From 0970785d4a8c20cbd577a3cd0d99489d95cb269d Mon Sep 17 00:00:00 2001
From: Arseny Kapoulkine <arseny.kapoulkine@gmail.com>
Date: Fri, 29 Mar 2024 20:46:23 -0700
Subject: [PATCH] clusterizer: Implement experimental meshlet optimizer

So far we were mostly concerned with meshlet clustering from the
perspective of treating meshlets as an unordered set of triangles; while
this matches the computational and documented model, this may not be
optimal for a given GPU.

Notably, NVidia GPUs are much more sensitive to the order of triangles
in the meshlet than to the number and fill percentage; so much so that
from pure rasterization performance, scan may win over proper clustering
because it implicitly generates a better order.

We do not know the precise criteria / mechanism that NV GPUs use here
but it helps to do locality optimization; most importantly, triangle
order, but also reordering meshlet-local vertices helps a little bit.

This change implements a simple meshlet optimizer; while this can also
be achieved by running existing optimization algorithms (vcache /
vfetch) on meshlet data, a custom optimizer is faster even when using
unoptimized quadratic implementation, and may allow us to implement
better locality reodering algorithms in the future assuming a small
input patch.

For now we just select the next triangle to maximize the number of
shared vertices with the previous triangle; this results in a
pseudo-strip order which seems reasonably optimal for NV; note that
neither the hardware nor this algorithm is concerned with the specifics
of edge matching as it doesn't seem to matter for performance.
---
 src/clusterizer.cpp | 78 +++++++++++++++++++++++++++++++++++++++++++++
 src/meshoptimizer.h | 10 ++++++
 2 files changed, 88 insertions(+)

diff --git a/src/clusterizer.cpp b/src/clusterizer.cpp
index c4672ad6..555cd108 100644
--- a/src/clusterizer.cpp
+++ b/src/clusterizer.cpp
@@ -882,3 +882,81 @@ meshopt_Bounds meshopt_computeMeshletBounds(const unsigned int* meshlet_vertices
 
 	return meshopt_computeClusterBounds(indices, triangle_count * 3, vertex_positions, vertex_count, vertex_positions_stride);
 }
+
+void meshopt_optimizeMeshlet(unsigned int* meshlet_vertices, unsigned char* meshlet_triangles, size_t triangle_count, size_t vertex_count)
+{
+	using namespace meshopt;
+
+	assert(triangle_count <= kMeshletMaxTriangles);
+	assert(vertex_count <= kMeshletMaxVertices);
+
+	unsigned char* indices = meshlet_triangles;
+	unsigned int* vertices = meshlet_vertices;
+
+	unsigned char result[kMeshletMaxTriangles * 3];
+
+	unsigned char visited[kMeshletMaxTriangles];
+	memset(visited, 0, triangle_count);
+
+	unsigned char lasta = 0, lastb = 0, lastc = 0;
+
+	for (size_t i = 0; i < triangle_count; ++i)
+	{
+		int next = -1;
+
+		for (int pass = 2; pass >= 0; pass--)
+		{
+			for (size_t j = 0; j < triangle_count; ++j)
+			{
+				if (visited[j])
+					continue;
+
+				unsigned char a = indices[j * 3 + 0], b = indices[j * 3 + 1], c = indices[j * 3 + 2];
+
+				int aok = (a == lasta || a == lastb || a == lastc);
+				int bok = (b == lasta || b == lastb || b == lastc);
+				int cok = (c == lasta || c == lastb || c == lastc);
+
+				if (aok + bok + cok >= pass)
+				{
+					next = (int)j;
+					break;
+				}
+			}
+
+			if (next >= 0)
+				break;
+		}
+
+		assert(next >= 0);
+		visited[next] = 1;
+		result[i * 3 + 0] = lasta = indices[next * 3 + 0];
+		result[i * 3 + 1] = lastb = indices[next * 3 + 1];
+		result[i * 3 + 2] = lastc = indices[next * 3 + 2];
+	}
+
+	memcpy(indices, result, triangle_count * 3);
+
+	unsigned int newv[kMeshletMaxVertices];
+
+	unsigned char remap[kMeshletMaxVertices];
+	memset(remap, -1, vertex_count);
+
+	size_t vertex_offset = 0;
+
+	for (size_t i = 0; i < triangle_count * 3; ++i)
+	{
+		unsigned char& r = remap[indices[i]];
+
+		if (r == 0xff)
+		{
+			r = unsigned(vertex_offset);
+			newv[vertex_offset] = vertices[indices[i]];
+			vertex_offset++;
+		}
+
+		indices[i] = r;
+	}
+
+	memcpy(vertices, newv, vertex_offset * sizeof(unsigned int));
+}
diff --git a/src/meshoptimizer.h b/src/meshoptimizer.h
index 0619cec8..6ab48c19 100644
--- a/src/meshoptimizer.h
+++ b/src/meshoptimizer.h
@@ -493,6 +493,16 @@ MESHOPTIMIZER_API size_t meshopt_buildMeshlets(struct meshopt_Meshlet* meshlets,
 MESHOPTIMIZER_API size_t meshopt_buildMeshletsScan(struct meshopt_Meshlet* meshlets, unsigned int* meshlet_vertices, unsigned char* meshlet_triangles, const unsigned int* indices, size_t index_count, size_t vertex_count, size_t max_vertices, size_t max_triangles);
 MESHOPTIMIZER_API size_t meshopt_buildMeshletsBound(size_t index_count, size_t max_vertices, size_t max_triangles);
 
+/**
+ * Experimental: Meshlet optimizer
+ * Reorders meshlet vertices and triangles to maximize locality to improve rasterizer throughput
+ *
+ * meshlet_triangles and meshlet_vertices must refer to meshlet triangle and vertex index data; when buildMeshlets* is used, these
+ * need to be computed from meshlet's vertex_offset and triangle_offset
+ * triangle_count and vertex_count must not exceed implementation limits (vertex_count <= 255 - not 256!, triangle_count <= 512)
+ */
+MESHOPTIMIZER_EXPERIMENTAL void meshopt_optimizeMeshlet(unsigned int* meshlet_vertices, unsigned char* meshlet_triangles, size_t triangle_count, size_t vertex_count);
+
 struct meshopt_Bounds
 {
 	/* bounding sphere, useful for frustum and occlusion culling */