Merge pull request #48629 from nekomatata/dynamic-bvh-broadphase-4.0

Dynamic BVH broadphase in 2D & 3D Godot Physics

33 files changed, 3735 insertions, 1538 deletions

diff --git a/core/math/bvh.h b/core/math/bvh.h
new file mode 100644
index 0000000000..cefbc9b0db
--- /dev/null
+++ b/core/math/bvh.h
@@ -0,0 +1,695 @@
+/*************************************************************************/
+/*  bvh.h                                                                */
+/*************************************************************************/
+/*                       This file is part of:                           */
+/*                           GODOT ENGINE                                */
+/*                      https://godotengine.org                          */
+/*************************************************************************/
+/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
+/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
+/*                                                                       */
+/* Permission is hereby granted, free of charge, to any person obtaining */
+/* a copy of this software and associated documentation files (the       */
+/* "Software"), to deal in the Software without restriction, including   */
+/* without limitation the rights to use, copy, modify, merge, publish,   */
+/* distribute, sublicense, and/or sell copies of the Software, and to    */
+/* permit persons to whom the Software is furnished to do so, subject to */
+/* the following conditions:                                             */
+/*                                                                       */
+/* The above copyright notice and this permission notice shall be        */
+/* included in all copies or substantial portions of the Software.       */
+/*                                                                       */
+/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
+/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
+/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
+/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
+/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
+/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
+/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
+/*************************************************************************/
+
+#ifndef BVH_H
+#define BVH_H
+
+// BVH
+// This class provides a wrapper around BVH tree, which contains most of the functionality
+// for a dynamic BVH with templated leaf size.
+// However BVH also adds facilities for pairing, to maintain compatibility with Godot 3.2.
+// Pairing is a collision pairing system, on top of the basic BVH.
+
+// Some notes on the use of BVH / Octree from Godot 3.2.
+// This is not well explained elsewhere.
+// The rendering tree mask and types that are sent to the BVH are NOT layer masks.
+// They are INSTANCE_TYPES (defined in visual_server.h), e.g. MESH, MULTIMESH, PARTICLES etc.
+// Thus the lights do no cull by layer mask in the BVH.
+
+// Layer masks are implemented in the renderers as a later step, and light_cull_mask appears to be
+// implemented in GLES3 but not GLES2. Layer masks are not yet implemented for directional lights.
+
+#include "bvh_tree.h"
+
+#define BVHTREE_CLASS BVH_Tree<T, 2, MAX_ITEMS, USE_PAIRS, Bounds, Point>
+
+template <class T, bool USE_PAIRS = false, int MAX_ITEMS = 32, class Bounds = AABB, class Point = Vector3>
+class BVH_Manager {
+public:
+	// note we are using uint32_t instead of BVHHandle, losing type safety, but this
+	// is for compatibility with octree
+	typedef void *(*PairCallback)(void *, uint32_t, T *, int, uint32_t, T *, int);
+	typedef void (*UnpairCallback)(void *, uint32_t, T *, int, uint32_t, T *, int, void *);
+
+	// these 2 are crucial for fine tuning, and can be applied manually
+	// see the variable declarations for more info.
+	void params_set_node_expansion(real_t p_value) {
+		if (p_value >= 0.0) {
+			tree._node_expansion = p_value;
+			tree._auto_node_expansion = false;
+		} else {
+			tree._auto_node_expansion = true;
+		}
+	}
+
+	void params_set_pairing_expansion(real_t p_value) {
+		if (p_value >= 0.0) {
+			tree._pairing_expansion = p_value;
+			tree._auto_pairing_expansion = false;
+		} else {
+			tree._auto_pairing_expansion = true;
+		}
+	}
+
+	void set_pair_callback(PairCallback p_callback, void *p_userdata) {
+		pair_callback = p_callback;
+		pair_callback_userdata = p_userdata;
+	}
+	void set_unpair_callback(UnpairCallback p_callback, void *p_userdata) {
+		unpair_callback = p_callback;
+		unpair_callback_userdata = p_userdata;
+	}
+
+	BVHHandle create(T *p_userdata, bool p_active, const Bounds &p_aabb = Bounds(), int p_subindex = 0, bool p_pairable = false, uint32_t p_pairable_type = 0, uint32_t p_pairable_mask = 1) {
+		// not sure if absolutely necessary to flush collisions here. It will cost performance to, instead
+		// of waiting for update, so only uncomment this if there are bugs.
+		if (USE_PAIRS) {
+			//_check_for_collisions();
+		}
+
+#ifdef TOOLS_ENABLED
+		if (!USE_PAIRS) {
+			if (p_pairable) {
+				WARN_PRINT_ONCE("creating pairable item in BVH with USE_PAIRS set to false");
+			}
+		}
+#endif
+
+		BVHHandle h = tree.item_add(p_userdata, p_active, p_aabb, p_subindex, p_pairable, p_pairable_type, p_pairable_mask);
+
+		if (USE_PAIRS) {
+			// for safety initialize the expanded AABB
+			Bounds &expanded_aabb = tree._pairs[h.id()].expanded_aabb;
+			expanded_aabb = p_aabb;
+			expanded_aabb.grow_by(tree._pairing_expansion);
+
+			// force a collision check no matter the AABB
+			if (p_active) {
+				_add_changed_item(h, p_aabb, false);
+				_check_for_collisions(true);
+			}
+		}
+
+		return h;
+	}
+
+	////////////////////////////////////////////////////
+	// wrapper versions that use uint32_t instead of handle
+	// for backward compatibility. Less type safe
+	void move(uint32_t p_handle, const Bounds &p_aabb) {
+		BVHHandle h;
+		h.set(p_handle);
+		move(h, p_aabb);
+	}
+
+	void erase(uint32_t p_handle) {
+		BVHHandle h;
+		h.set(p_handle);
+		erase(h);
+	}
+
+	void force_collision_check(uint32_t p_handle) {
+		BVHHandle h;
+		h.set(p_handle);
+		force_collision_check(h);
+	}
+
+	bool activate(uint32_t p_handle, const Bounds &p_aabb, bool p_delay_collision_check = false) {
+		BVHHandle h;
+		h.set(p_handle);
+		return activate(h, p_aabb, p_delay_collision_check);
+	}
+
+	bool deactivate(uint32_t p_handle) {
+		BVHHandle h;
+		h.set(p_handle);
+		return deactivate(h);
+	}
+
+	void set_pairable(uint32_t p_handle, bool p_pairable, uint32_t p_pairable_type, uint32_t p_pairable_mask, bool p_force_collision_check = true) {
+		BVHHandle h;
+		h.set(p_handle);
+		set_pairable(h, p_pairable, p_pairable_type, p_pairable_mask, p_force_collision_check);
+	}
+
+	bool is_pairable(uint32_t p_handle) const {
+		BVHHandle h;
+		h.set(p_handle);
+		return item_is_pairable(h);
+	}
+	int get_subindex(uint32_t p_handle) const {
+		BVHHandle h;
+		h.set(p_handle);
+		return item_get_subindex(h);
+	}
+
+	T *get(uint32_t p_handle) const {
+		BVHHandle h;
+		h.set(p_handle);
+		return item_get_userdata(h);
+	}
+
+	////////////////////////////////////////////////////
+
+	void move(BVHHandle p_handle, const Bounds &p_aabb) {
+		if (tree.item_move(p_handle, p_aabb)) {
+			if (USE_PAIRS) {
+				_add_changed_item(p_handle, p_aabb);
+			}
+		}
+	}
+
+	void erase(BVHHandle p_handle) {
+		// call unpair and remove all references to the item
+		// before deleting from the tree
+		if (USE_PAIRS) {
+			_remove_changed_item(p_handle);
+		}
+
+		tree.item_remove(p_handle);
+
+		_check_for_collisions(true);
+	}
+
+	// use in conjunction with activate if you have deferred the collision check, and
+	// set pairable has never been called.
+	// (deferred collision checks are a workaround for visual server for historical reasons)
+	void force_collision_check(BVHHandle p_handle) {
+		if (USE_PAIRS) {
+			// the aabb should already be up to date in the BVH
+			Bounds aabb;
+			item_get_AABB(p_handle, aabb);
+
+			// add it as changed even if aabb not different
+			_add_changed_item(p_handle, aabb, false);
+
+			// force an immediate full collision check, much like calls to set_pairable
+			_check_for_collisions(true);
+		}
+	}
+
+	// these should be read as set_visible for render trees,
+	// but generically this makes items add or remove from the
+	// tree internally, to speed things up by ignoring inactive items
+	bool activate(BVHHandle p_handle, const Bounds &p_aabb, bool p_delay_collision_check = false) {
+		// sending the aabb here prevents the need for the BVH to maintain
+		// a redundant copy of the aabb.
+		// returns success
+		if (tree.item_activate(p_handle, p_aabb)) {
+			if (USE_PAIRS) {
+				// in the special case of the render tree, when setting visibility we are using the combination of
+				// activate then set_pairable. This would case 2 sets of collision checks. For efficiency here we allow
+				// deferring to have a single collision check at the set_pairable call.
+				// Watch for bugs! This may cause bugs if set_pairable is not called.
+				if (!p_delay_collision_check) {
+					_add_changed_item(p_handle, p_aabb, false);
+
+					// force an immediate collision check, much like calls to set_pairable
+					_check_for_collisions(true);
+				}
+			}
+			return true;
+		}
+
+		return false;
+	}
+
+	bool deactivate(BVHHandle p_handle) {
+		// returns success
+		if (tree.item_deactivate(p_handle)) {
+			// call unpair and remove all references to the item
+			// before deleting from the tree
+			if (USE_PAIRS) {
+				_remove_changed_item(p_handle);
+
+				// force check for collisions, much like an erase was called
+				_check_for_collisions(true);
+			}
+			return true;
+		}
+
+		return false;
+	}
+
+	bool get_active(BVHHandle p_handle) const {
+		return tree.item_get_active(p_handle);
+	}
+
+	// call e.g. once per frame (this does a trickle optimize)
+	void update() {
+		tree.update();
+		_check_for_collisions();
+#ifdef BVH_INTEGRITY_CHECKS
+		tree.integrity_check_all();
+#endif
+	}
+
+	// this can be called more frequently than per frame if necessary
+	void update_collisions() {
+		_check_for_collisions();
+	}
+
+	// prefer calling this directly as type safe
+	void set_pairable(const BVHHandle &p_handle, bool p_pairable, uint32_t p_pairable_type, uint32_t p_pairable_mask, bool p_force_collision_check = true) {
+		// Returns true if the pairing state has changed.
+		bool state_changed = tree.item_set_pairable(p_handle, p_pairable, p_pairable_type, p_pairable_mask);
+
+		if (USE_PAIRS) {
+			// not sure if absolutely necessary to flush collisions here. It will cost performance to, instead
+			// of waiting for update, so only uncomment this if there are bugs.
+			//_check_for_collisions();
+
+			if ((p_force_collision_check || state_changed) && get_active(p_handle)) {
+				// when the pairable state changes, we need to force a collision check because newly pairable
+				// items may be in collision, and unpairable items might move out of collision.
+				// We cannot depend on waiting for the next update, because that may come much later.
+				Bounds aabb;
+				item_get_AABB(p_handle, aabb);
+
+				// passing false disables the optimization which prevents collision checks if
+				// the aabb hasn't changed
+				_add_changed_item(p_handle, aabb, false);
+
+				// force an immediate collision check (probably just for this one item)
+				// but it must be a FULL collision check, also checking pairable state and masks.
+				// This is because AABB intersecting objects may have changed pairable state / mask
+				// such that they should no longer be paired. E.g. lights.
+				_check_for_collisions(true);
+			} // only if active
+		}
+	}
+
+	// cull tests
+	int cull_aabb(const Bounds &p_aabb, T **p_result_array, int p_result_max, int *p_subindex_array = nullptr, uint32_t p_mask = 0xFFFFFFFF) {
+		typename BVHTREE_CLASS::CullParams params;
+
+		params.result_count_overall = 0;
+		params.result_max = p_result_max;
+		params.result_array = p_result_array;
+		params.subindex_array = p_subindex_array;
+		params.mask = p_mask;
+		params.pairable_type = 0;
+		params.test_pairable_only = false;
+		params.abb.from(p_aabb);
+
+		tree.cull_aabb(params);
+
+		return params.result_count_overall;
+	}
+
+	int cull_segment(const Point &p_from, const Point &p_to, T **p_result_array, int p_result_max, int *p_subindex_array = nullptr, uint32_t p_mask = 0xFFFFFFFF) {
+		typename BVHTREE_CLASS::CullParams params;
+
+		params.result_count_overall = 0;
+		params.result_max = p_result_max;
+		params.result_array = p_result_array;
+		params.subindex_array = p_subindex_array;
+		params.mask = p_mask;
+		params.pairable_type = 0;
+
+		params.segment.from = p_from;
+		params.segment.to = p_to;
+
+		tree.cull_segment(params);
+
+		return params.result_count_overall;
+	}
+
+	int cull_point(const Point &p_point, T **p_result_array, int p_result_max, int *p_subindex_array = nullptr, uint32_t p_mask = 0xFFFFFFFF) {
+		typename BVHTREE_CLASS::CullParams params;
+
+		params.result_count_overall = 0;
+		params.result_max = p_result_max;
+		params.result_array = p_result_array;
+		params.subindex_array = p_subindex_array;
+		params.mask = p_mask;
+		params.pairable_type = 0;
+
+		params.point = p_point;
+
+		tree.cull_point(params);
+		return params.result_count_overall;
+	}
+
+	int cull_convex(const Vector<Plane> &p_convex, T **p_result_array, int p_result_max, uint32_t p_mask = 0xFFFFFFFF) {
+		if (!p_convex.size()) {
+			return 0;
+		}
+
+		Vector<Vector3> convex_points = Geometry3D::compute_convex_mesh_points(&p_convex[0], p_convex.size());
+		if (convex_points.size() == 0) {
+			return 0;
+		}
+
+		typename BVHTREE_CLASS::CullParams params;
+		params.result_count_overall = 0;
+		params.result_max = p_result_max;
+		params.result_array = p_result_array;
+		params.subindex_array = nullptr;
+		params.mask = p_mask;
+		params.pairable_type = 0;
+
+		params.hull.planes = &p_convex[0];
+		params.hull.num_planes = p_convex.size();
+		params.hull.points = &convex_points[0];
+		params.hull.num_points = convex_points.size();
+
+		tree.cull_convex(params);
+
+		return params.result_count_overall;
+	}
+
+private:
+	// do this after moving etc.
+	void _check_for_collisions(bool p_full_check = false) {
+		if (!changed_items.size()) {
+			// noop
+			return;
+		}
+
+		Bounds bb;
+
+		typename BVHTREE_CLASS::CullParams params;
+
+		params.result_count_overall = 0;
+		params.result_max = INT_MAX;
+		params.result_array = nullptr;
+		params.subindex_array = nullptr;
+		params.mask = 0xFFFFFFFF;
+		params.pairable_type = 0;
+
+		for (unsigned int n = 0; n < changed_items.size(); n++) {
+			const BVHHandle &h = changed_items[n];
+
+			// use the expanded aabb for pairing
+			const Bounds &expanded_aabb = tree._pairs[h.id()].expanded_aabb;
+			BVHABB_CLASS abb;
+			abb.from(expanded_aabb);
+
+			// find all the existing paired aabbs that are no longer
+			// paired, and send callbacks
+			_find_leavers(h, abb, p_full_check);
+
+			uint32_t changed_item_ref_id = h.id();
+
+			// set up the test from this item.
+			// this includes whether to test the non pairable tree,
+			// and the item mask.
+			tree.item_fill_cullparams(h, params);
+
+			params.abb = abb;
+
+			params.result_count_overall = 0; // might not be needed
+			tree.cull_aabb(params, false);
+
+			for (unsigned int i = 0; i < tree._cull_hits.size(); i++) {
+				uint32_t ref_id = tree._cull_hits[i];
+
+				// don't collide against ourself
+				if (ref_id == changed_item_ref_id) {
+					continue;
+				}
+
+#ifdef BVH_CHECKS
+				// if neither are pairable, they should ignore each other
+				// THIS SHOULD NEVER HAPPEN .. now we only test the pairable tree
+				// if the changed item is not pairable
+				CRASH_COND(params.test_pairable_only && !tree._extra[ref_id].pairable);
+#endif
+
+				// checkmasks is already done in the cull routine.
+				BVHHandle h_collidee;
+				h_collidee.set_id(ref_id);
+
+				// find NEW enterers, and send callbacks for them only
+				_collide(h, h_collidee);
+			}
+		}
+		_reset();
+	}
+
+public:
+	void item_get_AABB(BVHHandle p_handle, Bounds &r_aabb) {
+		BVHABB_CLASS abb;
+		tree.item_get_ABB(p_handle, abb);
+		abb.to(r_aabb);
+	}
+
+private:
+	// supplemental funcs
+	bool item_is_pairable(BVHHandle p_handle) const { return _get_extra(p_handle).pairable; }
+	T *item_get_userdata(BVHHandle p_handle) const { return _get_extra(p_handle).userdata; }
+	int item_get_subindex(BVHHandle p_handle) const { return _get_extra(p_handle).subindex; }
+
+	void _unpair(BVHHandle p_from, BVHHandle p_to) {
+		tree._handle_sort(p_from, p_to);
+
+		typename BVHTREE_CLASS::ItemExtra &exa = tree._extra[p_from.id()];
+		typename BVHTREE_CLASS::ItemExtra &exb = tree._extra[p_to.id()];
+
+		// if the userdata is the same, no collisions should occur
+		if ((exa.userdata == exb.userdata) && exa.userdata) {
+			return;
+		}
+
+		typename BVHTREE_CLASS::ItemPairs &pairs_from = tree._pairs[p_from.id()];
+		typename BVHTREE_CLASS::ItemPairs &pairs_to = tree._pairs[p_to.id()];
+
+		void *ud_from = pairs_from.remove_pair_to(p_to);
+		pairs_to.remove_pair_to(p_from);
+
+		// callback
+		if (unpair_callback) {
+			unpair_callback(pair_callback_userdata, p_from, exa.userdata, exa.subindex, p_to, exb.userdata, exb.subindex, ud_from);
+		}
+	}
+
+	// returns true if unpair
+	bool _find_leavers_process_pair(typename BVHTREE_CLASS::ItemPairs &p_pairs_from, const BVHABB_CLASS &p_abb_from, BVHHandle p_from, BVHHandle p_to, bool p_full_check) {
+		BVHABB_CLASS abb_to;
+		tree.item_get_ABB(p_to, abb_to);
+
+		// do they overlap?
+		if (p_abb_from.intersects(abb_to)) {
+			// the full check for pairable / non pairable and mask changes is extra expense
+			// this need not be done in most cases (for speed) except in the case where set_pairable is called
+			// where the masks etc of the objects in question may have changed
+			if (!p_full_check) {
+				return false;
+			}
+			const typename BVHTREE_CLASS::ItemExtra &exa = _get_extra(p_from);
+			const typename BVHTREE_CLASS::ItemExtra &exb = _get_extra(p_to);
+
+			// one of the two must be pairable to still pair
+			// if neither are pairable, we always unpair
+			if (exa.pairable || exb.pairable) {
+				// the masks must still be compatible to pair
+				// i.e. if there is a hit between the two, then they should stay paired
+				if (tree._cull_pairing_mask_test_hit(exa.pairable_mask, exa.pairable_type, exb.pairable_mask, exb.pairable_type)) {
+					return false;
+				}
+			}
+		}
+
+		_unpair(p_from, p_to);
+		return true;
+	}
+
+	// find all the existing paired aabbs that are no longer
+	// paired, and send callbacks
+	void _find_leavers(BVHHandle p_handle, const BVHABB_CLASS &expanded_abb_from, bool p_full_check) {
+		typename BVHTREE_CLASS::ItemPairs &p_from = tree._pairs[p_handle.id()];
+
+		BVHABB_CLASS abb_from = expanded_abb_from;
+
+		// remove from pairing list for every partner
+		for (unsigned int n = 0; n < p_from.extended_pairs.size(); n++) {
+			BVHHandle h_to = p_from.extended_pairs[n].handle;
+			if (_find_leavers_process_pair(p_from, abb_from, p_handle, h_to, p_full_check)) {
+				// we need to keep the counter n up to date if we deleted a pair
+				// as the number of items in p_from.extended_pairs will have decreased by 1
+				// and we don't want to miss an item
+				n--;
+			}
+		}
+	}
+
+	// find NEW enterers, and send callbacks for them only
+	// handle a and b
+	void _collide(BVHHandle p_ha, BVHHandle p_hb) {
+		// only have to do this oneway, lower ID then higher ID
+		tree._handle_sort(p_ha, p_hb);
+
+		const typename BVHTREE_CLASS::ItemExtra &exa = _get_extra(p_ha);
+		const typename BVHTREE_CLASS::ItemExtra &exb = _get_extra(p_hb);
+
+		// if the userdata is the same, no collisions should occur
+		if ((exa.userdata == exb.userdata) && exa.userdata) {
+			return;
+		}
+
+		typename BVHTREE_CLASS::ItemPairs &p_from = tree._pairs[p_ha.id()];
+		typename BVHTREE_CLASS::ItemPairs &p_to = tree._pairs[p_hb.id()];
+
+		// does this pair exist already?
+		// or only check the one with lower number of pairs for greater speed
+		if (p_from.num_pairs <= p_to.num_pairs) {
+			if (p_from.contains_pair_to(p_hb)) {
+				return;
+			}
+		} else {
+			if (p_to.contains_pair_to(p_ha)) {
+				return;
+			}
+		}
+
+		// callback
+		void *callback_userdata = nullptr;
+
+		if (pair_callback) {
+			callback_userdata = pair_callback(pair_callback_userdata, p_ha, exa.userdata, exa.subindex, p_hb, exb.userdata, exb.subindex);
+		}
+
+		// new pair! .. only really need to store the userdata on the lower handle, but both have storage so...
+		p_from.add_pair_to(p_hb, callback_userdata);
+		p_to.add_pair_to(p_ha, callback_userdata);
+	}
+
+	// if we remove an item, we need to immediately remove the pairs, to prevent reading the pair after deletion
+	void _remove_pairs_containing(BVHHandle p_handle) {
+		typename BVHTREE_CLASS::ItemPairs &p_from = tree._pairs[p_handle.id()];
+
+		// remove from pairing list for every partner.
+		// can't easily use a for loop here, because removing changes the size of the list
+		while (p_from.extended_pairs.size()) {
+			BVHHandle h_to = p_from.extended_pairs[0].handle;
+			_unpair(p_handle, h_to);
+		}
+	}
+
+private:
+	const typename BVHTREE_CLASS::ItemExtra &_get_extra(BVHHandle p_handle) const {
+		return tree._extra[p_handle.id()];
+	}
+	const typename BVHTREE_CLASS::ItemRef &_get_ref(BVHHandle p_handle) const {
+		return tree._refs[p_handle.id()];
+	}
+
+	void _reset() {
+		changed_items.clear();
+		_tick++;
+	}
+
+	void _add_changed_item(BVHHandle p_handle, const Bounds &aabb, bool p_check_aabb = true) {
+		// Note that non pairable items can pair with pairable,
+		// so all types must be added to the list
+
+		// aabb check with expanded aabb. This greatly decreases processing
+		// at the cost of slightly less accurate pairing checks
+		// Note this pairing AABB is separate from the AABB in the actual tree
+		Bounds &expanded_aabb = tree._pairs[p_handle.id()].expanded_aabb;
+
+		// passing p_check_aabb false disables the optimization which prevents collision checks if
+		// the aabb hasn't changed. This is needed where set_pairable has been called, but the position
+		// has not changed.
+		if (p_check_aabb && expanded_aabb.encloses(aabb)) {
+			return;
+		}
+
+		// ALWAYS update the new expanded aabb, even if already changed once
+		// this tick, because it is vital that the AABB is kept up to date
+		expanded_aabb = aabb;
+		expanded_aabb.grow_by(tree._pairing_expansion);
+
+		// this code is to ensure that changed items only appear once on the updated list
+		// collision checking them multiple times is not needed, and repeats the same thing
+		uint32_t &last_updated_tick = tree._extra[p_handle.id()].last_updated_tick;
+
+		if (last_updated_tick == _tick) {
+			return; // already on changed list
+		}
+
+		// mark as on list
+		last_updated_tick = _tick;
+
+		// add to the list
+		changed_items.push_back(p_handle);
+	}
+
+	void _remove_changed_item(BVHHandle p_handle) {
+		// Care has to be taken here for items that are deleted. The ref ID
+		// could be reused on the same tick for new items. This is probably
+		// rare but should be taken into consideration
+
+		// callbacks
+		_remove_pairs_containing(p_handle);
+
+		// remove from changed items (not very efficient yet)
+		for (int n = 0; n < (int)changed_items.size(); n++) {
+			if (changed_items[n] == p_handle) {
+				changed_items.remove_unordered(n);
+
+				// because we are using an unordered remove,
+				// the last changed item will now be at spot 'n',
+				// and we need to redo it, so we prevent moving on to
+				// the next n at the next for iteration.
+				n--;
+			}
+		}
+
+		// reset the last updated tick (may not be necessary but just in case)
+		tree._extra[p_handle.id()].last_updated_tick = 0;
+	}
+
+	PairCallback pair_callback;
+	UnpairCallback unpair_callback;
+	void *pair_callback_userdata;
+	void *unpair_callback_userdata;
+
+	BVHTREE_CLASS tree;
+
+	// for collision pairing,
+	// maintain a list of all items moved etc on each frame / tick
+	LocalVector<BVHHandle, uint32_t, true> changed_items;
+	uint32_t _tick;
+
+public:
+	BVH_Manager() {
+		_tick = 1; // start from 1 so items with 0 indicate never updated
+		pair_callback = nullptr;
+		unpair_callback = nullptr;
+		pair_callback_userdata = nullptr;
+		unpair_callback_userdata = nullptr;
+	}
+};
+
+#undef BVHTREE_CLASS
+
+#endif // BVH_H
diff --git a/core/math/bvh_abb.h b/core/math/bvh_abb.h
new file mode 100644
index 0000000000..bd9a01a87e
--- /dev/null
+++ b/core/math/bvh_abb.h
@@ -0,0 +1,276 @@
+/*************************************************************************/
+/*  bvh_abb.h                                                            */
+/*************************************************************************/
+/*                       This file is part of:                           */
+/*                           GODOT ENGINE                                */
+/*                      https://godotengine.org                          */
+/*************************************************************************/
+/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
+/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
+/*                                                                       */
+/* Permission is hereby granted, free of charge, to any person obtaining */
+/* a copy of this software and associated documentation files (the       */
+/* "Software"), to deal in the Software without restriction, including   */
+/* without limitation the rights to use, copy, modify, merge, publish,   */
+/* distribute, sublicense, and/or sell copies of the Software, and to    */
+/* permit persons to whom the Software is furnished to do so, subject to */
+/* the following conditions:                                             */
+/*                                                                       */
+/* The above copyright notice and this permission notice shall be        */
+/* included in all copies or substantial portions of the Software.       */
+/*                                                                       */
+/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
+/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
+/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
+/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
+/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
+/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
+/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
+/*************************************************************************/
+
+#ifndef BVH_ABB_H
+#define BVH_ABB_H
+
+// special optimized version of axis aligned bounding box
+template <class Bounds = AABB, class Point = Vector3>
+struct BVH_ABB {
+	struct ConvexHull {
+		// convex hulls (optional)
+		const Plane *planes;
+		int num_planes;
+		const Vector3 *points;
+		int num_points;
+	};
+
+	struct Segment {
+		Point from;
+		Point to;
+	};
+
+	enum IntersectResult {
+		IR_MISS = 0,
+		IR_PARTIAL,
+		IR_FULL,
+	};
+
+	// we store mins with a negative value in order to test them with SIMD
+	Point min;
+	Point neg_max;
+
+	bool operator==(const BVH_ABB &o) const { return (min == o.min) && (neg_max == o.neg_max); }
+	bool operator!=(const BVH_ABB &o) const { return (*this == o) == false; }
+
+	void set(const Point &_min, const Point &_max) {
+		min = _min;
+		neg_max = -_max;
+	}
+
+	// to and from standard AABB
+	void from(const Bounds &p_aabb) {
+		min = p_aabb.position;
+		neg_max = -(p_aabb.position + p_aabb.size);
+	}
+
+	void to(Bounds &r_aabb) const {
+		r_aabb.position = min;
+		r_aabb.size = calculate_size();
+	}
+
+	void merge(const BVH_ABB &p_o) {
+		for (int axis = 0; axis < Point::AXIS_COUNT; ++axis) {
+			neg_max[axis] = MIN(neg_max[axis], p_o.neg_max[axis]);
+			min[axis] = MIN(min[axis], p_o.min[axis]);
+		}
+	}
+
+	Point calculate_size() const {
+		return -neg_max - min;
+	}
+
+	Point calculate_centre() const {
+		return Point((calculate_size() * 0.5) + min);
+	}
+
+	real_t get_proximity_to(const BVH_ABB &p_b) const {
+		const Point d = (min - neg_max) - (p_b.min - p_b.neg_max);
+		real_t proximity = 0.0;
+		for (int axis = 0; axis < Point::AXIS_COUNT; ++axis) {
+			proximity += Math::abs(d[axis]);
+		}
+		return proximity;
+	}
+
+	int select_by_proximity(const BVH_ABB &p_a, const BVH_ABB &p_b) const {
+		return (get_proximity_to(p_a) < get_proximity_to(p_b) ? 0 : 1);
+	}
+
+	uint32_t find_cutting_planes(const BVH_ABB::ConvexHull &p_hull, uint32_t *p_plane_ids) const {
+		uint32_t count = 0;
+
+		for (int n = 0; n < p_hull.num_planes; n++) {
+			const Plane &p = p_hull.planes[n];
+			if (intersects_plane(p)) {
+				p_plane_ids[count++] = n;
+			}
+		}
+
+		return count;
+	}
+
+	bool intersects_plane(const Plane &p_p) const {
+		Vector3 size = calculate_size();
+		Vector3 half_extents = size * 0.5;
+		Vector3 ofs = min + half_extents;
+
+		// forward side of plane?
+		Vector3 point_offset(
+				(p_p.normal.x < 0) ? -half_extents.x : half_extents.x,
+				(p_p.normal.y < 0) ? -half_extents.y : half_extents.y,
+				(p_p.normal.z < 0) ? -half_extents.z : half_extents.z);
+		Vector3 point = point_offset + ofs;
+
+		if (!p_p.is_point_over(point)) {
+			return false;
+		}
+
+		point = -point_offset + ofs;
+		if (p_p.is_point_over(point)) {
+			return false;
+		}
+
+		return true;
+	}
+
+	bool intersects_convex_optimized(const ConvexHull &p_hull, const uint32_t *p_plane_ids, uint32_t p_num_planes) const {
+		Vector3 size = calculate_size();
+		Vector3 half_extents = size * 0.5;
+		Vector3 ofs = min + half_extents;
+
+		for (unsigned int i = 0; i < p_num_planes; i++) {
+			const Plane &p = p_hull.planes[p_plane_ids[i]];
+			Vector3 point(
+					(p.normal.x > 0) ? -half_extents.x : half_extents.x,
+					(p.normal.y > 0) ? -half_extents.y : half_extents.y,
+					(p.normal.z > 0) ? -half_extents.z : half_extents.z);
+			point += ofs;
+			if (p.is_point_over(point)) {
+				return false;
+			}
+		}
+
+		return true;
+	}
+
+	bool intersects_convex_partial(const ConvexHull &p_hull) const {
+		Bounds bb;
+		to(bb);
+		return bb.intersects_convex_shape(p_hull.planes, p_hull.num_planes, p_hull.points, p_hull.num_points);
+	}
+
+	IntersectResult intersects_convex(const ConvexHull &p_hull) const {
+		if (intersects_convex_partial(p_hull)) {
+			// fully within? very important for tree checks
+			if (is_within_convex(p_hull)) {
+				return IR_FULL;
+			}
+
+			return IR_PARTIAL;
+		}
+
+		return IR_MISS;
+	}
+
+	bool is_within_convex(const ConvexHull &p_hull) const {
+		// use half extents routine
+		Bounds bb;
+		to(bb);
+		return bb.inside_convex_shape(p_hull.planes, p_hull.num_planes);
+	}
+
+	bool is_point_within_hull(const ConvexHull &p_hull, const Vector3 &p_pt) const {
+		for (int n = 0; n < p_hull.num_planes; n++) {
+			if (p_hull.planes[n].distance_to(p_pt) > 0.0f) {
+				return false;
+			}
+		}
+		return true;
+	}
+
+	bool intersects_segment(const Segment &p_s) const {
+		Bounds bb;
+		to(bb);
+		return bb.intersects_segment(p_s.from, p_s.to);
+	}
+
+	bool intersects_point(const Point &p_pt) const {
+		if (_any_lessthan(-p_pt, neg_max)) {
+			return false;
+		}
+		if (_any_lessthan(p_pt, min)) {
+			return false;
+		}
+		return true;
+	}
+
+	bool intersects(const BVH_ABB &p_o) const {
+		if (_any_morethan(p_o.min, -neg_max)) {
+			return false;
+		}
+		if (_any_morethan(min, -p_o.neg_max)) {
+			return false;
+		}
+		return true;
+	}
+
+	bool is_other_within(const BVH_ABB &p_o) const {
+		if (_any_lessthan(p_o.neg_max, neg_max)) {
+			return false;
+		}
+		if (_any_lessthan(p_o.min, min)) {
+			return false;
+		}
+		return true;
+	}
+
+	void grow(const Point &p_change) {
+		neg_max -= p_change;
+		min -= p_change;
+	}
+
+	void expand(real_t p_change) {
+		Point change;
+		change.set_all(p_change);
+		grow(change);
+	}
+
+	// Actually surface area metric.
+	float get_area() const {
+		Point d = calculate_size();
+		return 2.0f * (d.x * d.y + d.y * d.z + d.z * d.x);
+	}
+
+	void set_to_max_opposite_extents() {
+		neg_max.set_all(FLT_MAX);
+		min = neg_max;
+	}
+
+	bool _any_morethan(const Point &p_a, const Point &p_b) const {
+		for (int axis = 0; axis < Point::AXIS_COUNT; ++axis) {
+			if (p_a[axis] > p_b[axis]) {
+				return true;
+			}
+		}
+		return false;
+	}
+
+	bool _any_lessthan(const Point &p_a, const Point &p_b) const {
+		for (int axis = 0; axis < Point::AXIS_COUNT; ++axis) {
+			if (p_a[axis] < p_b[axis]) {
+				return true;
+			}
+		}
+		return false;
+	}
+};
+
+#endif // BVH_ABB_H
diff --git a/core/math/bvh_cull.inc b/core/math/bvh_cull.inc
new file mode 100644
index 0000000000..cba8ea6cb3
--- /dev/null
+++ b/core/math/bvh_cull.inc
@@ -0,0 +1,534 @@
+public:
+// cull parameters is a convenient way of passing a bunch
+// of arguments through the culling functions without
+// writing loads of code. Not all members are used for some cull checks
+struct CullParams {
+	int result_count_overall; // both trees
+	int result_count; // this tree only
+	int result_max;
+	T **result_array;
+	int *subindex_array;
+
+	// nobody truly understands how masks are intended to work.
+	uint32_t mask;
+	uint32_t pairable_type;
+
+	// optional components for different tests
+	Vector3 point;
+	BVHABB_CLASS abb;
+	typename BVHABB_CLASS::ConvexHull hull;
+	typename BVHABB_CLASS::Segment segment;
+
+	// when collision testing, non pairable moving items
+	// only need to be tested against the pairable tree.
+	// collisions with other non pairable items are irrelevant.
+	bool test_pairable_only;
+};
+
+private:
+void _cull_translate_hits(CullParams &p) {
+	int num_hits = _cull_hits.size();
+	int left = p.result_max - p.result_count_overall;
+
+	if (num_hits > left) {
+		num_hits = left;
+	}
+
+	int out_n = p.result_count_overall;
+
+	for (int n = 0; n < num_hits; n++) {
+		uint32_t ref_id = _cull_hits[n];
+
+		const ItemExtra &ex = _extra[ref_id];
+		p.result_array[out_n] = ex.userdata;
+
+		if (p.subindex_array) {
+			p.subindex_array[out_n] = ex.subindex;
+		}
+
+		out_n++;
+	}
+
+	p.result_count = num_hits;
+	p.result_count_overall += num_hits;
+}
+
+public:
+int cull_convex(CullParams &r_params, bool p_translate_hits = true) {
+	_cull_hits.clear();
+	r_params.result_count = 0;
+
+	for (int n = 0; n < NUM_TREES; n++) {
+		if (_root_node_id[n] == BVHCommon::INVALID) {
+			continue;
+		}
+
+		_cull_convex_iterative(_root_node_id[n], r_params);
+	}
+
+	if (p_translate_hits) {
+		_cull_translate_hits(r_params);
+	}
+
+	return r_params.result_count;
+}
+
+int cull_segment(CullParams &r_params, bool p_translate_hits = true) {
+	_cull_hits.clear();
+	r_params.result_count = 0;
+
+	for (int n = 0; n < NUM_TREES; n++) {
+		if (_root_node_id[n] == BVHCommon::INVALID) {
+			continue;
+		}
+
+		_cull_segment_iterative(_root_node_id[n], r_params);
+	}
+
+	if (p_translate_hits) {
+		_cull_translate_hits(r_params);
+	}
+
+	return r_params.result_count;
+}
+
+int cull_point(CullParams &r_params, bool p_translate_hits = true) {
+	_cull_hits.clear();
+	r_params.result_count = 0;
+
+	for (int n = 0; n < NUM_TREES; n++) {
+		if (_root_node_id[n] == BVHCommon::INVALID) {
+			continue;
+		}
+
+		_cull_point_iterative(_root_node_id[n], r_params);
+	}
+
+	if (p_translate_hits) {
+		_cull_translate_hits(r_params);
+	}
+
+	return r_params.result_count;
+}
+
+int cull_aabb(CullParams &r_params, bool p_translate_hits = true) {
+	_cull_hits.clear();
+	r_params.result_count = 0;
+
+	for (int n = 0; n < NUM_TREES; n++) {
+		if (_root_node_id[n] == BVHCommon::INVALID) {
+			continue;
+		}
+
+		if ((n == 0) && r_params.test_pairable_only) {
+			continue;
+		}
+
+		_cull_aabb_iterative(_root_node_id[n], r_params);
+	}
+
+	if (p_translate_hits) {
+		_cull_translate_hits(r_params);
+	}
+
+	return r_params.result_count;
+}
+
+bool _cull_hits_full(const CullParams &p) {
+	// instead of checking every hit, we can do a lazy check for this condition.
+	// it isn't a problem if we write too much _cull_hits because they only the
+	// result_max amount will be translated and outputted. But we might as
+	// well stop our cull checks after the maximum has been reached.
+	return (int)_cull_hits.size() >= p.result_max;
+}
+
+// write this logic once for use in all routines
+// double check this as a possible source of bugs in future.
+bool _cull_pairing_mask_test_hit(uint32_t p_maskA, uint32_t p_typeA, uint32_t p_maskB, uint32_t p_typeB) const {
+	// double check this as a possible source of bugs in future.
+	bool A_match_B = p_maskA & p_typeB;
+
+	if (!A_match_B) {
+		bool B_match_A = p_maskB & p_typeA;
+		if (!B_match_A) {
+			return false;
+		}
+	}
+
+	return true;
+}
+
+void _cull_hit(uint32_t p_ref_id, CullParams &p) {
+	// take into account masks etc
+	// this would be more efficient to do before plane checks,
+	// but done here for ease to get started
+	if (USE_PAIRS) {
+		const ItemExtra &ex = _extra[p_ref_id];
+
+		if (!_cull_pairing_mask_test_hit(p.mask, p.pairable_type, ex.pairable_mask, ex.pairable_type)) {
+			return;
+		}
+	}
+
+	_cull_hits.push_back(p_ref_id);
+}
+
+bool _cull_segment_iterative(uint32_t p_node_id, CullParams &r_params) {
+	// our function parameters to keep on a stack
+	struct CullSegParams {
+		uint32_t node_id;
+	};
+
+	// most of the iterative functionality is contained in this helper class
+	BVH_IterativeInfo<CullSegParams> ii;
+
+	// alloca must allocate the stack from this function, it cannot be allocated in the
+	// helper class
+	ii.stack = (CullSegParams *)alloca(ii.get_alloca_stacksize());
+
+	// seed the stack
+	ii.get_first()->node_id = p_node_id;
+
+	CullSegParams csp;
+
+	// while there are still more nodes on the stack
+	while (ii.pop(csp)) {
+		TNode &tnode = _nodes[csp.node_id];
+
+		if (tnode.is_leaf()) {
+			// lazy check for hits full up condition
+			if (_cull_hits_full(r_params)) {
+				return false;
+			}
+
+			TLeaf &leaf = _node_get_leaf(tnode);
+
+			// test children individually
+			for (int n = 0; n < leaf.num_items; n++) {
+				const BVHABB_CLASS &aabb = leaf.get_aabb(n);
+
+				if (aabb.intersects_segment(r_params.segment)) {
+					uint32_t child_id = leaf.get_item_ref_id(n);
+
+					// register hit
+					_cull_hit(child_id, r_params);
+				}
+			}
+		} else {
+			// test children individually
+			for (int n = 0; n < tnode.num_children; n++) {
+				uint32_t child_id = tnode.children[n];
+				const BVHABB_CLASS &child_abb = _nodes[child_id].aabb;
+
+				if (child_abb.intersects_segment(r_params.segment)) {
+					// add to the stack
+					CullSegParams *child = ii.request();
+					child->node_id = child_id;
+				}
+			}
+		}
+
+	} // while more nodes to pop
+
+	// true indicates results are not full
+	return true;
+}
+
+bool _cull_point_iterative(uint32_t p_node_id, CullParams &r_params) {
+	// our function parameters to keep on a stack
+	struct CullPointParams {
+		uint32_t node_id;
+	};
+
+	// most of the iterative functionality is contained in this helper class
+	BVH_IterativeInfo<CullPointParams> ii;
+
+	// alloca must allocate the stack from this function, it cannot be allocated in the
+	// helper class
+	ii.stack = (CullPointParams *)alloca(ii.get_alloca_stacksize());
+
+	// seed the stack
+	ii.get_first()->node_id = p_node_id;
+
+	CullPointParams cpp;
+
+	// while there are still more nodes on the stack
+	while (ii.pop(cpp)) {
+		TNode &tnode = _nodes[cpp.node_id];
+		// no hit with this node?
+		if (!tnode.aabb.intersects_point(r_params.point)) {
+			continue;
+		}
+
+		if (tnode.is_leaf()) {
+			// lazy check for hits full up condition
+			if (_cull_hits_full(r_params)) {
+				return false;
+			}
+
+			TLeaf &leaf = _node_get_leaf(tnode);
+
+			// test children individually
+			for (int n = 0; n < leaf.num_items; n++) {
+				if (leaf.get_aabb(n).intersects_point(r_params.point)) {
+					uint32_t child_id = leaf.get_item_ref_id(n);
+
+					// register hit
+					_cull_hit(child_id, r_params);
+				}
+			}
+		} else {
+			// test children individually
+			for (int n = 0; n < tnode.num_children; n++) {
+				uint32_t child_id = tnode.children[n];
+
+				// add to the stack
+				CullPointParams *child = ii.request();
+				child->node_id = child_id;
+			}
+		}
+
+	} // while more nodes to pop
+
+	// true indicates results are not full
+	return true;
+}
+
+bool _cull_aabb_iterative(uint32_t p_node_id, CullParams &r_params, bool p_fully_within = false) {
+	// our function parameters to keep on a stack
+	struct CullAABBParams {
+		uint32_t node_id;
+		bool fully_within;
+	};
+
+	// most of the iterative functionality is contained in this helper class
+	BVH_IterativeInfo<CullAABBParams> ii;
+
+	// alloca must allocate the stack from this function, it cannot be allocated in the
+	// helper class
+	ii.stack = (CullAABBParams *)alloca(ii.get_alloca_stacksize());
+
+	// seed the stack
+	ii.get_first()->node_id = p_node_id;
+	ii.get_first()->fully_within = p_fully_within;
+
+	CullAABBParams cap;
+
+	// while there are still more nodes on the stack
+	while (ii.pop(cap)) {
+		TNode &tnode = _nodes[cap.node_id];
+
+		if (tnode.is_leaf()) {
+			// lazy check for hits full up condition
+			if (_cull_hits_full(r_params)) {
+				return false;
+			}
+
+			TLeaf &leaf = _node_get_leaf(tnode);
+
+			// if fully within we can just add all items
+			// as long as they pass mask checks
+			if (cap.fully_within) {
+				for (int n = 0; n < leaf.num_items; n++) {
+					uint32_t child_id = leaf.get_item_ref_id(n);
+
+					// register hit
+					_cull_hit(child_id, r_params);
+				}
+			} else {
+				for (int n = 0; n < leaf.num_items; n++) {
+					const BVHABB_CLASS &aabb = leaf.get_aabb(n);
+
+					if (aabb.intersects(r_params.abb)) {
+						uint32_t child_id = leaf.get_item_ref_id(n);
+
+						// register hit
+						_cull_hit(child_id, r_params);
+					}
+				}
+			} // not fully within
+		} else {
+			if (!cap.fully_within) {
+				// test children individually
+				for (int n = 0; n < tnode.num_children; n++) {
+					uint32_t child_id = tnode.children[n];
+					const BVHABB_CLASS &child_abb = _nodes[child_id].aabb;
+
+					if (child_abb.intersects(r_params.abb)) {
+						// is the node totally within the aabb?
+						bool fully_within = r_params.abb.is_other_within(child_abb);
+
+						// add to the stack
+						CullAABBParams *child = ii.request();
+
+						// should always return valid child
+						child->node_id = child_id;
+						child->fully_within = fully_within;
+					}
+				}
+			} else {
+				for (int n = 0; n < tnode.num_children; n++) {
+					uint32_t child_id = tnode.children[n];
+
+					// add to the stack
+					CullAABBParams *child = ii.request();
+
+					// should always return valid child
+					child->node_id = child_id;
+					child->fully_within = true;
+				}
+			}
+		}
+
+	} // while more nodes to pop
+
+	// true indicates results are not full
+	return true;
+}
+
+// returns full up with results
+bool _cull_convex_iterative(uint32_t p_node_id, CullParams &r_params, bool p_fully_within = false) {
+	// our function parameters to keep on a stack
+	struct CullConvexParams {
+		uint32_t node_id;
+		bool fully_within;
+	};
+
+	// most of the iterative functionality is contained in this helper class
+	BVH_IterativeInfo<CullConvexParams> ii;
+
+	// alloca must allocate the stack from this function, it cannot be allocated in the
+	// helper class
+	ii.stack = (CullConvexParams *)alloca(ii.get_alloca_stacksize());
+
+	// seed the stack
+	ii.get_first()->node_id = p_node_id;
+	ii.get_first()->fully_within = p_fully_within;
+
+	// preallocate these as a once off to be reused
+	uint32_t max_planes = r_params.hull.num_planes;
+	uint32_t *plane_ids = (uint32_t *)alloca(sizeof(uint32_t) * max_planes);
+
+	CullConvexParams ccp;
+
+	// while there are still more nodes on the stack
+	while (ii.pop(ccp)) {
+		const TNode &tnode = _nodes[ccp.node_id];
+
+		if (!ccp.fully_within) {
+			typename BVHABB_CLASS::IntersectResult res = tnode.aabb.intersects_convex(r_params.hull);
+
+			switch (res) {
+				default: {
+					continue; // miss, just move on to the next node in the stack
+				} break;
+				case BVHABB_CLASS::IR_PARTIAL: {
+				} break;
+				case BVHABB_CLASS::IR_FULL: {
+					ccp.fully_within = true;
+				} break;
+			}
+
+		} // if not fully within already
+
+		if (tnode.is_leaf()) {
+			// lazy check for hits full up condition
+			if (_cull_hits_full(r_params)) {
+				return false;
+			}
+
+			const TLeaf &leaf = _node_get_leaf(tnode);
+
+			// if fully within, simply add all items to the result
+			// (taking into account masks)
+			if (ccp.fully_within) {
+				for (int n = 0; n < leaf.num_items; n++) {
+					uint32_t child_id = leaf.get_item_ref_id(n);
+
+					// register hit
+					_cull_hit(child_id, r_params);
+				}
+
+			} else {
+				// we can either use a naive check of all the planes against the AABB,
+				// or an optimized check, which finds in advance which of the planes can possibly
+				// cut the AABB, and only tests those. This can be much faster.
+#define BVH_CONVEX_CULL_OPTIMIZED
+#ifdef BVH_CONVEX_CULL_OPTIMIZED
+				// first find which planes cut the aabb
+				uint32_t num_planes = tnode.aabb.find_cutting_planes(r_params.hull, plane_ids);
+				BVH_ASSERT(num_planes <= max_planes);
+
+//#define BVH_CONVEX_CULL_OPTIMIZED_RIGOR_CHECK
+#ifdef BVH_CONVEX_CULL_OPTIMIZED_RIGOR_CHECK
+				// rigorous check
+				uint32_t results[MAX_ITEMS];
+				uint32_t num_results = 0;
+#endif
+
+				// test children individually
+				for (int n = 0; n < leaf.num_items; n++) {
+					//const Item &item = leaf.get_item(n);
+					const BVHABB_CLASS &aabb = leaf.get_aabb(n);
+
+					if (aabb.intersects_convex_optimized(r_params.hull, plane_ids, num_planes)) {
+						uint32_t child_id = leaf.get_item_ref_id(n);
+
+#ifdef BVH_CONVEX_CULL_OPTIMIZED_RIGOR_CHECK
+						results[num_results++] = child_id;
+#endif
+
+						// register hit
+						_cull_hit(child_id, r_params);
+					}
+				}
+
+#ifdef BVH_CONVEX_CULL_OPTIMIZED_RIGOR_CHECK
+				uint32_t test_count = 0;
+
+				for (int n = 0; n < leaf.num_items; n++) {
+					const BVHABB_CLASS &aabb = leaf.get_aabb(n);
+
+					if (aabb.intersects_convex_partial(r_params.hull)) {
+						uint32_t child_id = leaf.get_item_ref_id(n);
+
+						CRASH_COND(child_id != results[test_count++]);
+						CRASH_COND(test_count > num_results);
+					}
+				}
+#endif
+
+#else
+				// not BVH_CONVEX_CULL_OPTIMIZED
+				// test children individually
+				for (int n = 0; n < leaf.num_items; n++) {
+					const BVHABB_CLASS &aabb = leaf.get_aabb(n);
+
+					if (aabb.intersects_convex_partial(r_params.hull)) {
+						uint32_t child_id = leaf.get_item_ref_id(n);
+
+						// full up with results? exit early, no point in further testing
+						if (!_cull_hit(child_id, r_params))
+							return false;
+					}
+				}
+#endif // BVH_CONVEX_CULL_OPTIMIZED
+			} // if not fully within
+		} else {
+			for (int n = 0; n < tnode.num_children; n++) {
+				uint32_t child_id = tnode.children[n];
+
+				// add to the stack
+				CullConvexParams *child = ii.request();
+
+				// should always return valid child
+				child->node_id = child_id;
+				child->fully_within = ccp.fully_within;
+			}
+		}
+
+	} // while more nodes to pop
+
+	// true indicates results are not full
+	return true;
+}
diff --git a/core/math/bvh_debug.inc b/core/math/bvh_debug.inc
new file mode 100644
index 0000000000..a97304334c
--- /dev/null
+++ b/core/math/bvh_debug.inc
@@ -0,0 +1,68 @@
+public:
+#ifdef BVH_VERBOSE
+void _debug_recursive_print_tree(int p_tree_id) const {
+	if (_root_node_id[p_tree_id] != BVHCommon::INVALID)
+		_debug_recursive_print_tree_node(_root_node_id[p_tree_id]);
+}
+
+String _debug_aabb_to_string(const BVHABB_CLASS &aabb) const {
+	String sz = "(";
+	sz += itos(aabb.min.x);
+	sz += " ~ ";
+	sz += itos(-aabb.neg_max.x);
+	sz += ") (";
+
+	sz += itos(aabb.min.y);
+	sz += " ~ ";
+	sz += itos(-aabb.neg_max.y);
+	sz += ") (";
+
+	sz += itos(aabb.min.z);
+	sz += " ~ ";
+	sz += itos(-aabb.neg_max.z);
+	sz += ") ";
+
+	Vector3 size = aabb.calculate_size();
+	float vol = size.x * size.y * size.z;
+	sz += "vol " + itos(vol);
+
+	return sz;
+}
+
+void _debug_recursive_print_tree_node(uint32_t p_node_id, int depth = 0) const {
+	const TNode &tnode = _nodes[p_node_id];
+
+	String sz = "";
+	for (int n = 0; n < depth; n++) {
+		sz += "\t";
+	}
+	sz += itos(p_node_id);
+
+	if (tnode.is_leaf()) {
+		sz += " L";
+		sz += itos(tnode.height) + " ";
+		const TLeaf &leaf = _node_get_leaf(tnode);
+
+		sz += "[";
+		for (int n = 0; n < leaf.num_items; n++) {
+			if (n)
+				sz += ", ";
+			sz += "r";
+			sz += itos(leaf.get_item_ref_id(n));
+		}
+		sz += "]  ";
+	} else {
+		sz += " N";
+		sz += itos(tnode.height) + " ";
+	}
+
+	sz += _debug_aabb_to_string(tnode.aabb);
+	print_line(sz);
+
+	if (!tnode.is_leaf()) {
+		for (int n = 0; n < tnode.num_children; n++) {
+			_debug_recursive_print_tree_node(tnode.children[n], depth + 1);
+		}
+	}
+}
+#endif
diff --git a/core/math/bvh_integrity.inc b/core/math/bvh_integrity.inc
new file mode 100644
index 0000000000..02e9d30097
--- /dev/null
+++ b/core/math/bvh_integrity.inc
@@ -0,0 +1,42 @@
+void _integrity_check_all() {
+#ifdef BVH_INTEGRITY_CHECKS
+	for (int n = 0; n < NUM_TREES; n++) {
+		uint32_t root = _root_node_id[n];
+		if (root != BVHCommon::INVALID) {
+			_integrity_check_down(root);
+		}
+	}
+#endif
+}
+
+void _integrity_check_up(uint32_t p_node_id) {
+	TNode &node = _nodes[p_node_id];
+
+	BVHABB_CLASS abb = node.aabb;
+	node_update_aabb(node);
+
+	BVHABB_CLASS abb2 = node.aabb;
+	abb2.expand(-_node_expansion);
+
+	CRASH_COND(!abb.is_other_within(abb2));
+}
+
+void _integrity_check_down(uint32_t p_node_id) {
+	const TNode &node = _nodes[p_node_id];
+
+	if (node.is_leaf()) {
+		_integrity_check_up(p_node_id);
+	} else {
+		CRASH_COND(node.num_children != 2);
+
+		for (int n = 0; n < node.num_children; n++) {
+			uint32_t child_id = node.children[n];
+
+			// check the children parent pointers are correct
+			TNode &child = _nodes[child_id];
+			CRASH_COND(child.parent_id != p_node_id);
+
+			_integrity_check_down(child_id);
+		}
+	}
+}
diff --git a/core/math/bvh_logic.inc b/core/math/bvh_logic.inc
new file mode 100644
index 0000000000..d84c3f7830
--- /dev/null
+++ b/core/math/bvh_logic.inc
@@ -0,0 +1,230 @@
+
+// for slow incremental optimization, we will periodically remove each
+// item from the tree and reinsert, to give it a chance to find a better position
+void _logic_item_remove_and_reinsert(uint32_t p_ref_id) {
+	// get the reference
+	ItemRef &ref = _refs[p_ref_id];
+
+	// no need to optimize inactive items
+	if (!ref.is_active()) {
+		return;
+	}
+
+	// special case of debug draw
+	if (ref.item_id == BVHCommon::INVALID) {
+		return;
+	}
+
+	BVH_ASSERT(ref.tnode_id != BVHCommon::INVALID);
+
+	// some overlay elaborate way to find out which tree the node is in!
+	BVHHandle temp_handle;
+	temp_handle.set_id(p_ref_id);
+	_current_tree = _handle_get_tree_id(temp_handle);
+
+	// remove and reinsert
+	BVHABB_CLASS abb;
+	node_remove_item(p_ref_id, &abb);
+
+	// we must choose where to add to tree
+	ref.tnode_id = _logic_choose_item_add_node(_root_node_id[_current_tree], abb);
+	_node_add_item(ref.tnode_id, p_ref_id, abb);
+
+	refit_upward_and_balance(ref.tnode_id);
+}
+
+// from randy gaul balance function
+BVHABB_CLASS _logic_abb_merge(const BVHABB_CLASS &a, const BVHABB_CLASS &b) {
+	BVHABB_CLASS c = a;
+	c.merge(b);
+	return c;
+}
+
+//--------------------------------------------------------------------------------------------------
+/**
+@file	q3DynamicAABBTree.h
+@author	Randy Gaul
+@date	10/10/2014
+	Copyright (c) 2014 Randy Gaul http://www.randygaul.net
+	This software is provided 'as-is', without any express or implied
+	warranty. In no event will the authors be held liable for any damages
+	arising from the use of this software.
+	Permission is granted to anyone to use this software for any purpose,
+	including commercial applications, and to alter it and redistribute it
+	freely, subject to the following restrictions:
+	  1. The origin of this software must not be misrepresented; you must not
+	     claim that you wrote the original software. If you use this software
+	     in a product, an acknowledgment in the product documentation would be
+	     appreciated but is not required.
+	  2. Altered source versions must be plainly marked as such, and must not
+	     be misrepresented as being the original software.
+	  3. This notice may not be removed or altered from any source distribution.
+*/
+//--------------------------------------------------------------------------------------------------
+
+// This function is based on the 'Balance' function from Randy Gaul's qu3e
+// https://github.com/RandyGaul/qu3e
+// It is MODIFIED from qu3e version.
+// This is the only function used (and _logic_abb_merge helper function).
+int32_t _logic_balance(int32_t iA) {
+	//	return iA; // uncomment this to bypass balance
+
+	TNode *A = &_nodes[iA];
+
+	if (A->is_leaf() || A->height == 1) {
+		return iA;
+	}
+
+	/*      A
+	      /   \
+	     B     C
+	    / \   / \
+	   D   E F   G
+	*/
+
+	CRASH_COND(A->num_children != 2);
+	int32_t iB = A->children[0];
+	int32_t iC = A->children[1];
+	TNode *B = &_nodes[iB];
+	TNode *C = &_nodes[iC];
+
+	int32_t balance = C->height - B->height;
+
+	// C is higher, promote C
+	if (balance > 1) {
+		int32_t iF = C->children[0];
+		int32_t iG = C->children[1];
+		TNode *F = &_nodes[iF];
+		TNode *G = &_nodes[iG];
+
+		// grandParent point to C
+		if (A->parent_id != BVHCommon::INVALID) {
+			if (_nodes[A->parent_id].children[0] == iA) {
+				_nodes[A->parent_id].children[0] = iC;
+
+			} else {
+				_nodes[A->parent_id].children[1] = iC;
+			}
+		} else {
+			// check this .. seems dodgy
+			change_root_node(iC);
+		}
+
+		// Swap A and C
+		C->children[0] = iA;
+		C->parent_id = A->parent_id;
+		A->parent_id = iC;
+
+		// Finish rotation
+		if (F->height > G->height) {
+			C->children[1] = iF;
+			A->children[1] = iG;
+			G->parent_id = iA;
+			A->aabb = _logic_abb_merge(B->aabb, G->aabb);
+			C->aabb = _logic_abb_merge(A->aabb, F->aabb);
+
+			A->height = 1 + MAX(B->height, G->height);
+			C->height = 1 + MAX(A->height, F->height);
+		}
+
+		else {
+			C->children[1] = iG;
+			A->children[1] = iF;
+			F->parent_id = iA;
+			A->aabb = _logic_abb_merge(B->aabb, F->aabb);
+			C->aabb = _logic_abb_merge(A->aabb, G->aabb);
+
+			A->height = 1 + MAX(B->height, F->height);
+			C->height = 1 + MAX(A->height, G->height);
+		}
+
+		return iC;
+	}
+
+	// B is higher, promote B
+	else if (balance < -1) {
+		int32_t iD = B->children[0];
+		int32_t iE = B->children[1];
+		TNode *D = &_nodes[iD];
+		TNode *E = &_nodes[iE];
+
+		// grandParent point to B
+		if (A->parent_id != BVHCommon::INVALID) {
+			if (_nodes[A->parent_id].children[0] == iA) {
+				_nodes[A->parent_id].children[0] = iB;
+			} else {
+				_nodes[A->parent_id].children[1] = iB;
+			}
+		}
+
+		else {
+			// check this .. seems dodgy
+			change_root_node(iB);
+		}
+
+		// Swap A and B
+		B->children[1] = iA;
+		B->parent_id = A->parent_id;
+		A->parent_id = iB;
+
+		// Finish rotation
+		if (D->height > E->height) {
+			B->children[0] = iD;
+			A->children[0] = iE;
+			E->parent_id = iA;
+			A->aabb = _logic_abb_merge(C->aabb, E->aabb);
+			B->aabb = _logic_abb_merge(A->aabb, D->aabb);
+
+			A->height = 1 + MAX(C->height, E->height);
+			B->height = 1 + MAX(A->height, D->height);
+		}
+
+		else {
+			B->children[0] = iE;
+			A->children[0] = iD;
+			D->parent_id = iA;
+			A->aabb = _logic_abb_merge(C->aabb, D->aabb);
+			B->aabb = _logic_abb_merge(A->aabb, E->aabb);
+
+			A->height = 1 + MAX(C->height, D->height);
+			B->height = 1 + MAX(A->height, E->height);
+		}
+
+		return iB;
+	}
+
+	return iA;
+}
+
+// either choose an existing node to add item to, or create a new node and return this
+uint32_t _logic_choose_item_add_node(uint32_t p_node_id, const BVHABB_CLASS &p_aabb) {
+	while (true) {
+		BVH_ASSERT(p_node_id != BVHCommon::INVALID);
+		TNode &tnode = _nodes[p_node_id];
+
+		if (tnode.is_leaf()) {
+			// if a leaf, and non full, use this to add to
+			if (!node_is_leaf_full(tnode)) {
+				return p_node_id;
+			}
+
+			// else split the leaf, and use one of the children to add to
+			return split_leaf(p_node_id, p_aabb);
+		}
+
+		// this should not happen???
+		// is still happening, need to debug and find circumstances. Is not that serious
+		// but would be nice to prevent. I think it only happens with the root node.
+		if (tnode.num_children == 1) {
+			WARN_PRINT_ONCE("BVH::recursive_choose_item_add_node, node with 1 child, recovering");
+			p_node_id = tnode.children[0];
+		} else {
+			BVH_ASSERT(tnode.num_children == 2);
+			TNode &childA = _nodes[tnode.children[0]];
+			TNode &childB = _nodes[tnode.children[1]];
+			int which = p_aabb.select_by_proximity(childA.aabb, childB.aabb);
+
+			p_node_id = tnode.children[which];
+		}
+	}
+}
diff --git a/core/math/bvh_misc.inc b/core/math/bvh_misc.inc
new file mode 100644
index 0000000000..71aa0e4fe0
--- /dev/null
+++ b/core/math/bvh_misc.inc
@@ -0,0 +1,55 @@
+
+int _handle_get_tree_id(BVHHandle p_handle) const {
+	if (USE_PAIRS) {
+		int tree = 0;
+		if (_extra[p_handle.id()].pairable) {
+			tree = 1;
+		}
+		return tree;
+	}
+	return 0;
+}
+
+public:
+void _handle_sort(BVHHandle &p_ha, BVHHandle &p_hb) const {
+	if (p_ha.id() > p_hb.id()) {
+		BVHHandle temp = p_hb;
+		p_hb = p_ha;
+		p_ha = temp;
+	}
+}
+
+private:
+void create_root_node(int p_tree) {
+	// if there is no root node, create one
+	if (_root_node_id[p_tree] == BVHCommon::INVALID) {
+		uint32_t root_node_id;
+		TNode *node = _nodes.request(root_node_id);
+		node->clear();
+		_root_node_id[p_tree] = root_node_id;
+
+		// make the root node a leaf
+		uint32_t leaf_id;
+		TLeaf *leaf = _leaves.request(leaf_id);
+		leaf->clear();
+		node->neg_leaf_id = -(int)leaf_id;
+	}
+}
+
+bool node_is_leaf_full(TNode &tnode) const {
+	const TLeaf &leaf = _node_get_leaf(tnode);
+	return leaf.is_full();
+}
+
+public:
+TLeaf &_node_get_leaf(TNode &tnode) {
+	BVH_ASSERT(tnode.is_leaf());
+	return _leaves[tnode.get_leaf_id()];
+}
+
+const TLeaf &_node_get_leaf(const TNode &tnode) const {
+	BVH_ASSERT(tnode.is_leaf());
+	return _leaves[tnode.get_leaf_id()];
+}
+
+private:
diff --git a/core/math/bvh_pair.inc b/core/math/bvh_pair.inc
new file mode 100644
index 0000000000..839db59a3a
--- /dev/null
+++ b/core/math/bvh_pair.inc
@@ -0,0 +1,62 @@
+public:
+// note .. maybe this can be attached to another node structure?
+// depends which works best for cache.
+struct ItemPairs {
+	struct Link {
+		void set(BVHHandle h, void *ud) {
+			handle = h;
+			userdata = ud;
+		}
+		BVHHandle handle;
+		void *userdata;
+	};
+
+	void clear() {
+		num_pairs = 0;
+		extended_pairs.reset();
+		expanded_aabb = Bounds();
+	}
+
+	Bounds expanded_aabb;
+
+	// maybe we can just use the number in the vector TODO
+	int32_t num_pairs;
+	LocalVector<Link> extended_pairs;
+
+	void add_pair_to(BVHHandle h, void *p_userdata) {
+		Link temp;
+		temp.set(h, p_userdata);
+
+		extended_pairs.push_back(temp);
+		num_pairs++;
+	}
+
+	uint32_t find_pair_to(BVHHandle h) const {
+		for (int n = 0; n < num_pairs; n++) {
+			if (extended_pairs[n].handle == h) {
+				return n;
+			}
+		}
+		return -1;
+	}
+
+	bool contains_pair_to(BVHHandle h) const {
+		return find_pair_to(h) != BVHCommon::INVALID;
+	}
+
+	// return success
+	void *remove_pair_to(BVHHandle h) {
+		void *userdata = nullptr;
+
+		for (int n = 0; n < num_pairs; n++) {
+			if (extended_pairs[n].handle == h) {
+				userdata = extended_pairs[n].userdata;
+				extended_pairs.remove_unordered(n);
+				num_pairs--;
+				break;
+			}
+		}
+
+		return userdata;
+	}
+};
diff --git a/core/math/bvh_public.inc b/core/math/bvh_public.inc
new file mode 100644
index 0000000000..f1b6d6b1bf
--- /dev/null
+++ b/core/math/bvh_public.inc
@@ -0,0 +1,421 @@
+public:
+BVHHandle item_add(T *p_userdata, bool p_active, const Bounds &p_aabb, int32_t p_subindex, bool p_pairable, uint32_t p_pairable_type, uint32_t p_pairable_mask, bool p_invisible = false) {
+#ifdef BVH_VERBOSE_TREE
+	VERBOSE_PRINT("\nitem_add BEFORE");
+	_debug_recursive_print_tree(0);
+	VERBOSE_PRINT("\n");
+#endif
+
+	BVHABB_CLASS abb;
+	abb.from(p_aabb);
+
+	// handle to be filled with the new item ref
+	BVHHandle handle;
+
+	// ref id easier to pass around than handle
+	uint32_t ref_id;
+
+	// this should never fail
+	ItemRef *ref = _refs.request(ref_id);
+
+	// the extra data should be parallel list to the references
+	uint32_t extra_id;
+	ItemExtra *extra = _extra.request(extra_id);
+	BVH_ASSERT(extra_id == ref_id);
+
+	// pairs info
+	if (USE_PAIRS) {
+		uint32_t pairs_id;
+		ItemPairs *pairs = _pairs.request(pairs_id);
+		pairs->clear();
+		BVH_ASSERT(pairs_id == ref_id);
+	}
+
+	extra->subindex = p_subindex;
+	extra->userdata = p_userdata;
+	extra->last_updated_tick = 0;
+
+	// add an active reference to the list for slow incremental optimize
+	// this list must be kept in sync with the references as they are added or removed.
+	extra->active_ref_id = _active_refs.size();
+	_active_refs.push_back(ref_id);
+
+	if (USE_PAIRS) {
+		extra->pairable_mask = p_pairable_mask;
+		extra->pairable_type = p_pairable_type;
+		extra->pairable = p_pairable;
+	} else {
+		// just for safety, in case this gets queried etc
+		extra->pairable = 0;
+		p_pairable = false;
+	}
+
+	// assign to handle to return
+	handle.set_id(ref_id);
+
+	_current_tree = 0;
+	if (p_pairable) {
+		_current_tree = 1;
+	}
+
+	create_root_node(_current_tree);
+
+	// we must choose where to add to tree
+	if (p_active) {
+		ref->tnode_id = _logic_choose_item_add_node(_root_node_id[_current_tree], abb);
+
+		bool refit = _node_add_item(ref->tnode_id, ref_id, abb);
+
+		if (refit) {
+			// only need to refit from the parent
+			const TNode &add_node = _nodes[ref->tnode_id];
+			if (add_node.parent_id != BVHCommon::INVALID) {
+				refit_upward_and_balance(add_node.parent_id);
+			}
+		}
+	} else {
+		ref->set_inactive();
+	}
+
+#ifdef BVH_VERBOSE
+	// memory use
+	int mem = _refs.estimate_memory_use();
+	mem += _nodes.estimate_memory_use();
+
+	String sz = _debug_aabb_to_string(abb);
+	VERBOSE_PRINT("\titem_add [" + itos(ref_id) + "] " + itos(_refs.size()) + " refs,\t" + itos(_nodes.size()) + " nodes " + sz);
+	VERBOSE_PRINT("mem use : " + itos(mem) + ", num nodes : " + itos(_nodes.size()));
+
+#endif
+
+	return handle;
+}
+
+void _debug_print_refs() {
+#ifdef BVH_VERBOSE_TREE
+	print_line("refs.....");
+	for (int n = 0; n < _refs.size(); n++) {
+		const ItemRef &ref = _refs[n];
+		print_line("tnode_id " + itos(ref.tnode_id) + ", item_id " + itos(ref.item_id));
+	}
+
+#endif
+}
+
+// returns false if noop
+bool item_move(BVHHandle p_handle, const Bounds &p_aabb) {
+	uint32_t ref_id = p_handle.id();
+
+	// get the reference
+	ItemRef &ref = _refs[ref_id];
+	if (!ref.is_active()) {
+		return false;
+	}
+
+	BVHABB_CLASS abb;
+	abb.from(p_aabb);
+
+	BVH_ASSERT(ref.tnode_id != BVHCommon::INVALID);
+	TNode &tnode = _nodes[ref.tnode_id];
+
+	// does it fit within the current aabb?
+	if (tnode.aabb.is_other_within(abb)) {
+		// do nothing .. fast path .. not moved enough to need refit
+
+		// however we WILL update the exact aabb in the leaf, as this will be needed
+		// for accurate collision detection
+		TLeaf &leaf = _node_get_leaf(tnode);
+
+		BVHABB_CLASS &leaf_abb = leaf.get_aabb(ref.item_id);
+
+		// no change?
+		if (leaf_abb == abb) {
+			return false;
+		}
+
+		leaf_abb = abb;
+		_integrity_check_all();
+
+		return true;
+	}
+
+	_current_tree = _handle_get_tree_id(p_handle);
+
+	// remove and reinsert
+	node_remove_item(ref_id);
+
+	// we must choose where to add to tree
+	ref.tnode_id = _logic_choose_item_add_node(_root_node_id[_current_tree], abb);
+
+	// add to the tree
+	bool needs_refit = _node_add_item(ref.tnode_id, ref_id, abb);
+
+	// only need to refit from the PARENT
+	if (needs_refit) {
+		// only need to refit from the parent
+		const TNode &add_node = _nodes[ref.tnode_id];
+		if (add_node.parent_id != BVHCommon::INVALID) {
+			// not sure we need to rebalance all the time, this can be done less often
+			refit_upward(add_node.parent_id);
+		}
+		//refit_upward_and_balance(add_node.parent_id);
+	}
+
+	return true;
+}
+
+void item_remove(BVHHandle p_handle) {
+	uint32_t ref_id = p_handle.id();
+
+	_current_tree = _handle_get_tree_id(p_handle);
+
+	VERBOSE_PRINT("item_remove [" + itos(ref_id) + "] ");
+
+	////////////////////////////////////////
+	// remove the active reference from the list for slow incremental optimize
+	// this list must be kept in sync with the references as they are added or removed.
+	uint32_t active_ref_id = _extra[ref_id].active_ref_id;
+	uint32_t ref_id_moved_back = _active_refs[_active_refs.size() - 1];
+
+	// swap back and decrement for fast unordered remove
+	_active_refs[active_ref_id] = ref_id_moved_back;
+	_active_refs.resize(_active_refs.size() - 1);
+
+	// keep the moved active reference up to date
+	_extra[ref_id_moved_back].active_ref_id = active_ref_id;
+	////////////////////////////////////////
+
+	// remove the item from the node (only if active)
+	if (_refs[ref_id].is_active()) {
+		node_remove_item(ref_id);
+	}
+
+	// remove the item reference
+	_refs.free(ref_id);
+	_extra.free(ref_id);
+	if (USE_PAIRS) {
+		_pairs.free(ref_id);
+	}
+
+	// don't think refit_all is necessary?
+	//refit_all(_current_tree);
+
+#ifdef BVH_VERBOSE_TREE
+	_debug_recursive_print_tree(_current_tree);
+#endif
+}
+
+// returns success
+bool item_activate(BVHHandle p_handle, const Bounds &p_aabb) {
+	uint32_t ref_id = p_handle.id();
+	ItemRef &ref = _refs[ref_id];
+	if (ref.is_active()) {
+		// noop
+		return false;
+	}
+
+	// add to tree
+	BVHABB_CLASS abb;
+	abb.from(p_aabb);
+
+	_current_tree = _handle_get_tree_id(p_handle);
+
+	// we must choose where to add to tree
+	ref.tnode_id = _logic_choose_item_add_node(_root_node_id[_current_tree], abb);
+	_node_add_item(ref.tnode_id, ref_id, abb);
+
+	refit_upward_and_balance(ref.tnode_id);
+
+	return true;
+}
+
+// returns success
+bool item_deactivate(BVHHandle p_handle) {
+	uint32_t ref_id = p_handle.id();
+	ItemRef &ref = _refs[ref_id];
+	if (!ref.is_active()) {
+		// noop
+		return false;
+	}
+
+	// remove from tree
+	BVHABB_CLASS abb;
+	node_remove_item(ref_id, &abb);
+
+	// mark as inactive
+	ref.set_inactive();
+	return true;
+}
+
+bool item_get_active(BVHHandle p_handle) const {
+	uint32_t ref_id = p_handle.id();
+	const ItemRef &ref = _refs[ref_id];
+	return ref.is_active();
+}
+
+// during collision testing, we want to set the mask and whether pairable for the item testing from
+void item_fill_cullparams(BVHHandle p_handle, CullParams &r_params) const {
+	uint32_t ref_id = p_handle.id();
+	const ItemExtra &extra = _extra[ref_id];
+
+	// testing from a non pairable item, we only want to test pairable items
+	r_params.test_pairable_only = extra.pairable == 0;
+
+	// we take into account the mask of the item testing from
+	r_params.mask = extra.pairable_mask;
+	r_params.pairable_type = extra.pairable_type;
+}
+
+bool item_is_pairable(const BVHHandle &p_handle) {
+	uint32_t ref_id = p_handle.id();
+	const ItemExtra &extra = _extra[ref_id];
+	return extra.pairable != 0;
+}
+
+void item_get_ABB(const BVHHandle &p_handle, BVHABB_CLASS &r_abb) {
+	// change tree?
+	uint32_t ref_id = p_handle.id();
+	const ItemRef &ref = _refs[ref_id];
+
+	TNode &tnode = _nodes[ref.tnode_id];
+	TLeaf &leaf = _node_get_leaf(tnode);
+
+	r_abb = leaf.get_aabb(ref.item_id);
+}
+
+bool item_set_pairable(const BVHHandle &p_handle, bool p_pairable, uint32_t p_pairable_type, uint32_t p_pairable_mask) {
+	// change tree?
+	uint32_t ref_id = p_handle.id();
+
+	ItemExtra &ex = _extra[ref_id];
+	ItemRef &ref = _refs[ref_id];
+
+	bool active = ref.is_active();
+	bool pairable_changed = (ex.pairable != 0) != p_pairable;
+	bool state_changed = pairable_changed || (ex.pairable_type != p_pairable_type) || (ex.pairable_mask != p_pairable_mask);
+
+	ex.pairable_type = p_pairable_type;
+	ex.pairable_mask = p_pairable_mask;
+
+	if (active && pairable_changed) {
+		// record abb
+		TNode &tnode = _nodes[ref.tnode_id];
+		TLeaf &leaf = _node_get_leaf(tnode);
+		BVHABB_CLASS abb = leaf.get_aabb(ref.item_id);
+
+		// make sure current tree is correct prior to changing
+		_current_tree = _handle_get_tree_id(p_handle);
+
+		// remove from old tree
+		node_remove_item(ref_id);
+
+		// we must set the pairable AFTER getting the current tree
+		// because the pairable status determines which tree
+		ex.pairable = p_pairable;
+
+		// add to new tree
+		_current_tree = _handle_get_tree_id(p_handle);
+		create_root_node(_current_tree);
+
+		// we must choose where to add to tree
+		ref.tnode_id = _logic_choose_item_add_node(_root_node_id[_current_tree], abb);
+		bool needs_refit = _node_add_item(ref.tnode_id, ref_id, abb);
+
+		// only need to refit from the PARENT
+		if (needs_refit) {
+			// only need to refit from the parent
+			const TNode &add_node = _nodes[ref.tnode_id];
+			if (add_node.parent_id != BVHCommon::INVALID) {
+				refit_upward_and_balance(add_node.parent_id);
+			}
+		}
+	} else {
+		// always keep this up to date
+		ex.pairable = p_pairable;
+	}
+
+	return state_changed;
+}
+
+void incremental_optimize() {
+	// first update all aabbs as one off step..
+	// this is cheaper than doing it on each move as each leaf may get touched multiple times
+	// in a frame.
+	for (int n = 0; n < NUM_TREES; n++) {
+		if (_root_node_id[n] != BVHCommon::INVALID) {
+			refit_branch(_root_node_id[n]);
+		}
+	}
+
+	// now do small section reinserting to get things moving
+	// gradually, and keep items in the right leaf
+	if (_current_active_ref >= _active_refs.size()) {
+		_current_active_ref = 0;
+	}
+
+	// special case
+	if (!_active_refs.size()) {
+		return;
+	}
+
+	uint32_t ref_id = _active_refs[_current_active_ref++];
+
+	_logic_item_remove_and_reinsert(ref_id);
+
+#ifdef BVH_VERBOSE
+	/*
+	// memory use
+	int mem_refs = _refs.estimate_memory_use();
+	int mem_nodes = _nodes.estimate_memory_use();
+	int mem_leaves = _leaves.estimate_memory_use();
+
+	String sz;
+	sz += "mem_refs : " + itos(mem_refs) + " ";
+	sz += "mem_nodes : " + itos(mem_nodes) + " ";
+	sz += "mem_leaves : " + itos(mem_leaves) + " ";
+	sz += ", num nodes : " + itos(_nodes.size());
+	print_line(sz);
+	*/
+#endif
+}
+
+void update() {
+	incremental_optimize();
+
+	// keep the expansion values up to date with the world bound
+//#define BVH_ALLOW_AUTO_EXPANSION
+#ifdef BVH_ALLOW_AUTO_EXPANSION
+	if (_auto_node_expansion || _auto_pairing_expansion) {
+		BVHABB_CLASS world_bound;
+		world_bound.set_to_max_opposite_extents();
+
+		bool bound_valid = false;
+
+		for (int n = 0; n < NUM_TREES; n++) {
+			uint32_t node_id = _root_node_id[n];
+			if (node_id != BVHCommon::INVALID) {
+				world_bound.merge(_nodes[node_id].aabb);
+				bound_valid = true;
+			}
+		}
+
+		// if there are no nodes, do nothing, but if there are...
+		if (bound_valid) {
+			Bounds bb;
+			world_bound.to(bb);
+			real_t size = bb.get_longest_axis_size();
+
+			// automatic AI decision for best parameters.
+			// These can be overridden in project settings.
+
+			// these magic numbers are determined by experiment
+			if (_auto_node_expansion) {
+				_node_expansion = size * 0.025;
+			}
+			if (_auto_pairing_expansion) {
+				_pairing_expansion = size * 0.009;
+			}
+		}
+	}
+#endif
+}
diff --git a/core/math/bvh_refit.inc b/core/math/bvh_refit.inc
new file mode 100644
index 0000000000..514c853ac5
--- /dev/null
+++ b/core/math/bvh_refit.inc
@@ -0,0 +1,141 @@
+void _debug_node_verify_bound(uint32_t p_node_id) {
+	TNode &node = _nodes[p_node_id];
+	BVHABB_CLASS abb_before = node.aabb;
+
+	node_update_aabb(node);
+
+	BVHABB_CLASS abb_after = node.aabb;
+	CRASH_COND(abb_before != abb_after);
+}
+
+void node_update_aabb(TNode &tnode) {
+	tnode.aabb.set_to_max_opposite_extents();
+	tnode.height = 0;
+
+	if (!tnode.is_leaf()) {
+		for (int n = 0; n < tnode.num_children; n++) {
+			uint32_t child_node_id = tnode.children[n];
+
+			// merge with child aabb
+			const TNode &tchild = _nodes[child_node_id];
+			tnode.aabb.merge(tchild.aabb);
+
+			// do heights at the same time
+			if (tchild.height > tnode.height) {
+				tnode.height = tchild.height;
+			}
+		}
+
+		// the height of a non leaf is always 1 bigger than the biggest child
+		tnode.height++;
+
+#ifdef BVH_CHECKS
+		if (!tnode.num_children) {
+			// the 'blank' aabb will screw up parent aabbs
+			WARN_PRINT("BVH_Tree::TNode no children, AABB is undefined");
+		}
+#endif
+	} else {
+		// leaf
+		const TLeaf &leaf = _node_get_leaf(tnode);
+
+		for (int n = 0; n < leaf.num_items; n++) {
+			tnode.aabb.merge(leaf.get_aabb(n));
+		}
+
+		// now the leaf items are unexpanded, we expand only in the node AABB
+		tnode.aabb.expand(_node_expansion);
+#ifdef BVH_CHECKS
+		if (!leaf.num_items) {
+			// the 'blank' aabb will screw up parent aabbs
+			WARN_PRINT("BVH_Tree::TLeaf no items, AABB is undefined");
+		}
+#endif
+	}
+}
+
+void refit_all(int p_tree_id) {
+	refit_downward(_root_node_id[p_tree_id]);
+}
+
+void refit_upward(uint32_t p_node_id) {
+	while (p_node_id != BVHCommon::INVALID) {
+		TNode &tnode = _nodes[p_node_id];
+		node_update_aabb(tnode);
+		p_node_id = tnode.parent_id;
+	}
+}
+
+void refit_upward_and_balance(uint32_t p_node_id) {
+	while (p_node_id != BVHCommon::INVALID) {
+		uint32_t before = p_node_id;
+		p_node_id = _logic_balance(p_node_id);
+
+		if (before != p_node_id) {
+			VERBOSE_PRINT("REBALANCED!");
+		}
+
+		TNode &tnode = _nodes[p_node_id];
+
+		// update overall aabb from the children
+		node_update_aabb(tnode);
+
+		p_node_id = tnode.parent_id;
+	}
+}
+
+void refit_downward(uint32_t p_node_id) {
+	TNode &tnode = _nodes[p_node_id];
+
+	// do children first
+	if (!tnode.is_leaf()) {
+		for (int n = 0; n < tnode.num_children; n++) {
+			refit_downward(tnode.children[n]);
+		}
+	}
+
+	node_update_aabb(tnode);
+}
+
+// go down to the leaves, then refit upward
+void refit_branch(uint32_t p_node_id) {
+	// our function parameters to keep on a stack
+	struct RefitParams {
+		uint32_t node_id;
+	};
+
+	// most of the iterative functionality is contained in this helper class
+	BVH_IterativeInfo<RefitParams> ii;
+
+	// alloca must allocate the stack from this function, it cannot be allocated in the
+	// helper class
+	ii.stack = (RefitParams *)alloca(ii.get_alloca_stacksize());
+
+	// seed the stack
+	ii.get_first()->node_id = p_node_id;
+
+	RefitParams rp;
+
+	// while there are still more nodes on the stack
+	while (ii.pop(rp)) {
+		TNode &tnode = _nodes[rp.node_id];
+
+		// do children first
+		if (!tnode.is_leaf()) {
+			for (int n = 0; n < tnode.num_children; n++) {
+				uint32_t child_id = tnode.children[n];
+
+				// add to the stack
+				RefitParams *child = ii.request();
+				child->node_id = child_id;
+			}
+		} else {
+			// leaf .. only refit upward if dirty
+			TLeaf &leaf = _node_get_leaf(tnode);
+			if (leaf.is_dirty()) {
+				leaf.set_dirty(false);
+				refit_upward(p_node_id);
+			}
+		}
+	} // while more nodes to pop
+}
diff --git a/core/math/bvh_split.inc b/core/math/bvh_split.inc
new file mode 100644
index 0000000000..3fcc4c7b10
--- /dev/null
+++ b/core/math/bvh_split.inc
@@ -0,0 +1,294 @@
+void _split_inform_references(uint32_t p_node_id) {
+	TNode &node = _nodes[p_node_id];
+	TLeaf &leaf = _node_get_leaf(node);
+
+	for (int n = 0; n < leaf.num_items; n++) {
+		uint32_t ref_id = leaf.get_item_ref_id(n);
+
+		ItemRef &ref = _refs[ref_id];
+		ref.tnode_id = p_node_id;
+		ref.item_id = n;
+	}
+}
+
+void _split_leaf_sort_groups_simple(int &num_a, int &num_b, uint16_t *group_a, uint16_t *group_b, const BVHABB_CLASS *temp_bounds, const BVHABB_CLASS full_bound) {
+	// special case for low leaf sizes .. should static compile out
+	if (MAX_ITEMS < 4) {
+		uint32_t ind = group_a[0];
+
+		// add to b
+		group_b[num_b++] = ind;
+
+		// remove from a
+		group_a[0] = group_a[num_a - 1];
+		num_a--;
+		return;
+	}
+
+	Point centre = full_bound.calculate_centre();
+	Point size = full_bound.calculate_size();
+
+	int order[3];
+
+	order[0] = size.min_axis();
+	order[2] = size.max_axis();
+	order[1] = 3 - (order[0] + order[2]);
+
+	// simplest case, split on the longest axis
+	int split_axis = order[0];
+	for (int a = 0; a < num_a; a++) {
+		uint32_t ind = group_a[a];
+
+		if (temp_bounds[ind].min.coord[split_axis] > centre.coord[split_axis]) {
+			// add to b
+			group_b[num_b++] = ind;
+
+			// remove from a
+			group_a[a] = group_a[num_a - 1];
+			num_a--;
+
+			// do this one again, as it has been replaced
+			a--;
+		}
+	}
+
+	// detect when split on longest axis failed
+	int min_threshold = MAX_ITEMS / 4;
+	int min_group_size[3];
+	min_group_size[0] = MIN(num_a, num_b);
+	if (min_group_size[0] < min_threshold) {
+		// slow but sure .. first move everything back into a
+		for (int b = 0; b < num_b; b++) {
+			group_a[num_a++] = group_b[b];
+		}
+		num_b = 0;
+
+		// now calculate the best split
+		for (int axis = 1; axis < 3; axis++) {
+			split_axis = order[axis];
+			int count = 0;
+
+			for (int a = 0; a < num_a; a++) {
+				uint32_t ind = group_a[a];
+
+				if (temp_bounds[ind].min.coord[split_axis] > centre.coord[split_axis]) {
+					count++;
+				}
+			}
+
+			min_group_size[axis] = MIN(count, num_a - count);
+		} // for axis
+
+		// best axis
+		int best_axis = 0;
+		int best_min = min_group_size[0];
+		for (int axis = 1; axis < 3; axis++) {
+			if (min_group_size[axis] > best_min) {
+				best_min = min_group_size[axis];
+				best_axis = axis;
+			}
+		}
+
+		// now finally do the split
+		if (best_min > 0) {
+			split_axis = order[best_axis];
+
+			for (int a = 0; a < num_a; a++) {
+				uint32_t ind = group_a[a];
+
+				if (temp_bounds[ind].min.coord[split_axis] > centre.coord[split_axis]) {
+					// add to b
+					group_b[num_b++] = ind;
+
+					// remove from a
+					group_a[a] = group_a[num_a - 1];
+					num_a--;
+
+					// do this one again, as it has been replaced
+					a--;
+				}
+			}
+		} // if there was a split!
+	} // if the longest axis wasn't a good split
+
+	// special case, none crossed threshold
+	if (!num_b) {
+		uint32_t ind = group_a[0];
+
+		// add to b
+		group_b[num_b++] = ind;
+
+		// remove from a
+		group_a[0] = group_a[num_a - 1];
+		num_a--;
+	}
+	// opposite problem! :)
+	if (!num_a) {
+		uint32_t ind = group_b[0];
+
+		// add to a
+		group_a[num_a++] = ind;
+
+		// remove from b
+		group_b[0] = group_b[num_b - 1];
+		num_b--;
+	}
+}
+
+void _split_leaf_sort_groups(int &num_a, int &num_b, uint16_t *group_a, uint16_t *group_b, const BVHABB_CLASS *temp_bounds) {
+	BVHABB_CLASS groupb_aabb;
+	groupb_aabb.set_to_max_opposite_extents();
+	for (int n = 0; n < num_b; n++) {
+		int which = group_b[n];
+		groupb_aabb.merge(temp_bounds[which]);
+	}
+	BVHABB_CLASS groupb_aabb_new;
+
+	BVHABB_CLASS rest_aabb;
+
+	float best_size = FLT_MAX;
+	int best_candidate = -1;
+
+	// find most likely from a to move into b
+	for (int check = 0; check < num_a; check++) {
+		rest_aabb.set_to_max_opposite_extents();
+		groupb_aabb_new = groupb_aabb;
+
+		// find aabb of all the rest
+		for (int rest = 0; rest < num_a; rest++) {
+			if (rest == check) {
+				continue;
+			}
+
+			int which = group_a[rest];
+			rest_aabb.merge(temp_bounds[which]);
+		}
+
+		groupb_aabb_new.merge(temp_bounds[group_a[check]]);
+
+		// now compare the sizes
+		float size = groupb_aabb_new.get_area() + rest_aabb.get_area();
+		if (size < best_size) {
+			best_size = size;
+			best_candidate = check;
+		}
+	}
+
+	// we should now have the best, move it from group a to group b
+	group_b[num_b++] = group_a[best_candidate];
+
+	// remove best candidate from group a
+	num_a--;
+	group_a[best_candidate] = group_a[num_a];
+}
+
+uint32_t split_leaf(uint32_t p_node_id, const BVHABB_CLASS &p_added_item_aabb) {
+	return split_leaf_complex(p_node_id, p_added_item_aabb);
+}
+
+// aabb is the new inserted node
+uint32_t split_leaf_complex(uint32_t p_node_id, const BVHABB_CLASS &p_added_item_aabb) {
+	VERBOSE_PRINT("split_leaf");
+
+	// note the tnode before and AFTER splitting may be a different address
+	// in memory because the vector could get relocated. So we need to reget
+	// the tnode after the split
+	BVH_ASSERT(_nodes[p_node_id].is_leaf());
+
+	// first create child leaf nodes
+	uint32_t *child_ids = (uint32_t *)alloca(sizeof(uint32_t) * MAX_CHILDREN);
+
+	for (int n = 0; n < MAX_CHILDREN; n++) {
+		// create node children
+		TNode *child_node = _nodes.request(child_ids[n]);
+
+		child_node->clear();
+
+		// back link to parent
+		child_node->parent_id = p_node_id;
+
+		// make each child a leaf node
+		node_make_leaf(child_ids[n]);
+	}
+
+	// don't get any leaves or nodes till AFTER the split
+	TNode &tnode = _nodes[p_node_id];
+	uint32_t orig_leaf_id = tnode.get_leaf_id();
+	const TLeaf &orig_leaf = _node_get_leaf(tnode);
+
+	// store the final child ids
+	for (int n = 0; n < MAX_CHILDREN; n++) {
+		tnode.children[n] = child_ids[n];
+	}
+
+	// mark as no longer a leaf node
+	tnode.num_children = MAX_CHILDREN;
+
+	// 2 groups, A and B, and assign children to each to split equally
+	int max_children = orig_leaf.num_items + 1; // plus 1 for the wildcard .. the item being added
+	//CRASH_COND(max_children > MAX_CHILDREN);
+
+	uint16_t *group_a = (uint16_t *)alloca(sizeof(uint16_t) * max_children);
+	uint16_t *group_b = (uint16_t *)alloca(sizeof(uint16_t) * max_children);
+
+	// we are copying the ABBs. This is ugly, but we need one extra for the inserted item...
+	BVHABB_CLASS *temp_bounds = (BVHABB_CLASS *)alloca(sizeof(BVHABB_CLASS) * max_children);
+
+	int num_a = max_children;
+	int num_b = 0;
+
+	// setup - start with all in group a
+	for (int n = 0; n < orig_leaf.num_items; n++) {
+		group_a[n] = n;
+		temp_bounds[n] = orig_leaf.get_aabb(n);
+	}
+	// wildcard
+	int wildcard = orig_leaf.num_items;
+
+	group_a[wildcard] = wildcard;
+	temp_bounds[wildcard] = p_added_item_aabb;
+
+	// we can choose here either an equal split, or just 1 in the new leaf
+	_split_leaf_sort_groups_simple(num_a, num_b, group_a, group_b, temp_bounds, tnode.aabb);
+
+	uint32_t wildcard_node = BVHCommon::INVALID;
+
+	// now there should be equal numbers in both groups
+	for (int n = 0; n < num_a; n++) {
+		int which = group_a[n];
+
+		if (which != wildcard) {
+			const BVHABB_CLASS &source_item_aabb = orig_leaf.get_aabb(which);
+			uint32_t source_item_ref_id = orig_leaf.get_item_ref_id(which);
+			//const Item &source_item = orig_leaf.get_item(which);
+			_node_add_item(tnode.children[0], source_item_ref_id, source_item_aabb);
+		} else {
+			wildcard_node = tnode.children[0];
+		}
+	}
+	for (int n = 0; n < num_b; n++) {
+		int which = group_b[n];
+
+		if (which != wildcard) {
+			const BVHABB_CLASS &source_item_aabb = orig_leaf.get_aabb(which);
+			uint32_t source_item_ref_id = orig_leaf.get_item_ref_id(which);
+			//const Item &source_item = orig_leaf.get_item(which);
+			_node_add_item(tnode.children[1], source_item_ref_id, source_item_aabb);
+		} else {
+			wildcard_node = tnode.children[1];
+		}
+	}
+
+	// now remove all items from the parent and replace with the child nodes
+	_leaves.free(orig_leaf_id);
+
+	// we should keep the references up to date!
+	for (int n = 0; n < MAX_CHILDREN; n++) {
+		_split_inform_references(tnode.children[n]);
+	}
+
+	refit_upward(p_node_id);
+
+	BVH_ASSERT(wildcard_node != BVHCommon::INVALID);
+	return wildcard_node;
+}
diff --git a/core/math/bvh_structs.inc b/core/math/bvh_structs.inc
new file mode 100644
index 0000000000..4133ba6c10
--- /dev/null
+++ b/core/math/bvh_structs.inc
@@ -0,0 +1,181 @@
+
+public:
+struct ItemRef {
+	uint32_t tnode_id; // -1 is invalid
+	uint32_t item_id; // in the leaf
+
+	bool is_active() const { return tnode_id != BVHCommon::INACTIVE; }
+	void set_inactive() {
+		tnode_id = BVHCommon::INACTIVE;
+		item_id = BVHCommon::INACTIVE;
+	}
+};
+
+// extra info kept in separate parallel list to the references,
+// as this is less used as keeps cache better
+struct ItemExtra {
+	uint32_t last_updated_tick;
+	uint32_t pairable;
+	uint32_t pairable_mask;
+	uint32_t pairable_type;
+
+	int32_t subindex;
+
+	// the active reference is a separate list of which references
+	// are active so that we can slowly iterate through it over many frames for
+	// slow optimize.
+	uint32_t active_ref_id;
+
+	T *userdata;
+};
+
+// this is an item OR a child node depending on whether a leaf node
+struct Item {
+	BVHABB_CLASS aabb;
+	uint32_t item_ref_id;
+};
+
+// tree leaf
+struct TLeaf {
+	uint16_t num_items;
+
+private:
+	uint16_t dirty;
+	// separate data orientated lists for faster SIMD traversal
+	uint32_t item_ref_ids[MAX_ITEMS];
+	BVHABB_CLASS aabbs[MAX_ITEMS];
+
+public:
+	// accessors
+	BVHABB_CLASS &get_aabb(uint32_t p_id) { return aabbs[p_id]; }
+	const BVHABB_CLASS &get_aabb(uint32_t p_id) const { return aabbs[p_id]; }
+
+	uint32_t &get_item_ref_id(uint32_t p_id) { return item_ref_ids[p_id]; }
+	const uint32_t &get_item_ref_id(uint32_t p_id) const { return item_ref_ids[p_id]; }
+
+	bool is_dirty() const { return dirty; }
+	void set_dirty(bool p) { dirty = p; }
+
+	void clear() {
+		num_items = 0;
+		set_dirty(true);
+	}
+	bool is_full() const { return num_items >= MAX_ITEMS; }
+
+	void remove_item_unordered(uint32_t p_id) {
+		BVH_ASSERT(p_id < num_items);
+		num_items--;
+		aabbs[p_id] = aabbs[num_items];
+		item_ref_ids[p_id] = item_ref_ids[num_items];
+	}
+
+	uint32_t request_item() {
+		if (num_items < MAX_ITEMS) {
+			uint32_t id = num_items;
+			num_items++;
+			return id;
+		}
+		return -1;
+	}
+};
+
+// tree node
+struct TNode {
+	BVHABB_CLASS aabb;
+	// either number of children if positive
+	// or leaf id if negative (leaf id 0 is disallowed)
+	union {
+		int32_t num_children;
+		int32_t neg_leaf_id;
+	};
+	uint32_t parent_id; // or -1
+	uint16_t children[MAX_CHILDREN];
+
+	// height in the tree, where leaves are 0, and all above are 1+
+	// (or the highest where there is a tie off)
+	int32_t height;
+
+	bool is_leaf() const { return num_children < 0; }
+	void set_leaf_id(int id) { neg_leaf_id = -id; }
+	int get_leaf_id() const { return -neg_leaf_id; }
+
+	void clear() {
+		num_children = 0;
+		parent_id = BVHCommon::INVALID;
+		height = 0; // or -1 for testing
+
+		// for safety set to improbable value
+		aabb.set_to_max_opposite_extents();
+
+		// other members are not blanked for speed .. they may be uninitialized
+	}
+
+	bool is_full_of_children() const { return num_children >= MAX_CHILDREN; }
+
+	void remove_child_internal(uint32_t child_num) {
+		children[child_num] = children[num_children - 1];
+		num_children--;
+	}
+
+	int find_child(uint32_t p_child_node_id) {
+		BVH_ASSERT(!is_leaf());
+
+		for (int n = 0; n < num_children; n++) {
+			if (children[n] == p_child_node_id) {
+				return n;
+			}
+		}
+
+		// not found
+		return -1;
+	}
+};
+
+// instead of using linked list we maintain
+// item references (for quick lookup)
+PooledList<ItemRef, true> _refs;
+PooledList<ItemExtra, true> _extra;
+PooledList<ItemPairs> _pairs;
+
+// these 2 are not in sync .. nodes != leaves!
+PooledList<TNode, true> _nodes;
+PooledList<TLeaf, true> _leaves;
+
+// we can maintain an un-ordered list of which references are active,
+// in order to do a slow incremental optimize of the tree over each frame.
+// This will work best if dynamic objects and static objects are in a different tree.
+LocalVector<uint32_t, uint32_t, true> _active_refs;
+uint32_t _current_active_ref = 0;
+
+// instead of translating directly to the userdata output,
+// we keep an intermediate list of hits as reference IDs, which can be used
+// for pairing collision detection
+LocalVector<uint32_t, uint32_t, true> _cull_hits;
+
+// we now have multiple root nodes, allowing us to store
+// more than 1 tree. This can be more efficient, while sharing the same
+// common lists
+enum { NUM_TREES = 2,
+};
+
+// Tree 0 - Non pairable
+// Tree 1 - Pairable
+// This is more efficient because in physics we only need check non pairable against the pairable tree.
+uint32_t _root_node_id[NUM_TREES];
+int _current_tree = 0;
+
+// these values may need tweaking according to the project
+// the bound of the world, and the average velocities of the objects
+
+// node expansion is important in the rendering tree
+// larger values give less re-insertion as items move...
+// but on the other hand over estimates the bounding box of nodes.
+// we can either use auto mode, where the expansion is based on the root node size, or specify manually
+real_t _node_expansion = 0.5;
+bool _auto_node_expansion = true;
+
+// pairing expansion important for physics pairing
+// larger values gives more 'sticky' pairing, and is less likely to exhibit tunneling
+// we can either use auto mode, where the expansion is based on the root node size, or specify manually
+real_t _pairing_expansion = 0.1;
+bool _auto_pairing_expansion = true;
diff --git a/core/math/bvh_tree.h b/core/math/bvh_tree.h
new file mode 100644
index 0000000000..64c5f6e254
--- /dev/null
+++ b/core/math/bvh_tree.h
@@ -0,0 +1,422 @@
+/*************************************************************************/
+/*  bvh_tree.h                                                           */
+/*************************************************************************/
+/*                       This file is part of:                           */
+/*                           GODOT ENGINE                                */
+/*                      https://godotengine.org                          */
+/*************************************************************************/
+/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
+/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
+/*                                                                       */
+/* Permission is hereby granted, free of charge, to any person obtaining */
+/* a copy of this software and associated documentation files (the       */
+/* "Software"), to deal in the Software without restriction, including   */
+/* without limitation the rights to use, copy, modify, merge, publish,   */
+/* distribute, sublicense, and/or sell copies of the Software, and to    */
+/* permit persons to whom the Software is furnished to do so, subject to */
+/* the following conditions:                                             */
+/*                                                                       */
+/* The above copyright notice and this permission notice shall be        */
+/* included in all copies or substantial portions of the Software.       */
+/*                                                                       */
+/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
+/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
+/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
+/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
+/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
+/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
+/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
+/*************************************************************************/
+
+#ifndef BVH_TREE_H
+#define BVH_TREE_H
+
+// BVH Tree
+// This is an implementation of a dynamic BVH with templated leaf size.
+// This differs from most dynamic BVH in that it can handle more than 1 object
+// in leaf nodes. This can make it far more efficient in certain circumstances.
+// It also means that the splitting logic etc have to be completely different
+// to a simpler tree.
+// Note that MAX_CHILDREN should be fixed at 2 for now.
+
+#include "core/math/aabb.h"
+#include "core/math/bvh_abb.h"
+#include "core/math/geometry_3d.h"
+#include "core/math/vector3.h"
+#include "core/string/print_string.h"
+#include "core/templates/local_vector.h"
+#include "core/templates/pooled_list.h"
+#include <limits.h>
+
+#define BVHABB_CLASS BVH_ABB<Bounds, Point>
+
+// never do these checks in release
+#if defined(TOOLS_ENABLED) && defined(DEBUG_ENABLED)
+//#define BVH_VERBOSE
+//#define BVH_VERBOSE_TREE
+
+//#define BVH_VERBOSE_FRAME
+//#define BVH_CHECKS
+//#define BVH_INTEGRITY_CHECKS
+#endif
+
+// debug only assert
+#ifdef BVH_CHECKS
+#define BVH_ASSERT(a) CRASH_COND((a) == false)
+#else
+#define BVH_ASSERT(a)
+#endif
+
+#ifdef BVH_VERBOSE
+#define VERBOSE_PRINT print_line
+#else
+#define VERBOSE_PRINT(a)
+#endif
+
+// really just a namespace
+struct BVHCommon {
+	// these could possibly also be the same constant,
+	// although this may be useful for debugging.
+	// or use zero for invalid and +1 based indices.
+	static const uint32_t INVALID = (0xffffffff);
+	static const uint32_t INACTIVE = (0xfffffffe);
+};
+
+// really a handle, can be anything
+// note that zero is a valid reference for the BVH .. this may involve using
+// a plus one based ID for clients that expect 0 to be invalid.
+struct BVHHandle {
+	// conversion operator
+	operator uint32_t() const { return _data; }
+	void set(uint32_t p_value) { _data = p_value; }
+
+	uint32_t _data;
+
+	void set_invalid() { _data = BVHCommon::INVALID; }
+	bool is_invalid() const { return _data == BVHCommon::INVALID; }
+	uint32_t id() const { return _data; }
+	void set_id(uint32_t p_id) { _data = p_id; }
+
+	bool operator==(const BVHHandle &p_h) const { return _data == p_h._data; }
+	bool operator!=(const BVHHandle &p_h) const { return (*this == p_h) == false; }
+};
+
+// helper class to make iterative versions of recursive functions
+template <class T>
+class BVH_IterativeInfo {
+public:
+	enum {
+		ALLOCA_STACK_SIZE = 128
+	};
+
+	int32_t depth = 1;
+	int32_t threshold = ALLOCA_STACK_SIZE - 2;
+	T *stack;
+	//only used in rare occasions when you run out of alloca memory
+	// because tree is too unbalanced.
+	LocalVector<T> aux_stack;
+	int32_t get_alloca_stacksize() const { return ALLOCA_STACK_SIZE * sizeof(T); }
+
+	T *get_first() const {
+		return &stack[0];
+	}
+
+	// pop the last member of the stack, or return false
+	bool pop(T &r_value) {
+		if (!depth) {
+			return false;
+		}
+
+		depth--;
+		r_value = stack[depth];
+		return true;
+	}
+
+	// request new addition to stack
+	T *request() {
+		if (depth > threshold) {
+			if (aux_stack.is_empty()) {
+				aux_stack.resize(ALLOCA_STACK_SIZE * 2);
+				memcpy(aux_stack.ptr(), stack, get_alloca_stacksize());
+			} else {
+				aux_stack.resize(aux_stack.size() * 2);
+			}
+			stack = aux_stack.ptr();
+			threshold = aux_stack.size() - 2;
+		}
+		return &stack[depth++];
+	}
+};
+
+template <class T, int MAX_CHILDREN, int MAX_ITEMS, bool USE_PAIRS = false, class Bounds = AABB, class Point = Vector3>
+class BVH_Tree {
+	friend class BVH;
+
+#include "bvh_pair.inc"
+#include "bvh_structs.inc"
+
+public:
+	BVH_Tree() {
+		for (int n = 0; n < NUM_TREES; n++) {
+			_root_node_id[n] = BVHCommon::INVALID;
+		}
+
+		// disallow zero leaf ids
+		// (as these ids are stored as negative numbers in the node)
+		uint32_t dummy_leaf_id;
+		_leaves.request(dummy_leaf_id);
+	}
+
+private:
+	bool node_add_child(uint32_t p_node_id, uint32_t p_child_node_id) {
+		TNode &tnode = _nodes[p_node_id];
+		if (tnode.is_full_of_children()) {
+			return false;
+		}
+
+		tnode.children[tnode.num_children] = p_child_node_id;
+		tnode.num_children += 1;
+
+		// back link in the child to the parent
+		TNode &tnode_child = _nodes[p_child_node_id];
+		tnode_child.parent_id = p_node_id;
+
+		return true;
+	}
+
+	void node_replace_child(uint32_t p_parent_id, uint32_t p_old_child_id, uint32_t p_new_child_id) {
+		TNode &parent = _nodes[p_parent_id];
+		BVH_ASSERT(!parent.is_leaf());
+
+		int child_num = parent.find_child(p_old_child_id);
+		BVH_ASSERT(child_num != BVHCommon::INVALID);
+		parent.children[child_num] = p_new_child_id;
+
+		TNode &new_child = _nodes[p_new_child_id];
+		new_child.parent_id = p_parent_id;
+	}
+
+	void node_remove_child(uint32_t p_parent_id, uint32_t p_child_id, bool p_prevent_sibling = false) {
+		TNode &parent = _nodes[p_parent_id];
+		BVH_ASSERT(!parent.is_leaf());
+
+		int child_num = parent.find_child(p_child_id);
+		BVH_ASSERT(child_num != BVHCommon::INVALID);
+
+		parent.remove_child_internal(child_num);
+
+		// no need to keep back references for children at the moment
+
+		uint32_t sibling_id; // always a node id, as tnode is never a leaf
+		bool sibling_present = false;
+
+		// if there are more children, or this is the root node, don't try and delete
+		if (parent.num_children > 1) {
+			return;
+		}
+
+		// if there is 1 sibling, it can be moved to be a child of the
+		if (parent.num_children == 1) {
+			// else there is now a redundant node with one child, which can be removed
+			sibling_id = parent.children[0];
+			sibling_present = true;
+		}
+
+		// now there may be no children in this node .. in which case it can be deleted
+		// remove node if empty
+		// remove link from parent
+		uint32_t grandparent_id = parent.parent_id;
+
+		// special case for root node
+		if (grandparent_id == BVHCommon::INVALID) {
+			if (sibling_present) {
+				// change the root node
+				change_root_node(sibling_id);
+
+				// delete the old root node as no longer needed
+				_nodes.free(p_parent_id);
+			}
+
+			return;
+		}
+
+		if (sibling_present) {
+			node_replace_child(grandparent_id, p_parent_id, sibling_id);
+		} else {
+			node_remove_child(grandparent_id, p_parent_id, true);
+		}
+
+		// put the node on the free list to recycle
+		_nodes.free(p_parent_id);
+	}
+
+	// this relies on _current_tree being accurate
+	void change_root_node(uint32_t p_new_root_id) {
+		_root_node_id[_current_tree] = p_new_root_id;
+		TNode &root = _nodes[p_new_root_id];
+
+		// mark no parent
+		root.parent_id = BVHCommon::INVALID;
+	}
+
+	void node_make_leaf(uint32_t p_node_id) {
+		uint32_t child_leaf_id;
+		TLeaf *child_leaf = _leaves.request(child_leaf_id);
+		child_leaf->clear();
+
+		// zero is reserved at startup, to prevent this id being used
+		// (as they are stored as negative values in the node, and zero is already taken)
+		BVH_ASSERT(child_leaf_id != 0);
+
+		TNode &node = _nodes[p_node_id];
+		node.neg_leaf_id = -(int)child_leaf_id;
+	}
+
+	void node_remove_item(uint32_t p_ref_id, BVHABB_CLASS *r_old_aabb = nullptr) {
+		// get the reference
+		ItemRef &ref = _refs[p_ref_id];
+		uint32_t owner_node_id = ref.tnode_id;
+
+		// debug draw special
+		// This may not be needed
+		if (owner_node_id == BVHCommon::INVALID) {
+			return;
+		}
+
+		TNode &tnode = _nodes[owner_node_id];
+		CRASH_COND(!tnode.is_leaf());
+
+		TLeaf &leaf = _node_get_leaf(tnode);
+
+		// if the aabb is not determining the corner size, then there is no need to refit!
+		// (optimization, as merging AABBs takes a lot of time)
+		const BVHABB_CLASS &old_aabb = leaf.get_aabb(ref.item_id);
+
+		// shrink a little to prevent using corner aabbs
+		// in order to miss the corners first we shrink by node_expansion
+		// (which is added to the overall bound of the leaf), then we also
+		// shrink by an epsilon, in order to miss out the very corner aabbs
+		// which are important in determining the bound. Any other aabb
+		// within this can be removed and not affect the overall bound.
+		BVHABB_CLASS node_bound = tnode.aabb;
+		node_bound.expand(-_node_expansion - 0.001f);
+		bool refit = true;
+
+		if (node_bound.is_other_within(old_aabb)) {
+			refit = false;
+		}
+
+		// record the old aabb if required (for incremental remove_and_reinsert)
+		if (r_old_aabb) {
+			*r_old_aabb = old_aabb;
+		}
+
+		leaf.remove_item_unordered(ref.item_id);
+
+		if (leaf.num_items) {
+			// the swapped item has to have its reference changed to, to point to the new item id
+			uint32_t swapped_ref_id = leaf.get_item_ref_id(ref.item_id);
+
+			ItemRef &swapped_ref = _refs[swapped_ref_id];
+
+			swapped_ref.item_id = ref.item_id;
+
+			// only have to refit if it is an edge item
+			// This is a VERY EXPENSIVE STEP
+			// we defer the refit updates until the update function is called once per frame
+			if (refit) {
+				leaf.set_dirty(true);
+			}
+		} else {
+			// remove node if empty
+			// remove link from parent
+			if (tnode.parent_id != BVHCommon::INVALID) {
+				// DANGER .. this can potentially end up with root node with 1 child ...
+				// we don't want this and must check for it
+
+				uint32_t parent_id = tnode.parent_id;
+
+				node_remove_child(parent_id, owner_node_id);
+				refit_upward(parent_id);
+
+				// put the node on the free list to recycle
+				_nodes.free(owner_node_id);
+			}
+
+			// else if no parent, it is the root node. Do not delete
+		}
+
+		ref.tnode_id = BVHCommon::INVALID;
+		ref.item_id = BVHCommon::INVALID; // unset
+	}
+
+	// returns true if needs refit of PARENT tree only, the node itself AABB is calculated
+	// within this routine
+	bool _node_add_item(uint32_t p_node_id, uint32_t p_ref_id, const BVHABB_CLASS &p_aabb) {
+		ItemRef &ref = _refs[p_ref_id];
+		ref.tnode_id = p_node_id;
+
+		TNode &node = _nodes[p_node_id];
+		BVH_ASSERT(node.is_leaf());
+		TLeaf &leaf = _node_get_leaf(node);
+
+		// optimization - we only need to do a refit
+		// if the added item is changing the AABB of the node.
+		// in most cases it won't.
+		bool needs_refit = true;
+
+		// expand bound now
+		BVHABB_CLASS expanded = p_aabb;
+		expanded.expand(_node_expansion);
+
+		// the bound will only be valid if there is an item in there already
+		if (leaf.num_items) {
+			if (node.aabb.is_other_within(expanded)) {
+				// no change to node AABBs
+				needs_refit = false;
+			} else {
+				node.aabb.merge(expanded);
+			}
+		} else {
+			// bound of the node = the new aabb
+			node.aabb = expanded;
+		}
+
+		ref.item_id = leaf.request_item();
+		BVH_ASSERT(ref.item_id != BVHCommon::INVALID);
+
+		// set the aabb of the new item
+		leaf.get_aabb(ref.item_id) = p_aabb;
+
+		// back reference on the item back to the item reference
+		leaf.get_item_ref_id(ref.item_id) = p_ref_id;
+
+		return needs_refit;
+	}
+
+	uint32_t _node_create_another_child(uint32_t p_node_id, const BVHABB_CLASS &p_aabb) {
+		uint32_t child_node_id;
+		TNode *child_node = _nodes.request(child_node_id);
+		child_node->clear();
+
+		// may not be necessary
+		child_node->aabb = p_aabb;
+
+		node_add_child(p_node_id, child_node_id);
+
+		return child_node_id;
+	}
+
+#include "bvh_cull.inc"
+#include "bvh_debug.inc"
+#include "bvh_integrity.inc"
+#include "bvh_logic.inc"
+#include "bvh_misc.inc"
+#include "bvh_public.inc"
+#include "bvh_refit.inc"
+#include "bvh_split.inc"
+};
+
+#undef VERBOSE_PRINT
+
+#endif // BVH_TREE_H
diff --git a/core/math/rect2.h b/core/math/rect2.h
index 512499bdb2..1dc027cf72 100644
--- a/core/math/rect2.h
+++ b/core/math/rect2.h
@@ -182,13 +182,17 @@ struct Rect2 {
 
 	inline Rect2 grow(real_t p_amount) const {
 		Rect2 g = *this;
-		g.position.x -= p_amount;
-		g.position.y -= p_amount;
-		g.size.width += p_amount * 2;
-		g.size.height += p_amount * 2;
+		g.grow_by(p_amount);
 		return g;
 	}
 
+	inline void grow_by(real_t p_amount) {
+		position.x -= p_amount;
+		position.y -= p_amount;
+		size.width += p_amount * 2;
+		size.height += p_amount * 2;
+	}
+
 	inline Rect2 grow_side(Side p_side, real_t p_amount) const {
 		Rect2 g = *this;
 		g = g.grow_individual((SIDE_LEFT == p_side) ? p_amount : 0,
diff --git a/core/math/vector2.h b/core/math/vector2.h
index edc6e3a3ef..6abe0f5ea9 100644
--- a/core/math/vector2.h
+++ b/core/math/vector2.h
@@ -37,18 +37,26 @@
 struct Vector2i;
 
 struct Vector2 {
+	static const int AXIS_COUNT = 2;
+
 	enum Axis {
 		AXIS_X,
 		AXIS_Y,
 	};
 
 	union {
-		real_t x = 0;
-		real_t width;
-	};
-	union {
-		real_t y = 0;
-		real_t height;
+		struct {
+			union {
+				real_t x;
+				real_t width;
+			};
+			union {
+				real_t y;
+				real_t height;
+			};
+		};
+
+		real_t coord[2] = { 0 };
 	};
 
 	_FORCE_INLINE_ real_t &operator[](int p_idx) {
@@ -58,6 +66,18 @@ struct Vector2 {
 		return p_idx ? y : x;
 	}
 
+	_FORCE_INLINE_ void set_all(real_t p_value) {
+		x = y = p_value;
+	}
+
+	_FORCE_INLINE_ int min_axis() const {
+		return x < y ? 0 : 1;
+	}
+
+	_FORCE_INLINE_ int max_axis() const {
+		return x < y ? 1 : 0;
+	}
+
 	void normalize();
 	Vector2 normalized() const;
 	bool is_normalized() const;
diff --git a/core/math/vector3.cpp b/core/math/vector3.cpp
index f0629d3db8..d4317d506c 100644
--- a/core/math/vector3.cpp
+++ b/core/math/vector3.cpp
@@ -52,14 +52,6 @@ real_t Vector3::get_axis(int p_axis) const {
 	return operator[](p_axis);
 }
 
-int Vector3::min_axis() const {
-	return x < y ? (x < z ? 0 : 2) : (y < z ? 1 : 2);
-}
-
-int Vector3::max_axis() const {
-	return x < y ? (y < z ? 2 : 1) : (x < z ? 2 : 0);
-}
-
 void Vector3::snap(Vector3 p_step) {
 	x = Math::snapped(x, p_step.x);
 	y = Math::snapped(y, p_step.y);
diff --git a/core/math/vector3.h b/core/math/vector3.h
index 377581bb45..b47c3cc916 100644
--- a/core/math/vector3.h
+++ b/core/math/vector3.h
@@ -38,6 +38,8 @@
 class Basis;
 
 struct Vector3 {
+	static const int AXIS_COUNT = 3;
+
 	enum Axis {
 		AXIS_X,
 		AXIS_Y,
@@ -65,8 +67,17 @@ struct Vector3 {
 	void set_axis(int p_axis, real_t p_value);
 	real_t get_axis(int p_axis) const;
 
-	int min_axis() const;
-	int max_axis() const;
+	_FORCE_INLINE_ void set_all(real_t p_value) {
+		x = y = z = p_value;
+	}
+
+	_FORCE_INLINE_ int min_axis() const {
+		return x < y ? (x < z ? 0 : 2) : (y < z ? 1 : 2);
+	}
+
+	_FORCE_INLINE_ int max_axis() const {
+		return x < y ? (y < z ? 2 : 1) : (x < z ? 2 : 0);
+	}
 
 	_FORCE_INLINE_ real_t length() const;
 	_FORCE_INLINE_ real_t length_squared() const;
diff --git a/servers/physics_3d/broad_phase_3d_basic.h b/core/templates/pooled_list.h
index 54d34e005f..b4a6d2d1dd 100644
--- a/servers/physics_3d/broad_phase_3d_basic.h
+++ b/core/templates/pooled_list.h
@@ -1,5 +1,5 @@
 /*************************************************************************/
-/*  broad_phase_3d_basic.h                                               */
+/*  pooled_list.h                                                        */
 /*************************************************************************/
 /*                       This file is part of:                           */
 /*                           GODOT ENGINE                                */
@@ -28,78 +28,68 @@
 /* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
 /*************************************************************************/
 
-#ifndef BROAD_PHASE_BASIC_H
-#define BROAD_PHASE_BASIC_H
+#pragma once
 
-#include "broad_phase_3d_sw.h"
-#include "core/templates/map.h"
+// Simple template to provide a pool with O(1) allocate and free.
+// The freelist could alternatively be a linked list placed within the unused elements
+// to use less memory, however a separate freelist is probably more cache friendly.
 
-class BroadPhase3DBasic : public BroadPhase3DSW {
-	struct Element {
-		CollisionObject3DSW *owner;
-		bool _static;
-		AABB aabb;
-		int subindex;
-	};
+// NOTE : Take great care when using this with non POD types. The construction and destruction
+// is done in the LocalVector, NOT as part of the pool. So requesting a new item does not guarantee
+// a constructor is run, and free does not guarantee a destructor.
+// You should generally handle clearing
+// an item explicitly after a request, as it may contain 'leftovers'.
+// This is by design for fastest use in the BVH. If you want a more general pool
+// that does call constructors / destructors on request / free, this should probably be
+// a separate template.
 
-	Map<ID, Element> element_map;
+#include "core/templates/local_vector.h"
 
-	ID current;
+template <class T, bool force_trivial = false>
+class PooledList {
+	LocalVector<T, uint32_t, force_trivial> list;
+	LocalVector<uint32_t, uint32_t, true> freelist;
 
-	struct PairKey {
-		union {
-			struct {
-				ID a;
-				ID b;
-			};
-			uint64_t key;
-		};
-
-		_FORCE_INLINE_ bool operator<(const PairKey &p_key) const {
-			return key < p_key.key;
-		}
-
-		PairKey() { key = 0; }
-		PairKey(ID p_a, ID p_b) {
-			if (p_a > p_b) {
-				a = p_b;
-				b = p_a;
-			} else {
-				a = p_a;
-				b = p_b;
-			}
-		}
-	};
-
-	Map<PairKey, void *> pair_map;
-
-	PairCallback pair_callback;
-	void *pair_userdata;
-	UnpairCallback unpair_callback;
-	void *unpair_userdata;
+	// not all list members are necessarily used
+	int _used_size;
 
 public:
-	// 0 is an invalid ID
-	virtual ID create(CollisionObject3DSW *p_object, int p_subindex = 0);
-	virtual void move(ID p_id, const AABB &p_aabb);
-	virtual void set_static(ID p_id, bool p_static);
-	virtual void remove(ID p_id);
-
-	virtual CollisionObject3DSW *get_object(ID p_id) const;
-	virtual bool is_static(ID p_id) const;
-	virtual int get_subindex(ID p_id) const;
-
-	virtual int cull_point(const Vector3 &p_point, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices = nullptr);
-	virtual int cull_segment(const Vector3 &p_from, const Vector3 &p_to, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices = nullptr);
-	virtual int cull_aabb(const AABB &p_aabb, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices = nullptr);
-
-	virtual void set_pair_callback(PairCallback p_pair_callback, void *p_userdata);
-	virtual void set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata);
-
-	virtual void update();
+	PooledList() {
+		_used_size = 0;
+	}
+
+	int estimate_memory_use() const {
+		return (list.size() * sizeof(T)) + (freelist.size() * sizeof(uint32_t));
+	}
+
+	const T &operator[](uint32_t p_index) const {
+		return list[p_index];
+	}
+	T &operator[](uint32_t p_index) {
+		return list[p_index];
+	}
+
+	int size() const { return _used_size; }
+
+	T *request(uint32_t &r_id) {
+		_used_size++;
+
+		if (freelist.size()) {
+			// pop from freelist
+			int new_size = freelist.size() - 1;
+			r_id = freelist[new_size];
+			freelist.resize(new_size);
+			return &list[r_id];
+		}
 
-	static BroadPhase3DSW *_create();
-	BroadPhase3DBasic();
+		r_id = list.size();
+		list.resize(r_id + 1);
+		return &list[r_id];
+	}
+	void free(const uint32_t &p_id) {
+		// should not be on free list already
+		CRASH_COND(p_id >= list.size());
+		freelist.push_back(p_id);
+		_used_size--;
+	}
 };
-
-#endif // BROAD_PHASE_BASIC_H
diff --git a/doc/classes/ProjectSettings.xml b/doc/classes/ProjectSettings.xml
index 3200f789b8..d38c3fc0d8 100644
--- a/doc/classes/ProjectSettings.xml
+++ b/doc/classes/ProjectSettings.xml
@@ -1197,12 +1197,6 @@
 			The CA certificates bundle to use for SSL connections. If this is set to a non-empty value, this will [i]override[/i] Godot's default [url=https://github.com/godotengine/godot/blob/master/thirdparty/certs/ca-certificates.crt]Mozilla certificate bundle[/url]. If left empty, the default certificate bundle will be used.
 			If in doubt, leave this setting empty.
 		</member>
-		<member name="physics/2d/bp_hash_table_size" type="int" setter="" getter="" default="4096">
-			Size of the hash table used for the broad-phase 2D hash grid algorithm.
-		</member>
-		<member name="physics/2d/cell_size" type="int" setter="" getter="" default="128">
-			Cell size used for the broad-phase 2D hash grid algorithm (in pixels).
-		</member>
 		<member name="physics/2d/default_angular_damp" type="float" setter="" getter="" default="1.0">
 			The default angular damp in 2D.
 			[b]Note:[/b] Good values are in the range [code]0[/code] to [code]1[/code]. At value [code]0[/code] objects will keep moving with the same velocity. Values greater than [code]1[/code] will aim to reduce the velocity to [code]0[/code] in less than a second e.g. a value of [code]2[/code] will aim to reduce the velocity to [code]0[/code] in half a second. A value equal to or greater than the physics frame rate ([member ProjectSettings.physics/common/physics_fps], [code]60[/code] by default) will bring the object to a stop in one iteration.
@@ -1239,9 +1233,6 @@
 			The default linear damp in 2D.
 			[b]Note:[/b] Good values are in the range [code]0[/code] to [code]1[/code]. At value [code]0[/code] objects will keep moving with the same velocity. Values greater than [code]1[/code] will aim to reduce the velocity to [code]0[/code] in less than a second e.g. a value of [code]2[/code] will aim to reduce the velocity to [code]0[/code] in half a second. A value equal to or greater than the physics frame rate ([member ProjectSettings.physics/common/physics_fps], [code]60[/code] by default) will bring the object to a stop in one iteration.
 		</member>
-		<member name="physics/2d/large_object_surface_threshold_in_cells" type="int" setter="" getter="" default="512">
-			Threshold defining the surface size that constitutes a large object with regard to cells in the broad-phase 2D hash grid algorithm.
-		</member>
 		<member name="physics/2d/physics_engine" type="String" setter="" getter="" default="&quot;DEFAULT&quot;">
 			Sets which physics engine to use for 2D physics.
 			"DEFAULT" and "GodotPhysics2D" are the same, as there is currently no alternative 2D physics server implemented.
diff --git a/servers/physics_2d/broad_phase_2d_basic.cpp b/servers/physics_2d/broad_phase_2d_basic.cpp
deleted file mode 100644
index 17424629a9..0000000000
--- a/servers/physics_2d/broad_phase_2d_basic.cpp
+++ /dev/null
@@ -1,174 +0,0 @@
-/*************************************************************************/
-/*  broad_phase_2d_basic.cpp                                             */
-/*************************************************************************/
-/*                       This file is part of:                           */
-/*                           GODOT ENGINE                                */
-/*                      https://godotengine.org                          */
-/*************************************************************************/
-/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
-/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
-/*                                                                       */
-/* Permission is hereby granted, free of charge, to any person obtaining */
-/* a copy of this software and associated documentation files (the       */
-/* "Software"), to deal in the Software without restriction, including   */
-/* without limitation the rights to use, copy, modify, merge, publish,   */
-/* distribute, sublicense, and/or sell copies of the Software, and to    */
-/* permit persons to whom the Software is furnished to do so, subject to */
-/* the following conditions:                                             */
-/*                                                                       */
-/* The above copyright notice and this permission notice shall be        */
-/* included in all copies or substantial portions of the Software.       */
-/*                                                                       */
-/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
-/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
-/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
-/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
-/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
-/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
-/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
-/*************************************************************************/
-
-#include "broad_phase_2d_basic.h"
-
-BroadPhase2DBasic::ID BroadPhase2DBasic::create(CollisionObject2DSW *p_object_, int p_subindex) {
-	current++;
-
-	Element e;
-	e.owner = p_object_;
-	e._static = false;
-	e.subindex = p_subindex;
-
-	element_map[current] = e;
-	return current;
-}
-
-void BroadPhase2DBasic::move(ID p_id, const Rect2 &p_aabb) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-	E->get().aabb = p_aabb;
-}
-
-void BroadPhase2DBasic::set_static(ID p_id, bool p_static) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-	E->get()._static = p_static;
-}
-
-void BroadPhase2DBasic::remove(ID p_id) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-	element_map.erase(E);
-}
-
-CollisionObject2DSW *BroadPhase2DBasic::get_object(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, nullptr);
-	return E->get().owner;
-}
-
-bool BroadPhase2DBasic::is_static(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, false);
-	return E->get()._static;
-}
-
-int BroadPhase2DBasic::get_subindex(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, -1);
-	return E->get().subindex;
-}
-
-int BroadPhase2DBasic::cull_segment(const Vector2 &p_from, const Vector2 &p_to, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices) {
-	int rc = 0;
-
-	for (Map<ID, Element>::Element *E = element_map.front(); E; E = E->next()) {
-		const Rect2 aabb = E->get().aabb;
-		if (aabb.intersects_segment(p_from, p_to)) {
-			p_results[rc] = E->get().owner;
-			p_result_indices[rc] = E->get().subindex;
-			rc++;
-			if (rc >= p_max_results) {
-				break;
-			}
-		}
-	}
-
-	return rc;
-}
-
-int BroadPhase2DBasic::cull_aabb(const Rect2 &p_aabb, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices) {
-	int rc = 0;
-
-	for (Map<ID, Element>::Element *E = element_map.front(); E; E = E->next()) {
-		const Rect2 aabb = E->get().aabb;
-		if (aabb.intersects(p_aabb)) {
-			p_results[rc] = E->get().owner;
-			p_result_indices[rc] = E->get().subindex;
-			rc++;
-			if (rc >= p_max_results) {
-				break;
-			}
-		}
-	}
-
-	return rc;
-}
-
-void BroadPhase2DBasic::set_pair_callback(PairCallback p_pair_callback, void *p_userdata) {
-	pair_userdata = p_userdata;
-	pair_callback = p_pair_callback;
-}
-
-void BroadPhase2DBasic::set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata) {
-	unpair_userdata = p_userdata;
-	unpair_callback = p_unpair_callback;
-}
-
-void BroadPhase2DBasic::update() {
-	// recompute pairs
-	for (Map<ID, Element>::Element *I = element_map.front(); I; I = I->next()) {
-		for (Map<ID, Element>::Element *J = I->next(); J; J = J->next()) {
-			Element *elem_A = &I->get();
-			Element *elem_B = &J->get();
-
-			if (elem_A->owner == elem_B->owner) {
-				continue;
-			}
-
-			bool pair_ok = elem_A->aabb.intersects(elem_B->aabb) && (!elem_A->_static || !elem_B->_static);
-
-			PairKey key(I->key(), J->key());
-
-			Map<PairKey, void *>::Element *E = pair_map.find(key);
-
-			if (!pair_ok && E) {
-				if (unpair_callback) {
-					unpair_callback(elem_A->owner, elem_A->subindex, elem_B->owner, elem_B->subindex, E->get(), unpair_userdata);
-				}
-				pair_map.erase(key);
-			}
-
-			if (pair_ok && !E) {
-				void *data = nullptr;
-				if (pair_callback) {
-					data = pair_callback(elem_A->owner, elem_A->subindex, elem_B->owner, elem_B->subindex, unpair_userdata);
-					if (data) {
-						pair_map.insert(key, data);
-					}
-				}
-			}
-		}
-	}
-}
-
-BroadPhase2DSW *BroadPhase2DBasic::_create() {
-	return memnew(BroadPhase2DBasic);
-}
-
-BroadPhase2DBasic::BroadPhase2DBasic() {
-	current = 1;
-	unpair_callback = nullptr;
-	unpair_userdata = nullptr;
-	pair_callback = nullptr;
-	pair_userdata = nullptr;
-}
diff --git a/servers/physics_2d/broad_phase_2d_bvh.cpp b/servers/physics_2d/broad_phase_2d_bvh.cpp
new file mode 100644
index 0000000000..5f53f4a012
--- /dev/null
+++ b/servers/physics_2d/broad_phase_2d_bvh.cpp
@@ -0,0 +1,116 @@
+/*************************************************************************/
+/*  broad_phase_2d_bvh.cpp                                               */
+/*************************************************************************/
+/*                       This file is part of:                           */
+/*                           GODOT ENGINE                                */
+/*                      https://godotengine.org                          */
+/*************************************************************************/
+/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
+/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
+/*                                                                       */
+/* Permission is hereby granted, free of charge, to any person obtaining */
+/* a copy of this software and associated documentation files (the       */
+/* "Software"), to deal in the Software without restriction, including   */
+/* without limitation the rights to use, copy, modify, merge, publish,   */
+/* distribute, sublicense, and/or sell copies of the Software, and to    */
+/* permit persons to whom the Software is furnished to do so, subject to */
+/* the following conditions:                                             */
+/*                                                                       */
+/* The above copyright notice and this permission notice shall be        */
+/* included in all copies or substantial portions of the Software.       */
+/*                                                                       */
+/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
+/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
+/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
+/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
+/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
+/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
+/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
+/*************************************************************************/
+
+#include "broad_phase_2d_bvh.h"
+#include "collision_object_2d_sw.h"
+
+BroadPhase2DSW::ID BroadPhase2DBVH::create(CollisionObject2DSW *p_object, int p_subindex, const Rect2 &p_aabb, bool p_static) {
+	ID oid = bvh.create(p_object, true, p_aabb, p_subindex, !p_static, 1 << p_object->get_type(), p_static ? 0 : 0xFFFFF); // Pair everything, don't care?
+	return oid + 1;
+}
+
+void BroadPhase2DBVH::move(ID p_id, const Rect2 &p_aabb) {
+	bvh.move(p_id - 1, p_aabb);
+}
+
+void BroadPhase2DBVH::set_static(ID p_id, bool p_static) {
+	CollisionObject2DSW *it = bvh.get(p_id - 1);
+	bvh.set_pairable(p_id - 1, !p_static, 1 << it->get_type(), p_static ? 0 : 0xFFFFF, false); // Pair everything, don't care?
+}
+
+void BroadPhase2DBVH::remove(ID p_id) {
+	bvh.erase(p_id - 1);
+}
+
+CollisionObject2DSW *BroadPhase2DBVH::get_object(ID p_id) const {
+	CollisionObject2DSW *it = bvh.get(p_id - 1);
+	ERR_FAIL_COND_V(!it, nullptr);
+	return it;
+}
+
+bool BroadPhase2DBVH::is_static(ID p_id) const {
+	return !bvh.is_pairable(p_id - 1);
+}
+
+int BroadPhase2DBVH::get_subindex(ID p_id) const {
+	return bvh.get_subindex(p_id - 1);
+}
+
+int BroadPhase2DBVH::cull_segment(const Vector2 &p_from, const Vector2 &p_to, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices) {
+	return bvh.cull_segment(p_from, p_to, p_results, p_max_results, p_result_indices);
+}
+
+int BroadPhase2DBVH::cull_aabb(const Rect2 &p_aabb, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices) {
+	return bvh.cull_aabb(p_aabb, p_results, p_max_results, p_result_indices);
+}
+
+void *BroadPhase2DBVH::_pair_callback(void *self, uint32_t p_A, CollisionObject2DSW *p_object_A, int subindex_A, uint32_t p_B, CollisionObject2DSW *p_object_B, int subindex_B) {
+	BroadPhase2DBVH *bpo = (BroadPhase2DBVH *)(self);
+	if (!bpo->pair_callback) {
+		return nullptr;
+	}
+
+	return bpo->pair_callback(p_object_A, subindex_A, p_object_B, subindex_B, bpo->pair_userdata);
+}
+
+void BroadPhase2DBVH::_unpair_callback(void *self, uint32_t p_A, CollisionObject2DSW *p_object_A, int subindex_A, uint32_t p_B, CollisionObject2DSW *p_object_B, int subindex_B, void *pairdata) {
+	BroadPhase2DBVH *bpo = (BroadPhase2DBVH *)(self);
+	if (!bpo->unpair_callback) {
+		return;
+	}
+
+	bpo->unpair_callback(p_object_A, subindex_A, p_object_B, subindex_B, pairdata, bpo->unpair_userdata);
+}
+
+void BroadPhase2DBVH::set_pair_callback(PairCallback p_pair_callback, void *p_userdata) {
+	pair_callback = p_pair_callback;
+	pair_userdata = p_userdata;
+}
+
+void BroadPhase2DBVH::set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata) {
+	unpair_callback = p_unpair_callback;
+	unpair_userdata = p_userdata;
+}
+
+void BroadPhase2DBVH::update() {
+	bvh.update();
+}
+
+BroadPhase2DSW *BroadPhase2DBVH::_create() {
+	return memnew(BroadPhase2DBVH);
+}
+
+BroadPhase2DBVH::BroadPhase2DBVH() {
+	bvh.set_pair_callback(_pair_callback, this);
+	bvh.set_unpair_callback(_unpair_callback, this);
+	pair_callback = nullptr;
+	pair_userdata = nullptr;
+	unpair_userdata = nullptr;
+}
diff --git a/servers/physics_2d/broad_phase_2d_basic.h b/servers/physics_2d/broad_phase_2d_bvh.h
index ca1db360fb..6c11d2561b 100644
--- a/servers/physics_2d/broad_phase_2d_basic.h
+++ b/servers/physics_2d/broad_phase_2d_bvh.h
@@ -1,5 +1,5 @@
 /*************************************************************************/
-/*  broad_phase_2d_basic.h                                               */
+/*  broad_phase_2d_bvh.h                                                 */
 /*************************************************************************/
 /*                       This file is part of:                           */
 /*                           GODOT ENGINE                                */
@@ -28,49 +28,19 @@
 /* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
 /*************************************************************************/
 
-#ifndef BROAD_PHASE_2D_BASIC_H
-#define BROAD_PHASE_2D_BASIC_H
+#ifndef BROAD_PHASE_2D_BVH_H
+#define BROAD_PHASE_2D_BVH_H
 
-#include "core/templates/map.h"
-#include "space_2d_sw.h"
-class BroadPhase2DBasic : public BroadPhase2DSW {
-	struct Element {
-		CollisionObject2DSW *owner;
-		bool _static;
-		Rect2 aabb;
-		int subindex;
-	};
+#include "broad_phase_2d_sw.h"
+#include "core/math/bvh.h"
+#include "core/math/rect2.h"
+#include "core/math/vector2.h"
 
-	Map<ID, Element> element_map;
+class BroadPhase2DBVH : public BroadPhase2DSW {
+	BVH_Manager<CollisionObject2DSW, true, 128, Rect2, Vector2> bvh;
 
-	ID current;
-
-	struct PairKey {
-		union {
-			struct {
-				ID a;
-				ID b;
-			};
-			uint64_t key;
-		};
-
-		_FORCE_INLINE_ bool operator<(const PairKey &p_key) const {
-			return key < p_key.key;
-		}
-
-		PairKey() { key = 0; }
-		PairKey(ID p_a, ID p_b) {
-			if (p_a > p_b) {
-				a = p_b;
-				b = p_a;
-			} else {
-				a = p_a;
-				b = p_b;
-			}
-		}
-	};
-
-	Map<PairKey, void *> pair_map;
+	static void *_pair_callback(void *, uint32_t, CollisionObject2DSW *, int, uint32_t, CollisionObject2DSW *, int);
+	static void _unpair_callback(void *, uint32_t, CollisionObject2DSW *, int, uint32_t, CollisionObject2DSW *, int, void *);
 
 	PairCallback pair_callback;
 	void *pair_userdata;
@@ -79,7 +49,7 @@ class BroadPhase2DBasic : public BroadPhase2DSW {
 
 public:
 	// 0 is an invalid ID
-	virtual ID create(CollisionObject2DSW *p_object_, int p_subindex = 0);
+	virtual ID create(CollisionObject2DSW *p_object, int p_subindex = 0, const Rect2 &p_aabb = Rect2(), bool p_static = false);
 	virtual void move(ID p_id, const Rect2 &p_aabb);
 	virtual void set_static(ID p_id, bool p_static);
 	virtual void remove(ID p_id);
@@ -97,7 +67,7 @@ public:
 	virtual void update();
 
 	static BroadPhase2DSW *_create();
-	BroadPhase2DBasic();
+	BroadPhase2DBVH();
 };
 
-#endif // BROAD_PHASE_2D_BASIC_H
+#endif // BROAD_PHASE_2D_BVH_H
diff --git a/servers/physics_2d/broad_phase_2d_hash_grid.cpp b/servers/physics_2d/broad_phase_2d_hash_grid.cpp
deleted file mode 100644
index 35447c5389..0000000000
--- a/servers/physics_2d/broad_phase_2d_hash_grid.cpp
+++ /dev/null
@@ -1,738 +0,0 @@
-/*************************************************************************/
-/*  broad_phase_2d_hash_grid.cpp                                         */
-/*************************************************************************/
-/*                       This file is part of:                           */
-/*                           GODOT ENGINE                                */
-/*                      https://godotengine.org                          */
-/*************************************************************************/
-/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
-/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
-/*                                                                       */
-/* Permission is hereby granted, free of charge, to any person obtaining */
-/* a copy of this software and associated documentation files (the       */
-/* "Software"), to deal in the Software without restriction, including   */
-/* without limitation the rights to use, copy, modify, merge, publish,   */
-/* distribute, sublicense, and/or sell copies of the Software, and to    */
-/* permit persons to whom the Software is furnished to do so, subject to */
-/* the following conditions:                                             */
-/*                                                                       */
-/* The above copyright notice and this permission notice shall be        */
-/* included in all copies or substantial portions of the Software.       */
-/*                                                                       */
-/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
-/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
-/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
-/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
-/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
-/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
-/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
-/*************************************************************************/
-
-#include "broad_phase_2d_hash_grid.h"
-#include "collision_object_2d_sw.h"
-#include "core/config/project_settings.h"
-
-#define LARGE_ELEMENT_FI 1.01239812
-
-void BroadPhase2DHashGrid::_pair_attempt(Element *p_elem, Element *p_with) {
-	if (p_elem->owner == p_with->owner) {
-		return;
-	}
-	if (!_test_collision_mask(p_elem->collision_mask, p_elem->collision_layer, p_with->collision_mask, p_with->collision_layer)) {
-		return;
-	}
-	Map<Element *, PairData *>::Element *E = p_elem->paired.find(p_with);
-
-	ERR_FAIL_COND(p_elem->_static && p_with->_static);
-
-	if (!E) {
-		PairData *pd = memnew(PairData);
-		p_elem->paired[p_with] = pd;
-		p_with->paired[p_elem] = pd;
-	} else {
-		E->get()->rc++;
-	}
-}
-
-void BroadPhase2DHashGrid::_unpair_attempt(Element *p_elem, Element *p_with) {
-	if (p_elem->owner == p_with->owner) {
-		return;
-	}
-	if (!_test_collision_mask(p_elem->collision_mask, p_elem->collision_layer, p_with->collision_mask, p_with->collision_layer)) {
-		return;
-	}
-	Map<Element *, PairData *>::Element *E = p_elem->paired.find(p_with);
-
-	ERR_FAIL_COND(!E); //this should really be paired..
-
-	E->get()->rc--;
-
-	if (E->get()->rc == 0) {
-		if (E->get()->colliding) {
-			//uncollide
-			if (unpair_callback) {
-				unpair_callback(p_elem->owner, p_elem->subindex, p_with->owner, p_with->subindex, E->get()->ud, unpair_userdata);
-			}
-		}
-
-		memdelete(E->get());
-		p_elem->paired.erase(E);
-		p_with->paired.erase(p_elem);
-	}
-}
-
-void BroadPhase2DHashGrid::_check_motion(Element *p_elem) {
-	for (Map<Element *, PairData *>::Element *E = p_elem->paired.front(); E; E = E->next()) {
-		bool physical_collision = p_elem->aabb.intersects(E->key()->aabb);
-		bool logical_collision = p_elem->owner->test_collision_mask(E->key()->owner);
-
-		if (physical_collision && logical_collision) {
-			if (!E->get()->colliding && pair_callback) {
-				E->get()->ud = pair_callback(p_elem->owner, p_elem->subindex, E->key()->owner, E->key()->subindex, pair_userdata);
-			}
-			E->get()->colliding = true;
-		} else { // No collision
-			if (E->get()->colliding && unpair_callback) {
-				unpair_callback(p_elem->owner, p_elem->subindex, E->key()->owner, E->key()->subindex, E->get()->ud, unpair_userdata);
-				E->get()->ud = nullptr;
-			}
-			E->get()->colliding = false;
-		}
-	}
-}
-
-void BroadPhase2DHashGrid::_enter_grid(Element *p_elem, const Rect2 &p_rect, bool p_static, bool p_force_enter) {
-	Vector2 sz = (p_rect.size / cell_size * LARGE_ELEMENT_FI); //use magic number to avoid floating point issues
-	if (sz.width * sz.height > large_object_min_surface) {
-		//large object, do not use grid, must check against all elements
-		for (Map<ID, Element>::Element *E = element_map.front(); E; E = E->next()) {
-			if (E->key() == p_elem->self) {
-				continue; // do not pair against itself
-			}
-			if (E->get()._static && p_static) {
-				continue;
-			}
-
-			_pair_attempt(p_elem, &E->get());
-		}
-
-		large_elements[p_elem].inc();
-		return;
-	}
-
-	Point2i from = (p_rect.position / cell_size).floor();
-	Point2i to = ((p_rect.position + p_rect.size) / cell_size).floor();
-
-	for (int i = from.x; i <= to.x; i++) {
-		for (int j = from.y; j <= to.y; j++) {
-			PosKey pk;
-			pk.x = i;
-			pk.y = j;
-
-			uint32_t idx = pk.hash() % hash_table_size;
-			PosBin *pb = hash_table[idx];
-
-			while (pb) {
-				if (pb->key == pk) {
-					break;
-				}
-
-				pb = pb->next;
-			}
-
-			bool entered = p_force_enter;
-
-			if (!pb) {
-				//does not exist, create!
-				pb = memnew(PosBin);
-				pb->key = pk;
-				pb->next = hash_table[idx];
-				hash_table[idx] = pb;
-			}
-
-			if (p_static) {
-				if (pb->static_object_set[p_elem].inc() == 1) {
-					entered = true;
-				}
-			} else {
-				if (pb->object_set[p_elem].inc() == 1) {
-					entered = true;
-				}
-			}
-
-			if (entered) {
-				for (Map<Element *, RC>::Element *E = pb->object_set.front(); E; E = E->next()) {
-					_pair_attempt(p_elem, E->key());
-				}
-
-				if (!p_static) {
-					for (Map<Element *, RC>::Element *E = pb->static_object_set.front(); E; E = E->next()) {
-						_pair_attempt(p_elem, E->key());
-					}
-				}
-			}
-		}
-	}
-
-	//pair separatedly with large elements
-
-	for (Map<Element *, RC>::Element *E = large_elements.front(); E; E = E->next()) {
-		if (E->key() == p_elem) {
-			continue; // do not pair against itself
-		}
-		if (E->key()->_static && p_static) {
-			continue;
-		}
-		_pair_attempt(E->key(), p_elem);
-	}
-}
-
-void BroadPhase2DHashGrid::_exit_grid(Element *p_elem, const Rect2 &p_rect, bool p_static, bool p_force_exit) {
-	Vector2 sz = (p_rect.size / cell_size * LARGE_ELEMENT_FI);
-	if (sz.width * sz.height > large_object_min_surface) {
-		//unpair all elements, instead of checking all, just check what is already paired, so we at least save from checking static vs static
-		Map<Element *, PairData *>::Element *E = p_elem->paired.front();
-		while (E) {
-			Map<Element *, PairData *>::Element *next = E->next();
-			_unpair_attempt(p_elem, E->key());
-			E = next;
-		}
-
-		if (large_elements[p_elem].dec() == 0) {
-			large_elements.erase(p_elem);
-		}
-		return;
-	}
-
-	Point2i from = (p_rect.position / cell_size).floor();
-	Point2i to = ((p_rect.position + p_rect.size) / cell_size).floor();
-
-	for (int i = from.x; i <= to.x; i++) {
-		for (int j = from.y; j <= to.y; j++) {
-			PosKey pk;
-			pk.x = i;
-			pk.y = j;
-
-			uint32_t idx = pk.hash() % hash_table_size;
-			PosBin *pb = hash_table[idx];
-
-			while (pb) {
-				if (pb->key == pk) {
-					break;
-				}
-
-				pb = pb->next;
-			}
-
-			ERR_CONTINUE(!pb); //should exist!!
-
-			bool exited = p_force_exit;
-
-			if (p_static) {
-				if (pb->static_object_set[p_elem].dec() == 0) {
-					pb->static_object_set.erase(p_elem);
-					exited = true;
-				}
-			} else {
-				if (pb->object_set[p_elem].dec() == 0) {
-					pb->object_set.erase(p_elem);
-					exited = true;
-				}
-			}
-
-			if (exited) {
-				for (Map<Element *, RC>::Element *E = pb->object_set.front(); E; E = E->next()) {
-					_unpair_attempt(p_elem, E->key());
-				}
-
-				if (!p_static) {
-					for (Map<Element *, RC>::Element *E = pb->static_object_set.front(); E; E = E->next()) {
-						_unpair_attempt(p_elem, E->key());
-					}
-				}
-			}
-
-			if (pb->object_set.is_empty() && pb->static_object_set.is_empty()) {
-				if (hash_table[idx] == pb) {
-					hash_table[idx] = pb->next;
-				} else {
-					PosBin *px = hash_table[idx];
-
-					while (px) {
-						if (px->next == pb) {
-							px->next = pb->next;
-							break;
-						}
-
-						px = px->next;
-					}
-
-					ERR_CONTINUE(!px);
-				}
-
-				memdelete(pb);
-			}
-		}
-	}
-
-	for (Map<Element *, RC>::Element *E = large_elements.front(); E; E = E->next()) {
-		if (E->key() == p_elem) {
-			continue; // do not pair against itself
-		}
-		if (E->key()->_static && p_static) {
-			continue;
-		}
-
-		//unpair from large elements
-		_unpair_attempt(p_elem, E->key());
-	}
-}
-
-BroadPhase2DHashGrid::ID BroadPhase2DHashGrid::create(CollisionObject2DSW *p_object, int p_subindex) {
-	current++;
-
-	Element e;
-	e.owner = p_object;
-	e._static = false;
-	e.collision_mask = p_object->get_collision_mask();
-	e.collision_layer = p_object->get_collision_layer();
-	e.subindex = p_subindex;
-	e.self = current;
-	e.pass = 0;
-
-	element_map[current] = e;
-	return current;
-}
-
-void BroadPhase2DHashGrid::move(ID p_id, const Rect2 &p_aabb) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-
-	Element &e = E->get();
-	bool layer_changed = e.collision_mask != e.owner->get_collision_mask() || e.collision_layer != e.owner->get_collision_layer();
-
-	if (p_aabb != e.aabb || layer_changed) {
-		uint32_t old_mask = e.collision_mask;
-		uint32_t old_layer = e.collision_layer;
-		if (p_aabb != Rect2()) {
-			e.collision_mask = e.owner->get_collision_mask();
-			e.collision_layer = e.owner->get_collision_layer();
-
-			_enter_grid(&e, p_aabb, e._static, layer_changed);
-		}
-		if (e.aabb != Rect2()) {
-			// Need _exit_grid to remove from cells based on the old layer values.
-			e.collision_mask = old_mask;
-			e.collision_layer = old_layer;
-
-			_exit_grid(&e, e.aabb, e._static, layer_changed);
-
-			e.collision_mask = e.owner->get_collision_mask();
-			e.collision_layer = e.owner->get_collision_layer();
-		}
-		e.aabb = p_aabb;
-	}
-
-	_check_motion(&e);
-}
-
-void BroadPhase2DHashGrid::set_static(ID p_id, bool p_static) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-
-	Element &e = E->get();
-
-	if (e._static == p_static) {
-		return;
-	}
-
-	if (e.aabb != Rect2()) {
-		_exit_grid(&e, e.aabb, e._static, false);
-	}
-
-	e._static = p_static;
-
-	if (e.aabb != Rect2()) {
-		_enter_grid(&e, e.aabb, e._static, false);
-		_check_motion(&e);
-	}
-}
-
-void BroadPhase2DHashGrid::remove(ID p_id) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-
-	Element &e = E->get();
-
-	if (e.aabb != Rect2()) {
-		_exit_grid(&e, e.aabb, e._static, false);
-	}
-
-	element_map.erase(p_id);
-}
-
-CollisionObject2DSW *BroadPhase2DHashGrid::get_object(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, nullptr);
-	return E->get().owner;
-}
-
-bool BroadPhase2DHashGrid::is_static(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, false);
-	return E->get()._static;
-}
-
-int BroadPhase2DHashGrid::get_subindex(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, -1);
-	return E->get().subindex;
-}
-
-template <bool use_aabb, bool use_segment>
-void BroadPhase2DHashGrid::_cull(const Point2i p_cell, const Rect2 &p_aabb, const Point2 &p_from, const Point2 &p_to, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices, int &index) {
-	PosKey pk;
-	pk.x = p_cell.x;
-	pk.y = p_cell.y;
-
-	uint32_t idx = pk.hash() % hash_table_size;
-	PosBin *pb = hash_table[idx];
-
-	while (pb) {
-		if (pb->key == pk) {
-			break;
-		}
-
-		pb = pb->next;
-	}
-
-	if (!pb) {
-		return;
-	}
-
-	for (Map<Element *, RC>::Element *E = pb->object_set.front(); E; E = E->next()) {
-		if (index >= p_max_results) {
-			break;
-		}
-		if (E->key()->pass == pass) {
-			continue;
-		}
-
-		E->key()->pass = pass;
-
-		if (use_aabb && !p_aabb.intersects(E->key()->aabb)) {
-			continue;
-		}
-
-		if (use_segment && !E->key()->aabb.intersects_segment(p_from, p_to)) {
-			continue;
-		}
-
-		p_results[index] = E->key()->owner;
-		p_result_indices[index] = E->key()->subindex;
-		index++;
-	}
-
-	for (Map<Element *, RC>::Element *E = pb->static_object_set.front(); E; E = E->next()) {
-		if (index >= p_max_results) {
-			break;
-		}
-		if (E->key()->pass == pass) {
-			continue;
-		}
-
-		if (use_aabb && !p_aabb.intersects(E->key()->aabb)) {
-			continue;
-		}
-
-		if (use_segment && !E->key()->aabb.intersects_segment(p_from, p_to)) {
-			continue;
-		}
-
-		E->key()->pass = pass;
-		p_results[index] = E->key()->owner;
-		p_result_indices[index] = E->key()->subindex;
-		index++;
-	}
-}
-
-int BroadPhase2DHashGrid::cull_segment(const Vector2 &p_from, const Vector2 &p_to, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices) {
-	pass++;
-
-	Vector2 dir = (p_to - p_from);
-	if (dir == Vector2()) {
-		return 0;
-	}
-	//avoid divisions by zero
-	dir.normalize();
-	if (dir.x == 0.0) {
-		dir.x = 0.000001;
-	}
-	if (dir.y == 0.0) {
-		dir.y = 0.000001;
-	}
-	Vector2 delta = dir.abs();
-
-	delta.x = cell_size / delta.x;
-	delta.y = cell_size / delta.y;
-
-	Point2i pos = (p_from / cell_size).floor();
-	Point2i end = (p_to / cell_size).floor();
-
-	Point2i step = Vector2(SGN(dir.x), SGN(dir.y));
-
-	Vector2 max;
-
-	if (dir.x < 0) {
-		max.x = (Math::floor((double)pos.x) * cell_size - p_from.x) / dir.x;
-	} else {
-		max.x = (Math::floor((double)pos.x + 1) * cell_size - p_from.x) / dir.x;
-	}
-
-	if (dir.y < 0) {
-		max.y = (Math::floor((double)pos.y) * cell_size - p_from.y) / dir.y;
-	} else {
-		max.y = (Math::floor((double)pos.y + 1) * cell_size - p_from.y) / dir.y;
-	}
-
-	int cullcount = 0;
-	_cull<false, true>(pos, Rect2(), p_from, p_to, p_results, p_max_results, p_result_indices, cullcount);
-
-	bool reached_x = false;
-	bool reached_y = false;
-
-	while (true) {
-		if (max.x < max.y) {
-			max.x += delta.x;
-			pos.x += step.x;
-		} else {
-			max.y += delta.y;
-			pos.y += step.y;
-		}
-
-		if (step.x > 0) {
-			if (pos.x >= end.x) {
-				reached_x = true;
-			}
-		} else if (pos.x <= end.x) {
-			reached_x = true;
-		}
-
-		if (step.y > 0) {
-			if (pos.y >= end.y) {
-				reached_y = true;
-			}
-		} else if (pos.y <= end.y) {
-			reached_y = true;
-		}
-
-		_cull<false, true>(pos, Rect2(), p_from, p_to, p_results, p_max_results, p_result_indices, cullcount);
-
-		if (reached_x && reached_y) {
-			break;
-		}
-	}
-
-	for (Map<Element *, RC>::Element *E = large_elements.front(); E; E = E->next()) {
-		if (cullcount >= p_max_results) {
-			break;
-		}
-		if (E->key()->pass == pass) {
-			continue;
-		}
-
-		E->key()->pass = pass;
-
-		/*
-		if (use_aabb && !p_aabb.intersects(E->key()->aabb))
-			continue;
-		*/
-
-		if (!E->key()->aabb.intersects_segment(p_from, p_to)) {
-			continue;
-		}
-
-		p_results[cullcount] = E->key()->owner;
-		p_result_indices[cullcount] = E->key()->subindex;
-		cullcount++;
-	}
-
-	return cullcount;
-}
-
-int BroadPhase2DHashGrid::cull_aabb(const Rect2 &p_aabb, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices) {
-	pass++;
-
-	Point2i from = (p_aabb.position / cell_size).floor();
-	Point2i to = ((p_aabb.position + p_aabb.size) / cell_size).floor();
-	int cullcount = 0;
-
-	for (int i = from.x; i <= to.x; i++) {
-		for (int j = from.y; j <= to.y; j++) {
-			_cull<true, false>(Point2i(i, j), p_aabb, Point2(), Point2(), p_results, p_max_results, p_result_indices, cullcount);
-		}
-	}
-
-	for (Map<Element *, RC>::Element *E = large_elements.front(); E; E = E->next()) {
-		if (cullcount >= p_max_results) {
-			break;
-		}
-		if (E->key()->pass == pass) {
-			continue;
-		}
-
-		E->key()->pass = pass;
-
-		if (!p_aabb.intersects(E->key()->aabb)) {
-			continue;
-		}
-
-		/*
-		if (!E->key()->aabb.intersects_segment(p_from,p_to))
-			continue;
-		*/
-
-		p_results[cullcount] = E->key()->owner;
-		p_result_indices[cullcount] = E->key()->subindex;
-		cullcount++;
-	}
-	return cullcount;
-}
-
-void BroadPhase2DHashGrid::set_pair_callback(PairCallback p_pair_callback, void *p_userdata) {
-	pair_callback = p_pair_callback;
-	pair_userdata = p_userdata;
-}
-
-void BroadPhase2DHashGrid::set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata) {
-	unpair_callback = p_unpair_callback;
-	unpair_userdata = p_userdata;
-}
-
-void BroadPhase2DHashGrid::update() {
-}
-
-BroadPhase2DSW *BroadPhase2DHashGrid::_create() {
-	return memnew(BroadPhase2DHashGrid);
-}
-
-BroadPhase2DHashGrid::BroadPhase2DHashGrid() {
-	hash_table_size = GLOBAL_DEF("physics/2d/bp_hash_table_size", 4096);
-	ProjectSettings::get_singleton()->set_custom_property_info("physics/2d/bp_hash_table_size", PropertyInfo(Variant::INT, "physics/2d/bp_hash_table_size", PROPERTY_HINT_RANGE, "0,8192,1,or_greater"));
-	hash_table_size = Math::larger_prime(hash_table_size);
-	hash_table = memnew_arr(PosBin *, hash_table_size);
-
-	cell_size = GLOBAL_DEF("physics/2d/cell_size", 128);
-	ProjectSettings::get_singleton()->set_custom_property_info("physics/2d/cell_size", PropertyInfo(Variant::INT, "physics/2d/cell_size", PROPERTY_HINT_RANGE, "0,512,1,or_greater"));
-
-	large_object_min_surface = GLOBAL_DEF("physics/2d/large_object_surface_threshold_in_cells", 512);
-	ProjectSettings::get_singleton()->set_custom_property_info("physics/2d/large_object_surface_threshold_in_cells", PropertyInfo(Variant::INT, "physics/2d/large_object_surface_threshold_in_cells", PROPERTY_HINT_RANGE, "0,1024,1,or_greater"));
-
-	for (uint32_t i = 0; i < hash_table_size; i++) {
-		hash_table[i] = nullptr;
-	}
-	pass = 1;
-
-	current = 0;
-}
-
-BroadPhase2DHashGrid::~BroadPhase2DHashGrid() {
-	for (uint32_t i = 0; i < hash_table_size; i++) {
-		while (hash_table[i]) {
-			PosBin *pb = hash_table[i];
-			hash_table[i] = pb->next;
-			memdelete(pb);
-		}
-	}
-
-	memdelete_arr(hash_table);
-}
-
-/* 3D version of voxel traversal:
-
-public IEnumerable<Point3D> GetCellsOnRay(Ray ray, int maxDepth)
-{
-    // Implementation is based on:
-    // "A Fast Voxel Traversal Algorithm for Ray Tracing"
-    // John Amanatides, Andrew Woo
-    // http://www.cse.yorku.ca/~amana/research/grid.pdf
-    // https://web.archive.org/web/20100616193049/http://www.devmaster.net/articles/raytracing_series/A%20faster%20voxel%20traversal%20algorithm%20for%20ray%20tracing.pdf
-
-    // NOTES:
-    // * This code assumes that the ray's position and direction are in 'cell coordinates', which means
-    //   that one unit equals one cell in all directions.
-    // * When the ray doesn't start within the voxel grid, calculate the first position at which the
-    //   ray could enter the grid. If it never enters the grid, there is nothing more to do here.
-    // * Also, it is important to test when the ray exits the voxel grid when the grid isn't infinite.
-    // * The Point3D structure is a simple structure having three integer fields (X, Y and Z).
-
-    // The cell in which the ray starts.
-    Point3D start = GetCellAt(ray.Position);        // Rounds the position's X, Y and Z down to the nearest integer values.
-    int x = start.X;
-    int y = start.Y;
-    int z = start.Z;
-
-    // Determine which way we go.
-    int stepX = Math.Sign(ray.Direction.X);
-    int stepY = Math.Sign(ray.Direction.Y);
-    int stepZ = Math.Sign(ray.Direction.Z);
-
-    // Calculate cell boundaries. When the step (i.e. direction sign) is positive,
-    // the next boundary is AFTER our current position, meaning that we have to add 1.
-    // Otherwise, it is BEFORE our current position, in which case we add nothing.
-    Point3D cellBoundary = new Point3D(
-	x + (stepX > 0 ? 1 : 0),
-	y + (stepY > 0 ? 1 : 0),
-	z + (stepZ > 0 ? 1 : 0));
-
-    // NOTE: For the following calculations, the result will be Single.PositiveInfinity
-    // when ray.Direction.X, Y or Z equals zero, which is OK. However, when the left-hand
-    // value of the division also equals zero, the result is Single.NaN, which is not OK.
-
-    // Determine how far we can travel along the ray before we hit a voxel boundary.
-    Vector3 tMax = new Vector3(
-	(cellBoundary.X - ray.Position.X) / ray.Direction.X,    // Boundary is a plane on the YZ axis.
-	(cellBoundary.Y - ray.Position.Y) / ray.Direction.Y,    // Boundary is a plane on the XZ axis.
-	(cellBoundary.Z - ray.Position.Z) / ray.Direction.Z);    // Boundary is a plane on the XY axis.
-    if (Single.IsNaN(tMax.X)) tMax.X = Single.PositiveInfinity;
-    if (Single.IsNaN(tMax.Y)) tMax.Y = Single.PositiveInfinity;
-    if (Single.IsNaN(tMax.Z)) tMax.Z = Single.PositiveInfinity;
-
-    // Determine how far we must travel along the ray before we have crossed a gridcell.
-    Vector3 tDelta = new Vector3(
-	stepX / ray.Direction.X,                    // Crossing the width of a cell.
-	stepY / ray.Direction.Y,                    // Crossing the height of a cell.
-	stepZ / ray.Direction.Z);                    // Crossing the depth of a cell.
-    if (Single.IsNaN(tDelta.X)) tDelta.X = Single.PositiveInfinity;
-    if (Single.IsNaN(tDelta.Y)) tDelta.Y = Single.PositiveInfinity;
-    if (Single.IsNaN(tDelta.Z)) tDelta.Z = Single.PositiveInfinity;
-
-    // For each step, determine which distance to the next voxel boundary is lowest (i.e.
-    // which voxel boundary is nearest) and walk that way.
-    for (int i = 0; i < maxDepth; i++)
-    {
-	// Return it.
-	yield return new Point3D(x, y, z);
-
-	// Do the next step.
-	if (tMax.X < tMax.Y && tMax.X < tMax.Z)
-	{
-	    // tMax.X is the lowest, an YZ cell boundary plane is nearest.
-	    x += stepX;
-	    tMax.X += tDelta.X;
-	}
-	else if (tMax.Y < tMax.Z)
-	{
-	    // tMax.Y is the lowest, an XZ cell boundary plane is nearest.
-	    y += stepY;
-	    tMax.Y += tDelta.Y;
-	}
-	else
-	{
-	    // tMax.Z is the lowest, an XY cell boundary plane is nearest.
-	    z += stepZ;
-	    tMax.Z += tDelta.Z;
-	}
-    }
-
-    */
diff --git a/servers/physics_2d/broad_phase_2d_hash_grid.h b/servers/physics_2d/broad_phase_2d_hash_grid.h
deleted file mode 100644
index bb7c03b989..0000000000
--- a/servers/physics_2d/broad_phase_2d_hash_grid.h
+++ /dev/null
@@ -1,194 +0,0 @@
-/*************************************************************************/
-/*  broad_phase_2d_hash_grid.h                                           */
-/*************************************************************************/
-/*                       This file is part of:                           */
-/*                           GODOT ENGINE                                */
-/*                      https://godotengine.org                          */
-/*************************************************************************/
-/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
-/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
-/*                                                                       */
-/* Permission is hereby granted, free of charge, to any person obtaining */
-/* a copy of this software and associated documentation files (the       */
-/* "Software"), to deal in the Software without restriction, including   */
-/* without limitation the rights to use, copy, modify, merge, publish,   */
-/* distribute, sublicense, and/or sell copies of the Software, and to    */
-/* permit persons to whom the Software is furnished to do so, subject to */
-/* the following conditions:                                             */
-/*                                                                       */
-/* The above copyright notice and this permission notice shall be        */
-/* included in all copies or substantial portions of the Software.       */
-/*                                                                       */
-/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
-/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
-/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
-/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
-/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
-/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
-/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
-/*************************************************************************/
-
-#ifndef BROAD_PHASE_2D_HASH_GRID_H
-#define BROAD_PHASE_2D_HASH_GRID_H
-
-#include "broad_phase_2d_sw.h"
-#include "core/templates/map.h"
-
-class BroadPhase2DHashGrid : public BroadPhase2DSW {
-	struct PairData {
-		bool colliding;
-		int rc;
-		void *ud;
-		PairData() {
-			colliding = false;
-			rc = 1;
-			ud = nullptr;
-		}
-	};
-
-	struct Element {
-		ID self;
-		CollisionObject2DSW *owner;
-		bool _static;
-		Rect2 aabb;
-		// Owner's collision_mask/layer, used to detect changes in layers.
-		uint32_t collision_mask;
-		uint32_t collision_layer;
-		int subindex;
-		uint64_t pass;
-		Map<Element *, PairData *> paired;
-	};
-
-	struct RC {
-		int ref;
-
-		_FORCE_INLINE_ int inc() {
-			ref++;
-			return ref;
-		}
-		_FORCE_INLINE_ int dec() {
-			ref--;
-			return ref;
-		}
-
-		_FORCE_INLINE_ RC() {
-			ref = 0;
-		}
-	};
-
-	Map<ID, Element> element_map;
-	Map<Element *, RC> large_elements;
-
-	ID current;
-
-	uint64_t pass;
-
-	struct PairKey {
-		union {
-			struct {
-				ID a;
-				ID b;
-			};
-			uint64_t key;
-		};
-
-		_FORCE_INLINE_ bool operator<(const PairKey &p_key) const {
-			return key < p_key.key;
-		}
-
-		PairKey() { key = 0; }
-		PairKey(ID p_a, ID p_b) {
-			if (p_a > p_b) {
-				a = p_b;
-				b = p_a;
-			} else {
-				a = p_a;
-				b = p_b;
-			}
-		}
-	};
-
-	Map<PairKey, PairData> pair_map;
-
-	int cell_size;
-	int large_object_min_surface;
-
-	PairCallback pair_callback;
-	void *pair_userdata;
-	UnpairCallback unpair_callback;
-	void *unpair_userdata;
-
-	static _FORCE_INLINE_ bool _test_collision_mask(uint32_t p_mask1, uint32_t p_layer1, uint32_t p_mask2, uint32_t p_layer2) {
-		return p_mask1 & p_layer2 || p_mask2 & p_layer1;
-	}
-
-	void _enter_grid(Element *p_elem, const Rect2 &p_rect, bool p_static, bool p_force_enter);
-	void _exit_grid(Element *p_elem, const Rect2 &p_rect, bool p_static, bool p_force_exit);
-	template <bool use_aabb, bool use_segment>
-	_FORCE_INLINE_ void _cull(const Point2i p_cell, const Rect2 &p_aabb, const Point2 &p_from, const Point2 &p_to, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices, int &index);
-
-	struct PosKey {
-		union {
-			struct {
-				int32_t x;
-				int32_t y;
-			};
-			uint64_t key;
-		};
-
-		_FORCE_INLINE_ uint32_t hash() const {
-			uint64_t k = key;
-			k = (~k) + (k << 18); // k = (k << 18) - k - 1;
-			k = k ^ (k >> 31);
-			k = k * 21; // k = (k + (k << 2)) + (k << 4);
-			k = k ^ (k >> 11);
-			k = k + (k << 6);
-			k = k ^ (k >> 22);
-			return k;
-		}
-
-		bool operator==(const PosKey &p_key) const { return key == p_key.key; }
-		_FORCE_INLINE_ bool operator<(const PosKey &p_key) const {
-			return key < p_key.key;
-		}
-	};
-
-	struct PosBin {
-		PosKey key;
-		Map<Element *, RC> object_set;
-		Map<Element *, RC> static_object_set;
-		PosBin *next;
-	};
-
-	uint32_t hash_table_size;
-	PosBin **hash_table;
-
-	void _pair_attempt(Element *p_elem, Element *p_with);
-	void _unpair_attempt(Element *p_elem, Element *p_with);
-	void _check_motion(Element *p_elem);
-
-public:
-	virtual ID create(CollisionObject2DSW *p_object, int p_subindex = 0);
-	virtual void move(ID p_id, const Rect2 &p_aabb);
-	virtual void set_static(ID p_id, bool p_static);
-	virtual void remove(ID p_id);
-
-	virtual CollisionObject2DSW *get_object(ID p_id) const;
-	virtual bool is_static(ID p_id) const;
-	virtual int get_subindex(ID p_id) const;
-
-	virtual int cull_segment(const Vector2 &p_from, const Vector2 &p_to, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices = nullptr);
-	virtual int cull_aabb(const Rect2 &p_aabb, CollisionObject2DSW **p_results, int p_max_results, int *p_result_indices = nullptr);
-
-	virtual void set_pair_callback(PairCallback p_pair_callback, void *p_userdata);
-	virtual void set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata);
-
-	virtual void update();
-
-	static BroadPhase2DSW *_create();
-
-	BroadPhase2DHashGrid();
-	~BroadPhase2DHashGrid();
-};
-
-#endif // BROAD_PHASE_2D_HASH_GRID_H
diff --git a/servers/physics_2d/broad_phase_2d_sw.h b/servers/physics_2d/broad_phase_2d_sw.h
index d17ee6e2d6..0f82f06b9c 100644
--- a/servers/physics_2d/broad_phase_2d_sw.h
+++ b/servers/physics_2d/broad_phase_2d_sw.h
@@ -48,7 +48,7 @@ public:
 	typedef void (*UnpairCallback)(CollisionObject2DSW *A, int p_subindex_A, CollisionObject2DSW *B, int p_subindex_B, void *p_data, void *p_userdata);
 
 	// 0 is an invalid ID
-	virtual ID create(CollisionObject2DSW *p_object_, int p_subindex = 0) = 0;
+	virtual ID create(CollisionObject2DSW *p_object_, int p_subindex = 0, const Rect2 &p_aabb = Rect2(), bool p_static = false) = 0;
 	virtual void move(ID p_id, const Rect2 &p_aabb) = 0;
 	virtual void set_static(ID p_id, bool p_static) = 0;
 	virtual void remove(ID p_id) = 0;
diff --git a/servers/physics_2d/collision_object_2d_sw.cpp b/servers/physics_2d/collision_object_2d_sw.cpp
index 7a2f312263..fa87dc1f3f 100644
--- a/servers/physics_2d/collision_object_2d_sw.cpp
+++ b/servers/physics_2d/collision_object_2d_sw.cpp
@@ -182,19 +182,19 @@ void CollisionObject2DSW::_update_shapes() {
 			continue;
 		}
 
-		if (s.bpid == 0) {
-			s.bpid = space->get_broadphase()->create(this, i);
-			space->get_broadphase()->set_static(s.bpid, _static);
-		}
-
 		//not quite correct, should compute the next matrix..
 		Rect2 shape_aabb = s.shape->get_aabb();
 		Transform2D xform = transform * s.xform;
 		shape_aabb = xform.xform(shape_aabb);
+		shape_aabb.grow_by((s.aabb_cache.size.x + s.aabb_cache.size.y) * 0.5 * 0.05);
 		s.aabb_cache = shape_aabb;
-		s.aabb_cache = s.aabb_cache.grow((s.aabb_cache.size.x + s.aabb_cache.size.y) * 0.5 * 0.05);
 
-		space->get_broadphase()->move(s.bpid, s.aabb_cache);
+		if (s.bpid == 0) {
+			s.bpid = space->get_broadphase()->create(this, i, shape_aabb, _static);
+			space->get_broadphase()->set_static(s.bpid, _static);
+		}
+
+		space->get_broadphase()->move(s.bpid, shape_aabb);
 	}
 }
 
@@ -209,11 +209,6 @@ void CollisionObject2DSW::_update_shapes_with_motion(const Vector2 &p_motion) {
 			continue;
 		}
 
-		if (s.bpid == 0) {
-			s.bpid = space->get_broadphase()->create(this, i);
-			space->get_broadphase()->set_static(s.bpid, _static);
-		}
-
 		//not quite correct, should compute the next matrix..
 		Rect2 shape_aabb = s.shape->get_aabb();
 		Transform2D xform = transform * s.xform;
@@ -221,6 +216,11 @@ void CollisionObject2DSW::_update_shapes_with_motion(const Vector2 &p_motion) {
 		shape_aabb = shape_aabb.merge(Rect2(shape_aabb.position + p_motion, shape_aabb.size)); //use motion
 		s.aabb_cache = shape_aabb;
 
+		if (s.bpid == 0) {
+			s.bpid = space->get_broadphase()->create(this, i, shape_aabb, _static);
+			space->get_broadphase()->set_static(s.bpid, _static);
+		}
+
 		space->get_broadphase()->move(s.bpid, shape_aabb);
 	}
 }
diff --git a/servers/physics_2d/physics_server_2d_sw.cpp b/servers/physics_2d/physics_server_2d_sw.cpp
index 6d64f4126c..425546e5ca 100644
--- a/servers/physics_2d/physics_server_2d_sw.cpp
+++ b/servers/physics_2d/physics_server_2d_sw.cpp
@@ -30,8 +30,7 @@
 
 #include "physics_server_2d_sw.h"
 
-#include "broad_phase_2d_basic.h"
-#include "broad_phase_2d_hash_grid.h"
+#include "broad_phase_2d_bvh.h"
 #include "collision_solver_2d_sw.h"
 #include "core/config/project_settings.h"
 #include "core/debugger/engine_debugger.h"
@@ -1356,8 +1355,7 @@ PhysicsServer2DSW *PhysicsServer2DSW::singletonsw = nullptr;
 
 PhysicsServer2DSW::PhysicsServer2DSW(bool p_using_threads) {
 	singletonsw = this;
-	BroadPhase2DSW::create_func = BroadPhase2DHashGrid::_create;
-	//BroadPhase2DSW::create_func=BroadPhase2DBasic::_create;
+	BroadPhase2DSW::create_func = BroadPhase2DBVH::_create;
 
 	active = true;
 	island_count = 0;
diff --git a/servers/physics_3d/broad_phase_3d_basic.cpp b/servers/physics_3d/broad_phase_3d_basic.cpp
deleted file mode 100644
index b41c2530da..0000000000
--- a/servers/physics_3d/broad_phase_3d_basic.cpp
+++ /dev/null
@@ -1,212 +0,0 @@
-/*************************************************************************/
-/*  broad_phase_3d_basic.cpp                                             */
-/*************************************************************************/
-/*                       This file is part of:                           */
-/*                           GODOT ENGINE                                */
-/*                      https://godotengine.org                          */
-/*************************************************************************/
-/* Copyright (c) 2007-2021 Juan Linietsky, Ariel Manzur.                 */
-/* Copyright (c) 2014-2021 Godot Engine contributors (cf. AUTHORS.md).   */
-/*                                                                       */
-/* Permission is hereby granted, free of charge, to any person obtaining */
-/* a copy of this software and associated documentation files (the       */
-/* "Software"), to deal in the Software without restriction, including   */
-/* without limitation the rights to use, copy, modify, merge, publish,   */
-/* distribute, sublicense, and/or sell copies of the Software, and to    */
-/* permit persons to whom the Software is furnished to do so, subject to */
-/* the following conditions:                                             */
-/*                                                                       */
-/* The above copyright notice and this permission notice shall be        */
-/* included in all copies or substantial portions of the Software.       */
-/*                                                                       */
-/* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,       */
-/* EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF    */
-/* MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.*/
-/* IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY  */
-/* CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,  */
-/* TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE     */
-/* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
-/*************************************************************************/
-
-#include "broad_phase_3d_basic.h"
-#include "core/string/print_string.h"
-#include "core/templates/list.h"
-
-BroadPhase3DSW::ID BroadPhase3DBasic::create(CollisionObject3DSW *p_object, int p_subindex) {
-	ERR_FAIL_COND_V(p_object == nullptr, 0);
-
-	current++;
-
-	Element e;
-	e.owner = p_object;
-	e._static = false;
-	e.subindex = p_subindex;
-
-	element_map[current] = e;
-	return current;
-}
-
-void BroadPhase3DBasic::move(ID p_id, const AABB &p_aabb) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-	E->get().aabb = p_aabb;
-}
-
-void BroadPhase3DBasic::set_static(ID p_id, bool p_static) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-	E->get()._static = p_static;
-}
-
-void BroadPhase3DBasic::remove(ID p_id) {
-	Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND(!E);
-	List<PairKey> to_erase;
-	//unpair must be done immediately on removal to avoid potential invalid pointers
-	for (Map<PairKey, void *>::Element *F = pair_map.front(); F; F = F->next()) {
-		if (F->key().a == p_id || F->key().b == p_id) {
-			if (unpair_callback) {
-				Element *elem_A = &element_map[F->key().a];
-				Element *elem_B = &element_map[F->key().b];
-				unpair_callback(elem_A->owner, elem_A->subindex, elem_B->owner, elem_B->subindex, F->get(), unpair_userdata);
-			}
-			to_erase.push_back(F->key());
-		}
-	}
-	while (to_erase.size()) {
-		pair_map.erase(to_erase.front()->get());
-		to_erase.pop_front();
-	}
-	element_map.erase(E);
-}
-
-CollisionObject3DSW *BroadPhase3DBasic::get_object(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, nullptr);
-	return E->get().owner;
-}
-
-bool BroadPhase3DBasic::is_static(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, false);
-	return E->get()._static;
-}
-
-int BroadPhase3DBasic::get_subindex(ID p_id) const {
-	const Map<ID, Element>::Element *E = element_map.find(p_id);
-	ERR_FAIL_COND_V(!E, -1);
-	return E->get().subindex;
-}
-
-int BroadPhase3DBasic::cull_point(const Vector3 &p_point, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
-	int rc = 0;
-
-	for (Map<ID, Element>::Element *E = element_map.front(); E; E = E->next()) {
-		const AABB aabb = E->get().aabb;
-		if (aabb.has_point(p_point)) {
-			p_results[rc] = E->get().owner;
-			p_result_indices[rc] = E->get().subindex;
-			rc++;
-			if (rc >= p_max_results) {
-				break;
-			}
-		}
-	}
-
-	return rc;
-}
-
-int BroadPhase3DBasic::cull_segment(const Vector3 &p_from, const Vector3 &p_to, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
-	int rc = 0;
-
-	for (Map<ID, Element>::Element *E = element_map.front(); E; E = E->next()) {
-		const AABB aabb = E->get().aabb;
-		if (aabb.intersects_segment(p_from, p_to)) {
-			p_results[rc] = E->get().owner;
-			p_result_indices[rc] = E->get().subindex;
-			rc++;
-			if (rc >= p_max_results) {
-				break;
-			}
-		}
-	}
-
-	return rc;
-}
-
-int BroadPhase3DBasic::cull_aabb(const AABB &p_aabb, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
-	int rc = 0;
-
-	for (Map<ID, Element>::Element *E = element_map.front(); E; E = E->next()) {
-		const AABB aabb = E->get().aabb;
-		if (aabb.intersects(p_aabb)) {
-			p_results[rc] = E->get().owner;
-			p_result_indices[rc] = E->get().subindex;
-			rc++;
-			if (rc >= p_max_results) {
-				break;
-			}
-		}
-	}
-
-	return rc;
-}
-
-void BroadPhase3DBasic::set_pair_callback(PairCallback p_pair_callback, void *p_userdata) {
-	pair_userdata = p_userdata;
-	pair_callback = p_pair_callback;
-}
-
-void BroadPhase3DBasic::set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata) {
-	unpair_userdata = p_userdata;
-	unpair_callback = p_unpair_callback;
-}
-
-void BroadPhase3DBasic::update() {
-	// recompute pairs
-	for (Map<ID, Element>::Element *I = element_map.front(); I; I = I->next()) {
-		for (Map<ID, Element>::Element *J = I->next(); J; J = J->next()) {
-			Element *elem_A = &I->get();
-			Element *elem_B = &J->get();
-
-			if (elem_A->owner == elem_B->owner) {
-				continue;
-			}
-
-			bool pair_ok = elem_A->aabb.intersects(elem_B->aabb) && (!elem_A->_static || !elem_B->_static);
-
-			PairKey key(I->key(), J->key());
-
-			Map<PairKey, void *>::Element *E = pair_map.find(key);
-
-			if (!pair_ok && E) {
-				if (unpair_callback) {
-					unpair_callback(elem_A->owner, elem_A->subindex, elem_B->owner, elem_B->subindex, E->get(), unpair_userdata);
-				}
-				pair_map.erase(key);
-			}
-
-			if (pair_ok && !E) {
-				void *data = nullptr;
-				if (pair_callback) {
-					data = pair_callback(elem_A->owner, elem_A->subindex, elem_B->owner, elem_B->subindex, unpair_userdata);
-					if (data) {
-						pair_map.insert(key, data);
-					}
-				}
-			}
-		}
-	}
-}
-
-BroadPhase3DSW *BroadPhase3DBasic::_create() {
-	return memnew(BroadPhase3DBasic);
-}
-
-BroadPhase3DBasic::BroadPhase3DBasic() {
-	current = 1;
-	unpair_callback = nullptr;
-	unpair_userdata = nullptr;
-	pair_callback = nullptr;
-	pair_userdata = nullptr;
-}
diff --git a/servers/physics_3d/broad_phase_octree.cpp b/servers/physics_3d/broad_phase_3d_bvh.cpp
index 11324fa4e4..f9f64f786d 100644
--- a/servers/physics_3d/broad_phase_octree.cpp
+++ b/servers/physics_3d/broad_phase_3d_bvh.cpp
@@ -1,5 +1,5 @@
 /*************************************************************************/
-/*  broad_phase_octree.cpp                                               */
+/*  broad_phase_3d_bvh.cpp                                               */
 /*************************************************************************/
 /*                       This file is part of:                           */
 /*                           GODOT ENGINE                                */
@@ -28,55 +28,55 @@
 /* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
 /*************************************************************************/
 
-#include "broad_phase_octree.h"
+#include "broad_phase_3d_bvh.h"
 #include "collision_object_3d_sw.h"
 
-BroadPhase3DSW::ID BroadPhaseOctree::create(CollisionObject3DSW *p_object, int p_subindex) {
-	ID oid = octree.create(p_object, AABB(), p_subindex, false, 1 << p_object->get_type(), 0);
-	return oid;
+BroadPhase3DBVH::ID BroadPhase3DBVH::create(CollisionObject3DSW *p_object, int p_subindex, const AABB &p_aabb, bool p_static) {
+	ID oid = bvh.create(p_object, true, p_aabb, p_subindex, !p_static, 1 << p_object->get_type(), p_static ? 0 : 0xFFFFF); // Pair everything, don't care?
+	return oid + 1;
 }
 
-void BroadPhaseOctree::move(ID p_id, const AABB &p_aabb) {
-	octree.move(p_id, p_aabb);
+void BroadPhase3DBVH::move(ID p_id, const AABB &p_aabb) {
+	bvh.move(p_id - 1, p_aabb);
 }
 
-void BroadPhaseOctree::set_static(ID p_id, bool p_static) {
-	CollisionObject3DSW *it = octree.get(p_id);
-	octree.set_pairable(p_id, !p_static, 1 << it->get_type(), p_static ? 0 : 0xFFFFF); //pair everything, don't care 1?
+void BroadPhase3DBVH::set_static(ID p_id, bool p_static) {
+	CollisionObject3DSW *it = bvh.get(p_id - 1);
+	bvh.set_pairable(p_id - 1, !p_static, 1 << it->get_type(), p_static ? 0 : 0xFFFFF, false); // Pair everything, don't care?
 }
 
-void BroadPhaseOctree::remove(ID p_id) {
-	octree.erase(p_id);
+void BroadPhase3DBVH::remove(ID p_id) {
+	bvh.erase(p_id - 1);
 }
 
-CollisionObject3DSW *BroadPhaseOctree::get_object(ID p_id) const {
-	CollisionObject3DSW *it = octree.get(p_id);
+CollisionObject3DSW *BroadPhase3DBVH::get_object(ID p_id) const {
+	CollisionObject3DSW *it = bvh.get(p_id - 1);
 	ERR_FAIL_COND_V(!it, nullptr);
 	return it;
 }
 
-bool BroadPhaseOctree::is_static(ID p_id) const {
-	return !octree.is_pairable(p_id);
+bool BroadPhase3DBVH::is_static(ID p_id) const {
+	return !bvh.is_pairable(p_id - 1);
 }
 
-int BroadPhaseOctree::get_subindex(ID p_id) const {
-	return octree.get_subindex(p_id);
+int BroadPhase3DBVH::get_subindex(ID p_id) const {
+	return bvh.get_subindex(p_id - 1);
 }
 
-int BroadPhaseOctree::cull_point(const Vector3 &p_point, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
-	return octree.cull_point(p_point, p_results, p_max_results, p_result_indices);
+int BroadPhase3DBVH::cull_point(const Vector3 &p_point, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
+	return bvh.cull_point(p_point, p_results, p_max_results, p_result_indices);
 }
 
-int BroadPhaseOctree::cull_segment(const Vector3 &p_from, const Vector3 &p_to, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
-	return octree.cull_segment(p_from, p_to, p_results, p_max_results, p_result_indices);
+int BroadPhase3DBVH::cull_segment(const Vector3 &p_from, const Vector3 &p_to, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
+	return bvh.cull_segment(p_from, p_to, p_results, p_max_results, p_result_indices);
 }
 
-int BroadPhaseOctree::cull_aabb(const AABB &p_aabb, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
-	return octree.cull_aabb(p_aabb, p_results, p_max_results, p_result_indices);
+int BroadPhase3DBVH::cull_aabb(const AABB &p_aabb, CollisionObject3DSW **p_results, int p_max_results, int *p_result_indices) {
+	return bvh.cull_aabb(p_aabb, p_results, p_max_results, p_result_indices);
 }
 
-void *BroadPhaseOctree::_pair_callback(void *self, OctreeElementID p_A, CollisionObject3DSW *p_object_A, int subindex_A, OctreeElementID p_B, CollisionObject3DSW *p_object_B, int subindex_B) {
-	BroadPhaseOctree *bpo = (BroadPhaseOctree *)(self);
+void *BroadPhase3DBVH::_pair_callback(void *self, uint32_t p_A, CollisionObject3DSW *p_object_A, int subindex_A, uint32_t p_B, CollisionObject3DSW *p_object_B, int subindex_B) {
+	BroadPhase3DBVH *bpo = (BroadPhase3DBVH *)(self);
 	if (!bpo->pair_callback) {
 		return nullptr;
 	}
@@ -84,8 +84,8 @@ void *BroadPhaseOctree::_pair_callback(void *self, OctreeElementID p_A, Collisio
 	return bpo->pair_callback(p_object_A, subindex_A, p_object_B, subindex_B, bpo->pair_userdata);
 }
 
-void BroadPhaseOctree::_unpair_callback(void *self, OctreeElementID p_A, CollisionObject3DSW *p_object_A, int subindex_A, OctreeElementID p_B, CollisionObject3DSW *p_object_B, int subindex_B, void *pairdata) {
-	BroadPhaseOctree *bpo = (BroadPhaseOctree *)(self);
+void BroadPhase3DBVH::_unpair_callback(void *self, uint32_t p_A, CollisionObject3DSW *p_object_A, int subindex_A, uint32_t p_B, CollisionObject3DSW *p_object_B, int subindex_B, void *pairdata) {
+	BroadPhase3DBVH *bpo = (BroadPhase3DBVH *)(self);
 	if (!bpo->unpair_callback) {
 		return;
 	}
@@ -93,27 +93,27 @@ void BroadPhaseOctree::_unpair_callback(void *self, OctreeElementID p_A, Collisi
 	bpo->unpair_callback(p_object_A, subindex_A, p_object_B, subindex_B, pairdata, bpo->unpair_userdata);
 }
 
-void BroadPhaseOctree::set_pair_callback(PairCallback p_pair_callback, void *p_userdata) {
+void BroadPhase3DBVH::set_pair_callback(PairCallback p_pair_callback, void *p_userdata) {
 	pair_callback = p_pair_callback;
 	pair_userdata = p_userdata;
 }
 
-void BroadPhaseOctree::set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata) {
+void BroadPhase3DBVH::set_unpair_callback(UnpairCallback p_unpair_callback, void *p_userdata) {
 	unpair_callback = p_unpair_callback;
 	unpair_userdata = p_userdata;
 }
 
-void BroadPhaseOctree::update() {
-	// does.. not?
+void BroadPhase3DBVH::update() {
+	bvh.update();
 }
 
-BroadPhase3DSW *BroadPhaseOctree::_create() {
-	return memnew(BroadPhaseOctree);
+BroadPhase3DSW *BroadPhase3DBVH::_create() {
+	return memnew(BroadPhase3DBVH);
 }
 
-BroadPhaseOctree::BroadPhaseOctree() {
-	octree.set_pair_callback(_pair_callback, this);
-	octree.set_unpair_callback(_unpair_callback, this);
+BroadPhase3DBVH::BroadPhase3DBVH() {
+	bvh.set_pair_callback(_pair_callback, this);
+	bvh.set_unpair_callback(_unpair_callback, this);
 	pair_callback = nullptr;
 	pair_userdata = nullptr;
 	unpair_userdata = nullptr;
diff --git a/servers/physics_3d/broad_phase_octree.h b/servers/physics_3d/broad_phase_3d_bvh.h
index ee681dda96..30b8b7f2aa 100644
--- a/servers/physics_3d/broad_phase_octree.h
+++ b/servers/physics_3d/broad_phase_3d_bvh.h
@@ -1,5 +1,5 @@
 /*************************************************************************/
-/*  broad_phase_octree.h                                                 */
+/*  broad_phase_3d_bvh.h                                                 */
 /*************************************************************************/
 /*                       This file is part of:                           */
 /*                           GODOT ENGINE                                */
@@ -28,17 +28,17 @@
 /* SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.                */
 /*************************************************************************/
 
-#ifndef BROAD_PHASE_OCTREE_H
-#define BROAD_PHASE_OCTREE_H
+#ifndef BROAD_PHASE_3D_BVH_H
+#define BROAD_PHASE_3D_BVH_H
 
 #include "broad_phase_3d_sw.h"
-#include "core/math/octree.h"
+#include "core/math/bvh.h"
 
-class BroadPhaseOctree : public BroadPhase3DSW {
-	Octree<CollisionObject3DSW, true> octree;
+class BroadPhase3DBVH : public BroadPhase3DSW {
+	BVH_Manager<CollisionObject3DSW, true, 128> bvh;
 
-	static void *_pair_callback(void *, OctreeElementID, CollisionObject3DSW *, int, OctreeElementID, CollisionObject3DSW *, int);
-	static void _unpair_callback(void *, OctreeElementID, CollisionObject3DSW *, int, OctreeElementID, CollisionObject3DSW *, int, void *);
+	static void *_pair_callback(void *, uint32_t, CollisionObject3DSW *, int, uint32_t, CollisionObject3DSW *, int);
+	static void _unpair_callback(void *, uint32_t, CollisionObject3DSW *, int, uint32_t, CollisionObject3DSW *, int, void *);
 
 	PairCallback pair_callback;
 	void *pair_userdata;
@@ -47,7 +47,7 @@ class BroadPhaseOctree : public BroadPhase3DSW {
 
 public:
 	// 0 is an invalid ID
-	virtual ID create(CollisionObject3DSW *p_object, int p_subindex = 0);
+	virtual ID create(CollisionObject3DSW *p_object, int p_subindex = 0, const AABB &p_aabb = AABB(), bool p_static = false);
 	virtual void move(ID p_id, const AABB &p_aabb);
 	virtual void set_static(ID p_id, bool p_static);
 	virtual void remove(ID p_id);
@@ -66,7 +66,7 @@ public:
 	virtual void update();
 
 	static BroadPhase3DSW *_create();
-	BroadPhaseOctree();
+	BroadPhase3DBVH();
 };
 
-#endif // BROAD_PHASE_OCTREE_H
+#endif // BROAD_PHASE_3D_BVH_H
diff --git a/servers/physics_3d/broad_phase_3d_sw.h b/servers/physics_3d/broad_phase_3d_sw.h
index 283c087b96..98313cb216 100644
--- a/servers/physics_3d/broad_phase_3d_sw.h
+++ b/servers/physics_3d/broad_phase_3d_sw.h
@@ -48,7 +48,7 @@ public:
 	typedef void (*UnpairCallback)(CollisionObject3DSW *A, int p_subindex_A, CollisionObject3DSW *B, int p_subindex_B, void *p_data, void *p_userdata);
 
 	// 0 is an invalid ID
-	virtual ID create(CollisionObject3DSW *p_object_, int p_subindex = 0) = 0;
+	virtual ID create(CollisionObject3DSW *p_object_, int p_subindex = 0, const AABB &p_aabb = AABB(), bool p_static = false) = 0;
 	virtual void move(ID p_id, const AABB &p_aabb) = 0;
 	virtual void set_static(ID p_id, bool p_static) = 0;
 	virtual void remove(ID p_id) = 0;
diff --git a/servers/physics_3d/collision_object_3d_sw.cpp b/servers/physics_3d/collision_object_3d_sw.cpp
index 293a7e6606..459deb1356 100644
--- a/servers/physics_3d/collision_object_3d_sw.cpp
+++ b/servers/physics_3d/collision_object_3d_sw.cpp
@@ -146,22 +146,23 @@ void CollisionObject3DSW::_update_shapes() {
 
 	for (int i = 0; i < shapes.size(); i++) {
 		Shape &s = shapes.write[i];
-		if (s.bpid == 0) {
-			s.bpid = space->get_broadphase()->create(this, i);
-			space->get_broadphase()->set_static(s.bpid, _static);
-		}
 
 		//not quite correct, should compute the next matrix..
 		AABB shape_aabb = s.shape->get_aabb();
 		Transform xform = transform * s.xform;
 		shape_aabb = xform.xform(shape_aabb);
+		shape_aabb.grow_by((s.aabb_cache.size.x + s.aabb_cache.size.y) * 0.5 * 0.05);
 		s.aabb_cache = shape_aabb;
-		s.aabb_cache = s.aabb_cache.grow((s.aabb_cache.size.x + s.aabb_cache.size.y) * 0.5 * 0.05);
 
 		Vector3 scale = xform.get_basis().get_scale();
 		s.area_cache = s.shape->get_area() * scale.x * scale.y * scale.z;
 
-		space->get_broadphase()->move(s.bpid, s.aabb_cache);
+		if (s.bpid == 0) {
+			s.bpid = space->get_broadphase()->create(this, i, shape_aabb, _static);
+			space->get_broadphase()->set_static(s.bpid, _static);
+		}
+
+		space->get_broadphase()->move(s.bpid, shape_aabb);
 	}
 }
 
@@ -172,18 +173,19 @@ void CollisionObject3DSW::_update_shapes_with_motion(const Vector3 &p_motion) {
 
 	for (int i = 0; i < shapes.size(); i++) {
 		Shape &s = shapes.write[i];
-		if (s.bpid == 0) {
-			s.bpid = space->get_broadphase()->create(this, i);
-			space->get_broadphase()->set_static(s.bpid, _static);
-		}
 
 		//not quite correct, should compute the next matrix..
 		AABB shape_aabb = s.shape->get_aabb();
 		Transform xform = transform * s.xform;
 		shape_aabb = xform.xform(shape_aabb);
-		shape_aabb = shape_aabb.merge(AABB(shape_aabb.position + p_motion, shape_aabb.size)); //use motion
+		shape_aabb.merge_with(AABB(shape_aabb.position + p_motion, shape_aabb.size)); //use motion
 		s.aabb_cache = shape_aabb;
 
+		if (s.bpid == 0) {
+			s.bpid = space->get_broadphase()->create(this, i, shape_aabb, _static);
+			space->get_broadphase()->set_static(s.bpid, _static);
+		}
+
 		space->get_broadphase()->move(s.bpid, shape_aabb);
 	}
 }
diff --git a/servers/physics_3d/physics_server_3d_sw.cpp b/servers/physics_3d/physics_server_3d_sw.cpp
index c08e2b5794..f3eb1ae48f 100644
--- a/servers/physics_3d/physics_server_3d_sw.cpp
+++ b/servers/physics_3d/physics_server_3d_sw.cpp
@@ -30,8 +30,7 @@
 
 #include "physics_server_3d_sw.h"
 
-#include "broad_phase_3d_basic.h"
-#include "broad_phase_octree.h"
+#include "broad_phase_3d_bvh.h"
 #include "core/debugger/engine_debugger.h"
 #include "core/os/os.h"
 #include "joints/cone_twist_joint_3d_sw.h"
@@ -1755,7 +1754,8 @@ void PhysicsServer3DSW::_shape_col_cbk(const Vector3 &p_point_A, int p_index_A,
 PhysicsServer3DSW *PhysicsServer3DSW::singletonsw = nullptr;
 PhysicsServer3DSW::PhysicsServer3DSW(bool p_using_threads) {
 	singletonsw = this;
-	BroadPhase3DSW::create_func = BroadPhaseOctree::_create;
+	BroadPhase3DSW::create_func = BroadPhase3DBVH::_create;
+
 	island_count = 0;
 	active_objects = 0;
 	collision_pairs = 0;