Occlusion culling algorithm based on recursive descend (#13104)

Co-authored-by: DS <vorunbekannt75@web.de>

builtin/settingtypes.txt

@@ -623,6 +623,15 @@ update_last_checked (Last update check) string
 #    Ex: 5.5.0 is 005005000
 update_last_known (Last known version update) int 0
 
+#    Type of occlusion_culler
+#    "loops" is the legacy algorithm with nested loops and O(N^3) complexity
+#    "bfs" is the new algorithm based on breadth-first-search and side culling
+occlusion_culler (Occlusion Culler) enum bfs bfs,loops
+
+#    Use raytraced occlusion culling in the new culler.
+#	 This flag enables use of raytraced occlusion culling test
+enable_raytraced_culling (Enable Raytraced Culling) bool true
+
 [*Server]
 
 #    Name of the player.

src/client/client.cpp

@@ -561,6 +561,7 @@ void Client::step(float dtime)
 				block->mesh = nullptr;
 
 				if (r.mesh) {
+					block->solid_sides = r.solid_sides;
 					minimap_mapblock = r.mesh->moveMinimapMapblock();
 					if (minimap_mapblock == NULL)
 						do_mapper_update = false;

src/client/clientmap.cpp

@@ -28,9 +28,10 @@ with this program; if not, write to the Free Software Foundation, Inc.,
 #include "settings.h"
 #include "camera.h"               // CameraModes
 #include "util/basic_macros.h"
-#include <algorithm>
 #include "client/renderingengine.h"
 
+#include <queue>
+
 // struct MeshBufListList
 void MeshBufListList::clear()
 {
@@ -60,6 +61,10 @@ void MeshBufListList::add(scene::IMeshBuffer *buf, v3s16 position, u8 layer)
 	list.emplace_back(l);
 }
 
+static void on_settings_changed(const std::string &name, void *data)
+{
+	static_cast<ClientMap*>(data)->onSettingChanged(name);
+}
 // ClientMap
 
 ClientMap::ClientMap(
@@ -98,7 +103,24 @@ ClientMap::ClientMap(
 	m_cache_bilinear_filter   = g_settings->getBool("bilinear_filter");
 	m_cache_anistropic_filter = g_settings->getBool("anisotropic_filter");
 	m_cache_transparency_sorting_distance = g_settings->getU16("transparency_sorting_distance");
+	m_new_occlusion_culler = g_settings->get("occlusion_culler") == "bfs";
+	g_settings->registerChangedCallback("occlusion_culler", on_settings_changed, this);
+	m_enable_raytraced_culling = g_settings->getBool("enable_raytraced_culling");
+	g_settings->registerChangedCallback("enable_raytraced_culling", on_settings_changed, this);
+}
 
+void ClientMap::onSettingChanged(const std::string &name)
+{
+	if (name == "occlusion_culler")
+		m_new_occlusion_culler = g_settings->get("occlusion_culler") == "bfs";
+	if (name == "enable_raytraced_culling")
+		m_enable_raytraced_culling = g_settings->getBool("enable_raytraced_culling");
+}
+
+ClientMap::~ClientMap()
+{
+	g_settings->deregisterChangedCallback("occlusion_culler", on_settings_changed, this);
+	g_settings->deregisterChangedCallback("enable_raytraced_culling", on_settings_changed, this);
 }
 
 void ClientMap::updateCamera(v3f pos, v3f dir, f32 fov, v3s16 offset)
@@ -180,8 +202,290 @@ void ClientMap::getBlocksInViewRange(v3s16 cam_pos_nodes,
 			p_nodes_max.Z / MAP_BLOCKSIZE + 1);
 }
 
+class MapBlockFlags
+{
+public:
+	static constexpr u16 CHUNK_EDGE = 8;
+	static constexpr u16 CHUNK_MASK = CHUNK_EDGE - 1;
+	static constexpr std::size_t CHUNK_VOLUME = CHUNK_EDGE * CHUNK_EDGE * CHUNK_EDGE; // volume of a chunk
+
+	MapBlockFlags(v3s16 min_pos, v3s16 max_pos)
+			: min_pos(min_pos), volume((max_pos - min_pos) / CHUNK_EDGE + 1)
+	{
+		chunks.resize(volume.X * volume.Y * volume.Z);
+	}
+
+	class Chunk
+	{
+	public:
+		inline u8 &getBits(v3s16 pos)
+		{
+			std::size_t address = getAddress(pos);
+			return bits[address];
+		}
+
+	private:
+		inline std::size_t getAddress(v3s16 pos) {
+			std::size_t address = (pos.X & CHUNK_MASK) + (pos.Y & CHUNK_MASK) * CHUNK_EDGE + (pos.Z & CHUNK_MASK) * (CHUNK_EDGE * CHUNK_EDGE);
+			return address;
+		}
+
+		std::array<u8, CHUNK_VOLUME> bits;
+	};
+
+	Chunk &getChunk(v3s16 pos)
+	{
+		v3s16 delta = (pos - min_pos) / CHUNK_EDGE;
+		std::size_t address = delta.X + delta.Y * volume.X + delta.Z * volume.X * volume.Y;
+		Chunk *chunk = chunks[address].get();
+		if (!chunk) {
+			chunk = new Chunk();
+			chunks[address].reset(chunk);
+		}
+		return *chunk;
+	}
+private:
+	std::vector<std::unique_ptr<Chunk>> chunks;
+	v3s16 min_pos;
+	v3s16 volume;
+};
+
 void ClientMap::updateDrawList()
 {
+	if (m_new_occlusion_culler) {
+		ScopeProfiler sp(g_profiler, "CM::updateDrawList()", SPT_AVG);
+
+		m_needs_update_drawlist = false;
+
+		for (auto &i : m_drawlist) {
+			MapBlock *block = i.second;
+			block->refDrop();
+		}
+		m_drawlist.clear();
+
+		v3s16 cam_pos_nodes = floatToInt(m_camera_position, BS);
+
+		v3s16 p_blocks_min;
+		v3s16 p_blocks_max;
+		getBlocksInViewRange(cam_pos_nodes, &p_blocks_min, &p_blocks_max);
+
+		// Number of blocks occlusion culled
+		u32 blocks_occlusion_culled = 0;
+		// Blocks visited by the algorithm
+		u32 blocks_visited = 0;
+		// Block sides that were not traversed
+		u32 sides_skipped = 0;
+
+		// No occlusion culling when free_move is on and camera is inside ground
+		bool occlusion_culling_enabled = true;
+		if (m_control.allow_noclip) {
+			MapNode n = getNode(cam_pos_nodes);
+			if (n.getContent() == CONTENT_IGNORE || m_nodedef->get(n).solidness == 2)
+				occlusion_culling_enabled = false;
+		}
+
+		v3s16 camera_block = getContainerPos(cam_pos_nodes, MAP_BLOCKSIZE);
+		m_drawlist = std::map<v3s16, MapBlock*, MapBlockComparer>(MapBlockComparer(camera_block));
+
+		auto is_frustum_culled = m_client->getCamera()->getFrustumCuller();
+
+		// Uncomment to debug occluded blocks in the wireframe mode
+		// TODO: Include this as a flag for an extended debugging setting
+		// if (occlusion_culling_enabled && m_control.show_wireframe)
+		// 	occlusion_culling_enabled = porting::getTimeS() & 1;
+
+		std::queue<v3s16> blocks_to_consider;
+
+		// Bits per block:
+		// [ visited | 0 | 0 | 0 | 0 | Z visible | Y visible | X visible ]
+		MapBlockFlags blocks_seen(p_blocks_min, p_blocks_max);
+
+		// Start breadth-first search with the block the camera is in
+		blocks_to_consider.push(camera_block);
+		blocks_seen.getChunk(camera_block).getBits(camera_block) = 0x07; // mark all sides as visible
+
+		// Recursively walk the space and pick mapblocks for drawing
+		while (blocks_to_consider.size() > 0) {
+
+			v3s16 block_coord = blocks_to_consider.front();
+			blocks_to_consider.pop();
+
+			auto &flags = blocks_seen.getChunk(block_coord).getBits(block_coord);
+
+			// Only visit each block once (it may have been queued up to three times)
+			if ((flags & 0x80) == 0x80)
+				continue;
+			flags |= 0x80;
+
+			blocks_visited++;
+
+			// Get the sector, block and mesh
+			MapSector *sector = this->getSectorNoGenerate(v2s16(block_coord.X, block_coord.Z));
+
+			if (!sector)
+				continue;
+
+			MapBlock *block = sector->getBlockNoCreateNoEx(block_coord.Y);
+
+			MapBlockMesh *mesh = block ? block->mesh : nullptr;
+
+
+			// Calculate the coordinates for range and frutum culling
+			v3f mesh_sphere_center;
+			f32 mesh_sphere_radius;
+
+			v3s16 block_pos_nodes = block_coord * MAP_BLOCKSIZE;
+
+			if (mesh) {
+				mesh_sphere_center = intToFloat(block_pos_nodes, BS)
+						+ mesh->getBoundingSphereCenter();
+				mesh_sphere_radius = mesh->getBoundingRadius();
+			}
+			else {
+				mesh_sphere_center = intToFloat(block_pos_nodes, BS) + v3f((MAP_BLOCKSIZE * 0.5f - 0.5f) * BS);
+				mesh_sphere_radius = 0.0f;
+			}
+
+			// First, perform a simple distance check.
+			if (!m_control.range_all &&
+				mesh_sphere_center.getDistanceFrom(intToFloat(cam_pos_nodes, BS)) >
+					m_control.wanted_range * BS + mesh_sphere_radius)
+				continue; // Out of range, skip.
+
+			// Frustum culling
+			// Only do coarse culling here, to account for fast camera movement.
+			// This is needed because this function is not called every frame.
+			float frustum_cull_extra_radius = 300.0f;
+			if (is_frustum_culled(mesh_sphere_center,
+					mesh_sphere_radius + frustum_cull_extra_radius))
+				continue;
+
+			// Calculate the vector from the camera block to the current block
+			// We use it to determine through which sides of the current block we can continue the search
+			v3s16 look = block_coord - camera_block;
+
+			// Occluded near sides will further occlude the far sides
+			u8 visible_outer_sides = flags & 0x07;
+
+			// Raytraced occlusion culling - send rays from the camera to the block's corners
+			if (occlusion_culling_enabled && m_enable_raytraced_culling &&
+					block && mesh &&
+					visible_outer_sides != 0x07 && isBlockOccluded(block, cam_pos_nodes)) {
+				blocks_occlusion_culled++;
+				continue;
+			}
+
+			// The block is visible, add to the draw list
+			if (mesh) {
+				// Add to set
+				block->refGrab();
+				m_drawlist[block_coord] = block;
+			}
+
+			// Decide which sides to traverse next or to block away
+
+			// First, find the near sides that would occlude the far sides
+			// * A near side can itself be occluded by a nearby block (the test above ^^)
+			// * A near side can be visible but fully opaque by itself (e.g. ground at the 0 level)
+
+			// mesh solid sides are +Z-Z+Y-Y+X-X
+			// if we are inside the block's coordinates on an axis, 
+			// treat these sides as opaque, as they should not allow to reach the far sides
+			u8 block_inner_sides = (look.X == 0 ? 3 : 0) |
+				(look.Y == 0 ? 12 : 0) |
+				(look.Z == 0 ? 48 : 0);
+
+			// get the mask for the sides that are relevant based on the direction
+			u8 near_inner_sides = (look.X > 0 ? 1 : 2) |
+					(look.Y > 0 ? 4 : 8) |
+					(look.Z > 0 ? 16 : 32);
+			
+			// This bitset is +Z-Z+Y-Y+X-X (See MapBlockMesh), and axis is XYZ.
+			// Get he block's transparent sides
+			u8 transparent_sides = (occlusion_culling_enabled && block) ? ~block->solid_sides : 0x3F;
+
+			// compress block transparent sides to ZYX mask of see-through axes
+			u8 near_transparency =  (block_inner_sides == 0x3F) ? near_inner_sides : (transparent_sides & near_inner_sides);
+
+			// when we are inside the camera block, do not block any sides
+			if (block_inner_sides == 0x3F)
+				block_inner_sides = 0;
+
+			near_transparency &= ~block_inner_sides & 0x3F;
+
+			near_transparency |= (near_transparency >> 1);
+			near_transparency = (near_transparency & 1) |
+					((near_transparency >> 1) & 2) |
+					((near_transparency >> 2) & 4);
+
+			// combine with known visible sides that matter
+			near_transparency &= visible_outer_sides;
+
+			// The rule for any far side to be visible:
+			// * Any of the adjacent near sides is transparent (different axes)
+			// * The opposite near side (same axis) is transparent, if it is the dominant axis of the look vector
+
+			// Calculate vector from camera to mapblock center. Because we only need relation between
+			// coordinates we scale by 2 to avoid precision loss.
+			v3s16 precise_look = 2 * (block_pos_nodes - cam_pos_nodes) + MAP_BLOCKSIZE - 1;
+
+			// dominant axis flag
+			u8 dominant_axis = (abs(precise_look.X) > abs(precise_look.Y) && abs(precise_look.X) > abs(precise_look.Z)) |
+						((abs(precise_look.Y) > abs(precise_look.Z) && abs(precise_look.Y) > abs(precise_look.X)) << 1) |
+						((abs(precise_look.Z) > abs(precise_look.X) && abs(precise_look.Z) > abs(precise_look.Y)) << 2);
+
+			// Queue next blocks for processing:
+			// - Examine "far" sides of the current blocks, i.e. never move towards the camera
+			// - Only traverse the sides that are not occluded
+			// - Only traverse the sides that are not opaque
+			// When queueing, mark the relevant side on the next block as 'visible'
+			for (s16 axis = 0; axis < 3; axis++) {
+
+				// Select a bit from transparent_sides for the side
+				u8 far_side_mask = 1 << (2 * axis);
+
+				// axis flag
+				u8 my_side = 1 << axis;
+				u8 adjacent_sides = my_side ^ 0x07;
+
+				auto traverse_far_side = [&](s8 next_pos_offset) {
+					// far side is visible if adjacent near sides are transparent, or if opposite side on dominant axis is transparent
+					bool side_visible = ((near_transparency & adjacent_sides) | (near_transparency & my_side & dominant_axis)) != 0;
+					side_visible = side_visible && ((far_side_mask & transparent_sides) != 0);
+
+					v3s16 next_pos = block_coord;
+					next_pos[axis] += next_pos_offset;
+
+					// If a side is a see-through, mark the next block's side as visible, and queue
+					if (side_visible) {
+						auto &next_flags = blocks_seen.getChunk(next_pos).getBits(next_pos);
+						next_flags |= my_side;
+						blocks_to_consider.push(next_pos);
+					}
+					else {
+						sides_skipped++;
+					}
+				};
+
+
+				// Test the '-' direction of the axis
+				if (look[axis] <= 0 && block_coord[axis] > p_blocks_min[axis])
+					traverse_far_side(-1);
+
+				// Test the '+' direction of the axis
+				far_side_mask <<= 1;
+
+				if (look[axis] >= 0 && block_coord[axis] < p_blocks_max[axis])
+					traverse_far_side(+1);
+			}
+		}
+
+		g_profiler->avg("MapBlocks occlusion culled [#]", blocks_occlusion_culled);
+		g_profiler->avg("MapBlocks sides skipped [#]", sides_skipped);
+		g_profiler->avg("MapBlocks examined [#]", blocks_visited);
+		g_profiler->avg("MapBlocks drawn [#]", m_drawlist.size());
+	}
+	else {
 		ScopeProfiler sp(g_profiler, "CM::updateDrawList()", SPT_AVG);
 
 		m_needs_update_drawlist = false;
@@ -297,6 +601,70 @@ void ClientMap::updateDrawList()
 		g_profiler->avg("MapBlocks occlusion culled [#]", blocks_occlusion_culled);
 		g_profiler->avg("MapBlocks drawn [#]", m_drawlist.size());
 		g_profiler->avg("MapBlocks loaded [#]", blocks_loaded);
+	}
+}
+
+void ClientMap::touchMapBlocks()
+{
+	if (!m_new_occlusion_culler)
+		return;
+
+	v3s16 cam_pos_nodes = floatToInt(m_camera_position, BS);
+
+	v3s16 p_blocks_min;
+	v3s16 p_blocks_max;
+	getBlocksInViewRange(cam_pos_nodes, &p_blocks_min, &p_blocks_max);
+
+	// Number of blocks currently loaded by the client
+	u32 blocks_loaded = 0;
+	// Number of blocks with mesh in rendering range
+	u32 blocks_in_range_with_mesh = 0;
+
+	for (const auto &sector_it : m_sectors) {
+		MapSector *sector = sector_it.second;
+		v2s16 sp = sector->getPos();
+
+		blocks_loaded += sector->size();
+		if (!m_control.range_all) {
+			if (sp.X < p_blocks_min.X || sp.X > p_blocks_max.X ||
+					sp.Y < p_blocks_min.Z || sp.Y > p_blocks_max.Z)
+				continue;
+		}
+
+		MapBlockVect sectorblocks;
+		sector->getBlocks(sectorblocks);
+
+		/*
+			Loop through blocks in sector
+		*/
+
+		for (MapBlock *block : sectorblocks) {
+			/*
+				Compare block position to camera position, skip
+				if not seen on display
+			*/
+
+			if (!block->mesh) {
+				// Ignore if mesh doesn't exist
+				continue;
+			}
+
+			v3f mesh_sphere_center = intToFloat(block->getPosRelative(), BS)
+					+ block->mesh->getBoundingSphereCenter();
+			f32 mesh_sphere_radius = block->mesh->getBoundingRadius();
+			// First, perform a simple distance check.
+			if (!m_control.range_all &&
+				mesh_sphere_center.getDistanceFrom(intToFloat(cam_pos_nodes, BS)) >
+					m_control.wanted_range * BS + mesh_sphere_radius)
+				continue; // Out of range, skip.
+
+			// Keep the block alive as long as it is in range.
+			block->resetUsageTimer();
+			blocks_in_range_with_mesh++;
+		}
+	}
+	g_profiler->avg("MapBlock meshes in range [#]", blocks_in_range_with_mesh);
+	g_profiler->avg("MapBlocks loaded [#]", blocks_loaded);
 }
 
 void ClientMap::renderMap(video::IVideoDriver* driver, s32 pass)

src/client/clientmap.h

@@ -76,7 +76,7 @@ public:
 			s32 id
 	);
 
-	virtual ~ClientMap() = default;
+	virtual ~ClientMap();
 
 	bool maySaveBlocks() override
 	{
@@ -116,6 +116,8 @@ public:
 	void getBlocksInViewRange(v3s16 cam_pos_nodes,
 		v3s16 *p_blocks_min, v3s16 *p_blocks_max, float range=-1.0f);
 	void updateDrawList();
+	// @brief Calculate statistics about the map and keep the blocks alive
+	void touchMapBlocks();
 	void updateDrawListShadow(v3f shadow_light_pos, v3f shadow_light_dir, float radius, float length);
 	// Returns true if draw list needs updating before drawing the next frame.
 	bool needsUpdateDrawList() { return m_needs_update_drawlist; }
@@ -136,11 +138,14 @@ public:
 	f32 getWantedRange() const { return m_control.wanted_range; }
 	f32 getCameraFov() const { return m_camera_fov; }
 
+	void onSettingChanged(const std::string &name);
+
 private:
 
 	// update the vertex order in transparent mesh buffers
 	void updateTransparentMeshBuffers();
 
+
 	// Orders blocks by distance to the camera
 	class MapBlockComparer
 	{
@@ -205,4 +210,7 @@ private:
 	bool m_cache_bilinear_filter;
 	bool m_cache_anistropic_filter;
 	u16 m_cache_transparency_sorting_distance;
+
+	bool m_new_occlusion_culler;
+	bool m_enable_raytraced_culling;
 };

src/client/game.cpp

@@ -696,6 +696,7 @@ struct GameRunData {
 
 	float damage_flash;
 	float update_draw_list_timer;
+	float touch_blocks_timer;
 
 	f32 fog_range;
 
@@ -4030,6 +4031,9 @@ void Game::updateFrame(ProfilerGraph *graph, RunStats *stats, f32 dtime,
 		changed much
 	*/
 	runData.update_draw_list_timer += dtime;
+	runData.touch_blocks_timer += dtime;
+
+	bool draw_list_updated = false;
 
 	float update_draw_list_delta = 0.2f;
 
@@ -4041,6 +4045,12 @@ void Game::updateFrame(ProfilerGraph *graph, RunStats *stats, f32 dtime,
 		runData.update_draw_list_timer = 0;
 		client->getEnv().getClientMap().updateDrawList();
 		runData.update_draw_list_last_cam_dir = camera_direction;
+		draw_list_updated = true;
+	}
+
+	if (runData.touch_blocks_timer > update_draw_list_delta && !draw_list_updated) {
+		client->getEnv().getClientMap().touchMapBlocks();
+		runData.touch_blocks_timer = 0;
 	}
 
 	if (RenderingEngine::get_shadow_renderer()) {

src/client/mapblock_mesh.cpp

@@ -1582,3 +1582,31 @@ video::SColor encode_light(u16 light, u8 emissive_light)
 	float b = (day + night) / 2;
 	return video::SColor(r, b, b, b);
 }
+
+u8 get_solid_sides(MeshMakeData *data)
+{
+	v3s16 blockpos_nodes = data->m_blockpos * MAP_BLOCKSIZE;
+	const NodeDefManager *ndef = data->m_client->ndef();
+
+	u8 result = 0x3F; // all sides solid;
+
+	for (s16 i = 0; i < MAP_BLOCKSIZE && result != 0; i++)
+	for (s16 j = 0; j < MAP_BLOCKSIZE && result != 0; j++) {
+		v3s16 positions[6] = {
+			v3s16(0, i, j),
+			v3s16(MAP_BLOCKSIZE - 1, i, j),
+			v3s16(i, 0, j),
+			v3s16(i, MAP_BLOCKSIZE - 1, j),
+			v3s16(i, j, 0),
+			v3s16(i, j, MAP_BLOCKSIZE - 1)
+		};
+
+		for (u8 k = 0; k < 6; k++) {
+			const MapNode &top = data->m_vmanip.getNodeRefUnsafe(blockpos_nodes + positions[k]);
+			if (ndef->get(top).solidness != 2)
+				result &= ~(1 << k);
+		}
+	}
+
+	return result;
+}

src/client/mapblock_mesh.h

@@ -340,3 +340,8 @@ void final_color_blend(video::SColor *result,
 // TileFrame vector copy cost very much to client
 void getNodeTileN(MapNode mn, const v3s16 &p, u8 tileindex, MeshMakeData *data, TileSpec &tile);
 void getNodeTile(MapNode mn, const v3s16 &p, const v3s16 &dir, MeshMakeData *data, TileSpec &tile);
+
+/// Return bitset of the sides of the mapblock that consist of solid nodes only
+/// Bits:
+/// 0 0 -Z +Z -X +X -Y +Y
+u8 get_solid_sides(MeshMakeData *data);

src/client/mesh_generator_thread.cpp

@@ -299,6 +299,7 @@ void MeshUpdateWorkerThread::doUpdate()
 		MeshUpdateResult r;
 		r.p = q->p;
 		r.mesh = mesh_new;
+		r.solid_sides = get_solid_sides(q->data);
 		r.ack_block_to_server = q->ack_block_to_server;
 		r.urgent = q->urgent;
 

src/client/mesh_generator_thread.h

@@ -111,6 +111,7 @@ struct MeshUpdateResult
 {
 	v3s16 p = v3s16(-1338, -1338, -1338);
 	MapBlockMesh *mesh = nullptr;
+	u8 solid_sides = 0;
 	bool ack_block_to_server = false;
 	bool urgent = false;
 

src/defaultsettings.cpp

@@ -66,6 +66,8 @@ void set_default_settings()
 	settings->setDefault("max_out_chat_queue_size", "20");
 	settings->setDefault("pause_on_lost_focus", "false");
 	settings->setDefault("enable_split_login_register", "true");
+	settings->setDefault("occlusion_culler", "bfs");
+	settings->setDefault("enable_raytraced_culling", "true");
 	settings->setDefault("chat_weblink_color", "#8888FF");
 
 	// Keymap

src/mapblock.h

@@ -469,6 +469,8 @@ public:
 	bool contents_cached = false;
 	// True if we never want to cache content types for this block
 	bool do_not_cache_contents = false;
+	// marks the sides which are opaque: 00+Z-Z+Y-Y+X-X
+	u8 solid_sides {0};
 
 private:
 	/*