2f ms; gpu: %.2f ms (cull: %.2f ms, pyramid: %.2f ms, cull late: %.2f); triangles %.1fM; %.1fB tri/sec, %.1fM draws/sec; mesh shading %s, frustum culling %s, occlusion culling %s, level-of-detail %s",
+		sprintf(title, "cpu: %.2f ms; gpu: %.2f ms (cull: %.2f ms, pyramid: %.2f ms, cull late: %.2f); triangles %.2fM; %.1fB tri/sec, %.1fM draws/sec; mesh shading %s, frustum culling %s, occlusion culling %s, level-of-detail %s",
 			frameCpuAvg, frameGpuAvg, cullGpuTime, pyramidGpuTime, culllateGpuTime,
 			double(triangleCount) * 1e-6, trianglesPerSec * 1e-9, drawsPerSec * 1e-6,
 			meshShadingSupported && meshShadingEnabled ? "ON" : "OFF", cullingEnabled ? "ON" : "OFF", occlusionEnabled ? "ON" : "OFF", lodEnabled ? "ON" : "OFF");

src/shaders/meshlet.mesh.glsl

@@ -145,7 +145,7 @@ void main()
 		float sbprec = 1.0 / 256.0; // note: this can be set to 1/2^subpixelPrecisionBits
 
 		// note: this is slightly imprecise (doesn't fully match hw behavior and is both too loose and too strict)
-		culled = culled || (round(bmin.x - sbprec) == round(bmax.x + sbprec) || round(bmin.y - sbprec) == round(bmax.y + sbprec));
+		culled = culled || (round(bmin.x - sbprec) == round(bmax.x) || round(bmin.y) == round(bmax.y + sbprec));
 
 		// the computations above are only valid if all vertices are in front of perspective plane
 		culled = culled && (vertexClip[a].z > 0 && vertexClip[b].z > 0 && vertexClip[c].z > 0);

src/shaders/meshlet.task.glsl

@@ -73,7 +73,7 @@ void main()
 	visible = visible && center.z * globals.frustum[3] - abs(center.y) * globals.frustum[2] > -radius;
 	// the near/far plane culling uses camera space Z directly
 	// note: because we use an infinite projection matrix, this may cull meshlets that belong to a mesh that straddles the "far" plane; we could optionally remove the far check to be conservative
-	visible = visible && center.z + radius > globals.znear;// && center.z - radius < globals.zfar;
+	visible = visible && center.z + radius > globals.znear && center.z - radius < globals.zfar;
 
 	if (visible)
 	{