Add motion vector support for animated surfaces like skeletons or blend shapes.

Extends mesh instances that required custom vertex buffers to create two alternating buffers that are written to and binds them to use them as the previous vertex buffer when generating motion vectors.

2 files changed, 127 insertions, 42 deletions

diff --git a/servers/rendering/renderer_rd/storage_rd/mesh_storage.cpp b/servers/rendering/renderer_rd/storage_rd/mesh_storage.cpp
index 56f2ea0b0c..439d0702f5 100644
--- a/servers/rendering/renderer_rd/storage_rd/mesh_storage.cpp
+++ b/servers/rendering/renderer_rd/storage_rd/mesh_storage.cpp
@@ -29,7 +29,6 @@
 /**************************************************************************/
 
 #include "mesh_storage.h"
-#include "../../rendering_server_globals.h"
 
 using namespace RendererRD;
 
@@ -854,8 +853,11 @@ void MeshStorage::_mesh_instance_clear(MeshInstance *mi) {
 			}
 			memfree(surface.versions);
 		}
-		if (surface.vertex_buffer.is_valid()) {
-			RD::get_singleton()->free(surface.vertex_buffer);
+
+		for (uint32_t i = 0; i < 2; i++) {
+			if (surface.vertex_buffer[i].is_valid()) {
+				RD::get_singleton()->free(surface.vertex_buffer[i]);
+			}
 		}
 	}
 	mi->surfaces.clear();
@@ -881,35 +883,38 @@ void MeshStorage::_mesh_instance_add_surface(MeshInstance *mi, Mesh *mesh, uint3
 
 	MeshInstance::Surface s;
 	if ((mesh->blend_shape_count > 0 || (mesh->surfaces[p_surface]->format & RS::ARRAY_FORMAT_BONES)) && mesh->surfaces[p_surface]->vertex_buffer_size > 0) {
-		//surface warrants transform
-		s.vertex_buffer = RD::get_singleton()->vertex_buffer_create(mesh->surfaces[p_surface]->vertex_buffer_size, Vector<uint8_t>(), true);
-
-		Vector<RD::Uniform> uniforms;
-		{
-			RD::Uniform u;
-			u.binding = 1;
-			u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
-			u.append_id(s.vertex_buffer);
-			uniforms.push_back(u);
-		}
-		{
-			RD::Uniform u;
-			u.binding = 2;
-			u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
-			if (mi->blend_weights_buffer.is_valid()) {
-				u.append_id(mi->blend_weights_buffer);
-			} else {
-				u.append_id(default_rd_storage_buffer);
-			}
-			uniforms.push_back(u);
-		}
-		s.uniform_set = RD::get_singleton()->uniform_set_create(uniforms, skeleton_shader.version_shader[0], SkeletonShader::UNIFORM_SET_INSTANCE);
+		_mesh_instance_add_surface_buffer(mi, mesh, &s, p_surface, 0);
 	}
 
 	mi->surfaces.push_back(s);
 	mi->dirty = true;
 }
 
+void MeshStorage::_mesh_instance_add_surface_buffer(MeshInstance *mi, Mesh *mesh, MeshInstance::Surface *s, uint32_t p_surface, uint32_t p_buffer_index) {
+	s->vertex_buffer[p_buffer_index] = RD::get_singleton()->vertex_buffer_create(mesh->surfaces[p_surface]->vertex_buffer_size, Vector<uint8_t>(), true);
+
+	Vector<RD::Uniform> uniforms;
+	{
+		RD::Uniform u;
+		u.binding = 1;
+		u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
+		u.append_id(s->vertex_buffer[p_buffer_index]);
+		uniforms.push_back(u);
+	}
+	{
+		RD::Uniform u;
+		u.binding = 2;
+		u.uniform_type = RD::UNIFORM_TYPE_STORAGE_BUFFER;
+		if (mi->blend_weights_buffer.is_valid()) {
+			u.append_id(mi->blend_weights_buffer);
+		} else {
+			u.append_id(default_rd_storage_buffer);
+		}
+		uniforms.push_back(u);
+	}
+	s->uniform_set[p_buffer_index] = RD::get_singleton()->uniform_set_create(uniforms, skeleton_shader.version_shader[0], SkeletonShader::UNIFORM_SET_INSTANCE);
+}
+
 void MeshStorage::mesh_instance_check_for_update(RID p_mesh_instance) {
 	MeshInstance *mi = mesh_instance_owner.get_or_null(p_mesh_instance);
 
@@ -956,6 +961,8 @@ void MeshStorage::update_mesh_instances() {
 	}
 
 	//process skeletons and blend shapes
+	uint64_t frame = RSG::rasterizer->get_frame_number();
+	bool uses_motion_vectors = (RSG::viewport->get_num_viewports_with_motion_vectors() > 0);
 	RD::ComputeListID compute_list = RD::get_singleton()->compute_list_begin();
 
 	while (dirty_mesh_instance_arrays.first()) {
@@ -964,7 +971,29 @@ void MeshStorage::update_mesh_instances() {
 		Skeleton *sk = skeleton_owner.get_or_null(mi->skeleton);
 
 		for (uint32_t i = 0; i < mi->surfaces.size(); i++) {
-			if (mi->surfaces[i].uniform_set == RID() || mi->mesh->surfaces[i]->uniform_set == RID()) {
+			if (mi->surfaces[i].uniform_set[0].is_null() || mi->mesh->surfaces[i]->uniform_set.is_null()) {
+				// Skip over mesh instances that don't require their own uniform buffers.
+				continue;
+			}
+
+			mi->surfaces[i].previous_buffer = mi->surfaces[i].current_buffer;
+
+			if (uses_motion_vectors && (frame - mi->surfaces[i].last_change) == 1) {
+				// Previous buffer's data can only be one frame old to be able to use motion vectors.
+				uint32_t new_buffer_index = mi->surfaces[i].current_buffer ^ 1;
+
+				if (mi->surfaces[i].uniform_set[new_buffer_index].is_null()) {
+					// Create the new vertex buffer on demand where the result for the current frame will be stored.
+					_mesh_instance_add_surface_buffer(mi, mi->mesh, &mi->surfaces[i], i, new_buffer_index);
+				}
+
+				mi->surfaces[i].current_buffer = new_buffer_index;
+			}
+
+			mi->surfaces[i].last_change = frame;
+
+			RID mi_surface_uniform_set = mi->surfaces[i].uniform_set[mi->surfaces[i].current_buffer];
+			if (mi_surface_uniform_set.is_null()) {
 				continue;
 			}
 
@@ -972,7 +1001,7 @@ void MeshStorage::update_mesh_instances() {
 
 			RD::get_singleton()->compute_list_bind_compute_pipeline(compute_list, skeleton_shader.pipeline[array_is_2d ? SkeletonShader::SHADER_MODE_2D : SkeletonShader::SHADER_MODE_3D]);
 
-			RD::get_singleton()->compute_list_bind_uniform_set(compute_list, mi->surfaces[i].uniform_set, SkeletonShader::UNIFORM_SET_INSTANCE);
+			RD::get_singleton()->compute_list_bind_uniform_set(compute_list, mi_surface_uniform_set, SkeletonShader::UNIFORM_SET_INSTANCE);
 			RD::get_singleton()->compute_list_bind_uniform_set(compute_list, mi->mesh->surfaces[i]->uniform_set, SkeletonShader::UNIFORM_SET_SURFACE);
 			if (sk && sk->uniform_set_mi.is_valid()) {
 				RD::get_singleton()->compute_list_bind_uniform_set(compute_list, sk->uniform_set_mi, SkeletonShader::UNIFORM_SET_SKELETON);
@@ -1032,7 +1061,7 @@ void MeshStorage::update_mesh_instances() {
 	RD::get_singleton()->compute_list_end();
 }
 
-void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::Version &v, Mesh::Surface *s, uint32_t p_input_mask, MeshInstance::Surface *mis) {
+void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::Version &v, Mesh::Surface *s, uint32_t p_input_mask, bool p_input_motion_vectors, MeshInstance::Surface *mis) {
 	Vector<RD::VertexAttribute> attributes;
 	Vector<RID> buffers;
 
@@ -1105,7 +1134,7 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 					}
 
 					if (mis) {
-						buffer = mis->vertex_buffer;
+						buffer = mis->vertex_buffer[mis->current_buffer];
 					} else {
 						buffer = s->vertex_buffer;
 					}
@@ -1117,7 +1146,7 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 					stride += sizeof(uint16_t) * 2;
 
 					if (mis) {
-						buffer = mis->vertex_buffer;
+						buffer = mis->vertex_buffer[mis->current_buffer];
 					} else {
 						buffer = s->vertex_buffer;
 					}
@@ -1128,7 +1157,7 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 					stride += sizeof(uint16_t) * 2;
 
 					if (mis) {
-						buffer = mis->vertex_buffer;
+						buffer = mis->vertex_buffer[mis->current_buffer];
 					} else {
 						buffer = s->vertex_buffer;
 					}
@@ -1193,6 +1222,32 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 
 		attributes.push_back(vd);
 		buffers.push_back(buffer);
+
+		if (p_input_motion_vectors) {
+			// Since the previous vertex, normal and tangent can't be part of the vertex format but they are required when motion
+			// vectors are enabled, we opt to push a copy of the vertex attribute with a different location and buffer (if it's
+			// part of an instance that has one).
+			switch (i) {
+				case RS::ARRAY_VERTEX: {
+					vd.location = ATTRIBUTE_LOCATION_PREV_VERTEX;
+				} break;
+				case RS::ARRAY_NORMAL: {
+					vd.location = ATTRIBUTE_LOCATION_PREV_NORMAL;
+				} break;
+				case RS::ARRAY_TANGENT: {
+					vd.location = ATTRIBUTE_LOCATION_PREV_TANGENT;
+				} break;
+			}
+
+			if (int(vd.location) != i) {
+				if (mis && buffer != mesh_default_rd_buffers[i]) {
+					buffer = mis->vertex_buffer[mis->previous_buffer];
+				}
+
+				attributes.push_back(vd);
+				buffers.push_back(buffer);
+			}
+		}
 	}
 
 	//update final stride
@@ -1202,7 +1257,7 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 		}
 		int loc = attributes[i].location;
 
-		if (loc < RS::ARRAY_COLOR) {
+		if ((loc < RS::ARRAY_COLOR) || ((loc >= ATTRIBUTE_LOCATION_PREV_VERTEX) && (loc <= ATTRIBUTE_LOCATION_PREV_TANGENT))) {
 			attributes.write[i].stride = stride;
 		} else if (loc < RS::ARRAY_BONES) {
 			attributes.write[i].stride = attribute_stride;
@@ -1212,6 +1267,9 @@ void MeshStorage::_mesh_surface_generate_version_for_input_mask(Mesh::Surface::V
 	}
 
 	v.input_mask = p_input_mask;
+	v.current_buffer = mis ? mis->current_buffer : 0;
+	v.previous_buffer = mis ? mis->previous_buffer : 0;
+	v.input_motion_vectors = p_input_motion_vectors;
 	v.vertex_format = RD::get_singleton()->vertex_format_create(attributes);
 	v.vertex_array = RD::get_singleton()->vertex_array_create(s->vertex_count, v.vertex_format, buffers);
 }
diff --git a/servers/rendering/renderer_rd/storage_rd/mesh_storage.h b/servers/rendering/renderer_rd/storage_rd/mesh_storage.h
index c23a5b1449..99ba69f98a 100644
--- a/servers/rendering/renderer_rd/storage_rd/mesh_storage.h
+++ b/servers/rendering/renderer_rd/storage_rd/mesh_storage.h
@@ -31,6 +31,7 @@
 #ifndef MESH_STORAGE_RD_H
 #define MESH_STORAGE_RD_H
 
+#include "../../rendering_server_globals.h"
 #include "core/templates/local_vector.h"
 #include "core/templates/rid_owner.h"
 #include "core/templates/self_list.h"
@@ -90,6 +91,9 @@ private:
 
 			struct Version {
 				uint32_t input_mask = 0;
+				uint32_t current_buffer = 0;
+				uint32_t previous_buffer = 0;
+				bool input_motion_vectors = false;
 				RD::VertexFormatID vertex_format = 0;
 				RID vertex_array;
 			};
@@ -162,8 +166,11 @@ private:
 		Mesh *mesh = nullptr;
 		RID skeleton;
 		struct Surface {
-			RID vertex_buffer;
-			RID uniform_set;
+			RID vertex_buffer[2];
+			RID uniform_set[2];
+			uint32_t current_buffer = 0;
+			uint32_t previous_buffer = 0;
+			uint64_t last_change = 0;
 
 			Mesh::Surface::Version *versions = nullptr; //allocated on demand
 			uint32_t version_count = 0;
@@ -183,10 +190,11 @@ private:
 				weight_update_list(this), array_update_list(this) {}
 	};
 
-	void _mesh_surface_generate_version_for_input_mask(Mesh::Surface::Version &v, Mesh::Surface *s, uint32_t p_input_mask, MeshInstance::Surface *mis = nullptr);
+	void _mesh_surface_generate_version_for_input_mask(Mesh::Surface::Version &v, Mesh::Surface *s, uint32_t p_input_mask, bool p_input_motion_vectors, MeshInstance::Surface *mis = nullptr);
 
 	void _mesh_instance_clear(MeshInstance *mi);
 	void _mesh_instance_add_surface(MeshInstance *mi, Mesh *mesh, uint32_t p_surface);
+	void _mesh_instance_add_surface_buffer(MeshInstance *mi, Mesh *mesh, MeshInstance::Surface *s, uint32_t p_surface, uint32_t p_buffer_index);
 
 	mutable RID_Owner<MeshInstance> mesh_instance_owner;
 
@@ -311,6 +319,12 @@ private:
 
 	Skeleton *skeleton_dirty_list = nullptr;
 
+	enum AttributeLocation {
+		ATTRIBUTE_LOCATION_PREV_VERTEX = 12,
+		ATTRIBUTE_LOCATION_PREV_NORMAL = 13,
+		ATTRIBUTE_LOCATION_PREV_TANGENT = 14
+	};
+
 public:
 	static MeshStorage *get_singleton();
 
@@ -437,7 +451,7 @@ public:
 		}
 	}
 
-	_FORCE_INLINE_ void mesh_surface_get_vertex_arrays_and_format(void *p_surface, uint32_t p_input_mask, RID &r_vertex_array_rd, RD::VertexFormatID &r_vertex_format) {
+	_FORCE_INLINE_ void mesh_surface_get_vertex_arrays_and_format(void *p_surface, uint32_t p_input_mask, bool p_input_motion_vectors, RID &r_vertex_array_rd, RD::VertexFormatID &r_vertex_format) {
 		Mesh::Surface *s = reinterpret_cast<Mesh::Surface *>(p_surface);
 
 		s->version_lock.lock();
@@ -445,9 +459,11 @@ public:
 		//there will never be more than, at much, 3 or 4 versions, so iterating is the fastest way
 
 		for (uint32_t i = 0; i < s->version_count; i++) {
-			if (s->versions[i].input_mask != p_input_mask) {
+			if (s->versions[i].input_mask != p_input_mask || s->versions[i].input_motion_vectors != p_input_motion_vectors) {
+				// Find the version that matches the inputs required.
 				continue;
 			}
+
 			//we have this version, hooray
 			r_vertex_format = s->versions[i].vertex_format;
 			r_vertex_array_rd = s->versions[i].vertex_array;
@@ -459,7 +475,7 @@ public:
 		s->version_count++;
 		s->versions = (Mesh::Surface::Version *)memrealloc(s->versions, sizeof(Mesh::Surface::Version) * s->version_count);
 
-		_mesh_surface_generate_version_for_input_mask(s->versions[version], s, p_input_mask);
+		_mesh_surface_generate_version_for_input_mask(s->versions[version], s, p_input_mask, p_input_motion_vectors);
 
 		r_vertex_format = s->versions[version].vertex_format;
 		r_vertex_array_rd = s->versions[version].vertex_array;
@@ -467,7 +483,7 @@ public:
 		s->version_lock.unlock();
 	}
 
-	_FORCE_INLINE_ void mesh_instance_surface_get_vertex_arrays_and_format(RID p_mesh_instance, uint32_t p_surface_index, uint32_t p_input_mask, RID &r_vertex_array_rd, RD::VertexFormatID &r_vertex_format) {
+	_FORCE_INLINE_ void mesh_instance_surface_get_vertex_arrays_and_format(RID p_mesh_instance, uint32_t p_surface_index, uint32_t p_input_mask, bool p_input_motion_vectors, RID &r_vertex_array_rd, RD::VertexFormatID &r_vertex_format) {
 		MeshInstance *mi = mesh_instance_owner.get_or_null(p_mesh_instance);
 		ERR_FAIL_COND(!mi);
 		Mesh *mesh = mi->mesh;
@@ -475,15 +491,26 @@ public:
 
 		MeshInstance::Surface *mis = &mi->surfaces[p_surface_index];
 		Mesh::Surface *s = mesh->surfaces[p_surface_index];
+		uint32_t current_buffer = mis->current_buffer;
+
+		// Using the previous buffer is only allowed if the surface was updated this frame and motion vectors are required.
+		uint32_t previous_buffer = p_input_motion_vectors && (RSG::rasterizer->get_frame_number() == mis->last_change) ? mis->previous_buffer : current_buffer;
 
 		s->version_lock.lock();
 
 		//there will never be more than, at much, 3 or 4 versions, so iterating is the fastest way
 
 		for (uint32_t i = 0; i < mis->version_count; i++) {
-			if (mis->versions[i].input_mask != p_input_mask) {
+			if (mis->versions[i].input_mask != p_input_mask || mis->versions[i].input_motion_vectors != p_input_motion_vectors) {
+				// Find the version that matches the inputs required.
 				continue;
 			}
+
+			if (mis->versions[i].current_buffer != current_buffer || mis->versions[i].previous_buffer != previous_buffer) {
+				// Find the version that corresponds to the correct buffers that should be used.
+				continue;
+			}
+
 			//we have this version, hooray
 			r_vertex_format = mis->versions[i].vertex_format;
 			r_vertex_array_rd = mis->versions[i].vertex_array;
@@ -495,7 +522,7 @@ public:
 		mis->version_count++;
 		mis->versions = (Mesh::Surface::Version *)memrealloc(mis->versions, sizeof(Mesh::Surface::Version) * mis->version_count);
 
-		_mesh_surface_generate_version_for_input_mask(mis->versions[version], s, p_input_mask, mis);
+		_mesh_surface_generate_version_for_input_mask(mis->versions[version], s, p_input_mask, p_input_motion_vectors, mis);
 
 		r_vertex_format = mis->versions[version].vertex_format;
 		r_vertex_array_rd = mis->versions[version].vertex_array;