.gitea/default_merge_message/REBASE_TEMPLATE.md

@@ -2,4 +2,4 @@ ${CommitTitle}
 
 ${CommitBody}
 
-Pull Request #${PullRequestIndex}
+Pull Request: https://projects.blender.org/blender/blender-addons/pulls/${PullRequestIndex}

.gitea/default_merge_message/SQUASH_TEMPLATE.md

@@ -1,3 +1,3 @@
 ${PullRequestTitle}
 
-Pull Request #${PullRequestIndex}
+Pull Request: https://projects.blender.org/blender/blender-addons/pulls/${PullRequestIndex}

archimesh/__init__.py

@@ -11,7 +11,7 @@ bl_info = {
     "name": "Archimesh",
     "author": "Antonio Vazquez (antonioya)",
     "location": "View3D > Add Mesh / Sidebar > Create Tab",
-    "version": (1, 2, 4),
+    "version": (1, 2, 5),
     "blender": (3, 0, 0),
     "description": "Generate rooms, doors, windows, and other architecture objects",
     "doc_url": "{BLENDER_MANUAL_URL}/addons/add_mesh/archimesh.html",

archimesh/achm_main_panel.py

@@ -108,8 +108,7 @@ class ARCHIMESH_OT_Hole(Operator):
             # ---------------------------------------
             for child in obj.parent.children:
                 # noinspection PyBroadException
-                try:
-                    if child["archimesh.ctrl_hole"]:
+                if "archimesh.ctrl_hole" in child and child["archimesh.ctrl_hole"]:
                     # apply scale
                     t = parentobj.RoomGenerator[0].wall_width
                     if t > 0:
@@ -117,11 +116,8 @@ class ARCHIMESH_OT_Hole(Operator):
                     else:
                         child.scale.y = 1
                     # add boolean modifier
-                        if isboolean(myroom, child) is False:
+                    if not isboolean(myroom, child):
                         set_modifier_boolean(myroom, child)
-                except:
-                    # print("Unexpected error:" + str(sys.exc_info()))
-                    pass
 
         # ---------------------------------------
         # Now add the modifiers to baseboard

archimesh/achm_tools.py

@@ -172,9 +172,7 @@ def set_modifier_solidify(myobject, width):
 # Add modifier (boolean)
 # --------------------------------------------------------------------
 def set_modifier_boolean(myobject, bolobject):
-    bpy.context.view_layer.objects.active = myobject
-    if bpy.context.view_layer.objects.active.name == myobject.name:
-        boolean_modifier = context.object.modifiers.new("", 'BOOLEAN')
+    boolean_modifier = myobject.modifiers.new("", 'BOOLEAN')
     boolean_modifier.operation = 'DIFFERENCE'
     boolean_modifier.object = bolobject
 

io_scene_fbx/__init__.py

@@ -2,9 +2,9 @@
 
 bl_info = {
     "name": "FBX format",
-    "author": "Campbell Barton, Bastien Montagne, Jens Restemeier",
-    "version": (4, 37, 5),
-    "blender": (3, 4, 0),
+    "author": "Campbell Barton, Bastien Montagne, Jens Restemeier, @Mysteryem",
+    "version": (5, 1, 0),
+    "blender": (3, 6, 0),
     "location": "File > Import-Export",
     "description": "FBX IO meshes, UVs, vertex colors, materials, textures, cameras, lamps and actions",
     "warning": "",
@@ -482,6 +482,12 @@ class ExportFBX(bpy.types.Operator, ExportHelper):
             description="Export vertex color attributes",
             default='SRGB',
             )
+    prioritize_active_color: BoolProperty(
+            name="Prioritize Active Color",
+            description="Make sure active color will be exported first. Could be important "
+                        "since some other software can discard other color attributes besides the first one",
+            default=False,
+            )
     use_subsurf: BoolProperty(
             name="Export Subdivision Surface",
             description="Export the last Catmull-Rom subdivision modifier as FBX subdivision "
@@ -787,6 +793,7 @@ class FBX_PT_export_geometry(bpy.types.Panel):
         #~ sub.enabled = operator.mesh_smooth_type in {'OFF'}
         sub.prop(operator, "use_tspace")
         layout.prop(operator, "colors_type")
+        layout.prop(operator, "prioritize_active_color")
 
 
 class FBX_PT_export_armature(bpy.types.Panel):

io_scene_fbx/export_fbx_bin.py

@@ -11,6 +11,7 @@ import os
 import time
 
 from itertools import zip_longest, chain
+from functools import cache
 
 if "bpy" in locals():
     import importlib
@@ -47,9 +48,10 @@ from .fbx_utils import (
     # Miscellaneous utils.
     PerfMon,
     units_blender_to_fbx_factor, units_convertor, units_convertor_iter,
-    matrix4_to_array, similar_values, similar_values_iter, astype_view_signedness, fast_first_axis_unique,
+    matrix4_to_array, similar_values, shape_difference_exclude_similar, astype_view_signedness, fast_first_axis_unique,
+    fast_first_axis_flat,
     # Mesh transform helpers.
-    vcos_transformed_gen, nors_transformed_gen, vcos_transformed, nors_transformed,
+    vcos_transformed_gen, vcos_transformed, nors_transformed,
     # UUID from key.
     get_fbx_uuid_from_key,
     # Key generators.
@@ -760,15 +762,19 @@ def fbx_data_mesh_shapes_elements(root, me_obj, me, scene_data, fbx_me_tmpl, fbx
     for shape, (channel_key, geom_key, shape_verts_co, shape_verts_idx) in shapes.items():
         # Use vgroups as weights, if defined.
         if shape.vertex_group and shape.vertex_group in me_obj.bdata.vertex_groups:
-            shape_verts_weights = array.array(data_types.ARRAY_FLOAT64, [0.0]) * (len(shape_verts_co) // 3)
+            shape_verts_weights = np.zeros(len(shape_verts_idx), dtype=np.float64)
+            # It's slightly faster to iterate and index the underlying memoryview objects
+            mv_shape_verts_weights = shape_verts_weights.data
+            mv_shape_verts_idx = shape_verts_idx.data
             vg_idx = me_obj.bdata.vertex_groups[shape.vertex_group].index
-            for sk_idx, v_idx in enumerate(shape_verts_idx):
+            for sk_idx, v_idx in enumerate(mv_shape_verts_idx):
                 for vg in vertices[v_idx].groups:
                     if vg.group == vg_idx:
-                        shape_verts_weights[sk_idx] = vg.weight * 100.0
+                        mv_shape_verts_weights[sk_idx] = vg.weight
                         break
+            shape_verts_weights *= 100.0
         else:
-            shape_verts_weights = array.array(data_types.ARRAY_FLOAT64, [100.0]) * (len(shape_verts_co) // 3)
+            shape_verts_weights = np.full(len(shape_verts_idx), 100.0, dtype=np.float64)
         channels.append((channel_key, shape, shape_verts_weights))
 
         geom = elem_data_single_int64(root, b"Geometry", get_fbx_uuid_from_key(geom_key))
@@ -784,8 +790,7 @@ def fbx_data_mesh_shapes_elements(root, me_obj, me, scene_data, fbx_me_tmpl, fbx
         elem_data_single_int32_array(geom, b"Indexes", shape_verts_idx)
         elem_data_single_float64_array(geom, b"Vertices", shape_verts_co)
         if write_normals:
-            elem_data_single_float64_array(geom, b"Normals",
-                                           array.array(data_types.ARRAY_FLOAT64, [0.0]) * len(shape_verts_co))
+            elem_data_single_float64_array(geom, b"Normals", np.zeros(len(shape_verts_idx) * 3, dtype=np.float64))
 
     # Yiha! BindPose for shapekeys too! Dodecasigh...
     # XXX Not sure yet whether several bindposes on same mesh are allowed, or not... :/
@@ -897,19 +902,60 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
     # We do loose edges as two-vertices faces, if enabled...
     #
     # Note we have to process Edges in the same time, as they are based on poly's loops...
+
+    # Total number of loops, including any extra added for loose edges.
     loop_nbr = len(me.loops)
-    t_pvi = array.array(data_types.ARRAY_INT32, (0,)) * loop_nbr
-    t_ls = [None] * len(me.polygons)
 
-    me.loops.foreach_get("vertex_index", t_pvi)
+    # dtypes matching the C data. Matching the C datatype avoids iteration and casting of every element in foreach_get's
+    # C code.
+    bl_vertex_index_dtype = bl_edge_index_dtype = bl_loop_index_dtype = np.uintc
+
+    # Start vertex indices of loops. May contain elements for loops added for the export of loose edges.
+    t_lvi = np.empty(len(me.loops), dtype=bl_vertex_index_dtype)
+
+    # Loop start indices of polygons. May contain elements for the polygons added for the export of loose edges.
+    t_ls = np.empty(len(me.polygons), dtype=bl_loop_index_dtype)
+
+    # Vertex indices of edges (unsorted, unlike Mesh.edge_keys), flattened into an array twice the length of the number
+    # of edges.
+    t_ev = np.empty(len(me.edges) * 2, dtype=bl_vertex_index_dtype)
+    # Each edge has two vertex indices, so it's useful to view the array as 2d where each element on the first axis is a
+    # pair of vertex indices
+    t_ev_pair_view = t_ev.view()
+    t_ev_pair_view.shape = (-1, 2)
+
+    # Edge indices of loops. May contain elements for loops added for the export of loose edges.
+    t_lei = np.empty(len(me.loops), dtype=bl_edge_index_dtype)
+
+    me.loops.foreach_get("vertex_index", t_lvi)
     me.polygons.foreach_get("loop_start", t_ls)
+    me.edges.foreach_get("vertices", t_ev)
+    me.loops.foreach_get("edge_index", t_lei)
 
-    # Add "fake" faces for loose edges.
+    # Add "fake" faces for loose edges. Each "fake" face consists of two loops creating a new 2-sided polygon.
     if scene_data.settings.use_mesh_edges:
-        t_le = tuple(e.vertices for e in me.edges if e.is_loose)
-        t_pvi.extend(chain(*t_le))
-        t_ls.extend(range(loop_nbr, loop_nbr + len(t_le) * 2, 2))
+        bl_edge_is_loose_dtype = bool
+        # Get the mask of edges that are loose
+        loose_mask = np.empty(len(me.edges), dtype=bl_edge_is_loose_dtype)
+        me.edges.foreach_get('is_loose', loose_mask)
+
+        indices_of_loose_edges = np.flatnonzero(loose_mask)
+        # Since we add two loops per loose edge, repeat the indices so that there's one for each new loop
+        new_loop_edge_indices = np.repeat(indices_of_loose_edges, 2)
+
+        # Get the loose edge vertex index pairs
+        t_le = t_ev_pair_view[loose_mask]
+
+        # append will automatically flatten the pairs in t_le
+        t_lvi = np.append(t_lvi, t_le)
+        t_lei = np.append(t_lei, new_loop_edge_indices)
+        # Two loops are added per loose edge
+        loop_nbr += 2 * len(t_le)
+        t_ls = np.append(t_ls, np.arange(len(me.loops), loop_nbr, 2, dtype=t_ls.dtype))
         del t_le
+        del loose_mask
+        del indices_of_loose_edges
+        del new_loop_edge_indices
 
     # Edges...
     # Note: Edges are represented as a loop here: each edge uses a single index, which refers to the polygon array.
@@ -919,83 +965,136 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
     #                 for loose edges).
     #       We also have to store a mapping from real edges to their indices in this array, for edge-mapped data
     #       (like e.g. crease).
-    t_eli = array.array(data_types.ARRAY_INT32)
-    edges_map = {}
-    edges_nbr = 0
-    if t_ls and t_pvi:
-        # t_ls is loop start indices of polygons, but we want to use it to indicate the end loop of each polygon.
-        # The loop end index of a polygon is the loop start index of the next polygon minus one, so the first element of
-        # t_ls will be ignored, and we need to add an extra element at the end to signify the end of the last polygon.
-        # If we were to add another polygon to the mesh, its loop start index would be the next loop index.
-        t_ls = set(t_ls[1:])
-        t_ls.add(loop_nbr)
-        todo_edges = [None] * len(me.edges) * 2
-        # Sigh, cannot access edge.key through foreach_get... :/
-        me.edges.foreach_get("vertices", todo_edges)
-        todo_edges = set((v1, v2) if v1 < v2 else (v2, v1) for v1, v2 in zip(*(iter(todo_edges),) * 2))
+    eli_fbx_dtype = np.int32
 
-        li = 0
-        vi = vi_start = t_pvi[0]
-        for li_next, vi_next in enumerate(t_pvi[1:] + t_pvi[:1], start=1):
-            if li_next in t_ls:  # End of a poly's loop.
-                vi2 = vi_start
-                vi_start = vi_next
-            else:
-                vi2 = vi_next
+    # Edge index of each unique edge-key, used to map per-edge data to unique edge-keys (t_pvi).
+    t_pvi_edge_indices = np.empty(0, dtype=t_lei.dtype)
 
-            e_key = (vi, vi2) if vi < vi2 else (vi2, vi)
-            if e_key in todo_edges:
-                t_eli.append(li)
-                todo_edges.remove(e_key)
-                edges_map[e_key] = edges_nbr
-                edges_nbr += 1
+    pvi_fbx_dtype = np.int32
+    if t_ls.size and t_lvi.size:
+        # Get unsorted edge keys by indexing the edge->vertex-indices array by the loop->edge-index array.
+        t_pvi_edge_keys = t_ev_pair_view[t_lei]
 
-            vi = vi_next
-            li = li_next
-    # End of edges!
+        # Sort each [edge_start_n, edge_end_n] pair to get edge keys. Heapsort seems to be the fastest for this specific
+        # use case.
+        t_pvi_edge_keys.sort(axis=1, kind='heapsort')
+
+        # Note that finding unique edge keys means that if there are multiple edges that share the same vertices (which
+        # shouldn't normally happen), only the first edge found in loops will be exported along with its per-edge data.
+        # To export separate edges that share the same vertices, fast_first_axis_unique can be replaced with np.unique
+        # with t_lei as the first argument, finding unique edges rather than unique edge keys.
+        #
+        # Since we want the unique values in their original order, the only part we care about is the indices of the
+        # first occurrence of the unique elements in t_pvi_edge_keys, so we can use our fast uniqueness helper function.
+        t_eli = fast_first_axis_unique(t_pvi_edge_keys, return_unique=False, return_index=True)
+
+        # To get the indices of the elements in t_pvi_edge_keys that produce unique values, but in the original order of
+        # t_pvi_edge_keys, t_eli must be sorted.
+        # Due to loops and their edge keys tending to have a partial ordering within meshes, sorting with kind='stable'
+        # with radix sort tends to be faster than the default of kind='quicksort' with introsort.
+        t_eli.sort(kind='stable')
+
+        # Edge index of each element in unique t_pvi_edge_keys, used to map per-edge data such as sharp and creases.
+        t_pvi_edge_indices = t_lei[t_eli]
 
         # We have to ^-1 last index of each loop.
-    for ls in t_ls:
-        t_pvi[ls - 1] ^= -1
+        # Ensure t_pvi is the correct number of bits before inverting.
+        t_pvi = astype_view_signedness(t_lvi, pvi_fbx_dtype)
+        # The index of the end of each loop is one before the index of the start of the next loop.
+        t_pvi[t_ls[1:] - 1] ^= -1
+        # The index of the end of the last loop will be the very last index.
+        t_pvi[-1] ^= -1
+        del t_pvi_edge_keys
+    else:
+        # Should be empty, but make sure it's the correct type.
+        t_pvi = np.empty(0, dtype=pvi_fbx_dtype)
+        t_eli = np.empty(0, dtype=eli_fbx_dtype)
 
     # And finally we can write data!
+    t_pvi = astype_view_signedness(t_pvi, pvi_fbx_dtype)
+    t_eli = astype_view_signedness(t_eli, eli_fbx_dtype)
     elem_data_single_int32_array(geom, b"PolygonVertexIndex", t_pvi)
     elem_data_single_int32_array(geom, b"Edges", t_eli)
+    del t_lvi
     del t_pvi
-    del t_ls
     del t_eli
+    del t_ev
+    del t_ev_pair_view
 
     # And now, layers!
 
     # Smoothing.
     if smooth_type in {'FACE', 'EDGE'}:
-        t_ps = None
+        ps_fbx_dtype = np.int32
+        poly_use_smooth_dtype = bool
+        edge_use_sharp_dtype = bool
         _map = b""
         if smooth_type == 'FACE':
-            t_ps = array.array(data_types.ARRAY_INT32, (0,)) * len(me.polygons)
+            t_ps = np.empty(len(me.polygons), dtype=poly_use_smooth_dtype)
             me.polygons.foreach_get("use_smooth", t_ps)
             _map = b"ByPolygon"
         else:  # EDGE
+            _map = b"ByEdge"
+            if t_pvi_edge_indices.size:
                 # Write Edge Smoothing.
                 # Note edge is sharp also if it's used by more than two faces, or one of its faces is flat.
-            t_ps = array.array(data_types.ARRAY_INT32, (0,)) * edges_nbr
-            sharp_edges = set()
-            temp_sharp_edges = {}
-            for p in me.polygons:
-                if not p.use_smooth:
-                    sharp_edges.update(p.edge_keys)
-                    continue
-                for k in p.edge_keys:
-                    if temp_sharp_edges.setdefault(k, 0) > 1:
-                        sharp_edges.add(k)
+                mesh_poly_nbr = len(me.polygons)
+                mesh_edge_nbr = len(me.edges)
+                mesh_loop_nbr = len(me.loops)
+                # t_ls and t_lei may contain extra polygons or loops added for loose edges that are not present in the
+                # mesh data, so create views that exclude the extra data added for loose edges.
+                mesh_t_ls_view = t_ls[:mesh_poly_nbr]
+                mesh_t_lei_view = t_lei[:mesh_loop_nbr]
+
+                # - Get sharp edges from flat shaded faces
+                # Get the 'use_smooth' attribute of all polygons.
+                p_use_smooth_mask = np.empty(mesh_poly_nbr, dtype=poly_use_smooth_dtype)
+                me.polygons.foreach_get('use_smooth', p_use_smooth_mask)
+                # Invert to get all flat shaded polygons.
+                p_flat_mask = np.invert(p_use_smooth_mask, out=p_use_smooth_mask)
+                # Convert flat shaded polygons to flat shaded loops by repeating each element by the number of sides of
+                # that polygon.
+                # Polygon sides can be calculated from the element-wise difference of loop starts appended by the number
+                # of loops. Alternatively, polygon sides can be retrieved directly from the 'loop_total' attribute of
+                # polygons, but since we already have t_ls, it tends to be quicker to calculate from t_ls when above
+                # around 10_000 polygons.
+                polygon_sides = np.diff(mesh_t_ls_view, append=mesh_loop_nbr)
+                p_flat_loop_mask = np.repeat(p_flat_mask, polygon_sides)
+                # Convert flat shaded loops to flat shaded (sharp) edge indices.
+                # Note that if an edge is in multiple loops that are part of flat shaded faces, its edge index will end
+                # up in sharp_edge_indices_from_polygons multiple times.
+                sharp_edge_indices_from_polygons = mesh_t_lei_view[p_flat_loop_mask]
+
+                # - Get sharp edges from edges marked as sharp
+                e_use_sharp_mask = np.empty(mesh_edge_nbr, dtype=edge_use_sharp_dtype)
+                me.edges.foreach_get('use_edge_sharp', e_use_sharp_mask)
+
+                # - Get sharp edges from edges used by more than two loops (and therefore more than two faces)
+                e_more_than_two_faces_mask = np.bincount(mesh_t_lei_view, minlength=mesh_edge_nbr) > 2
+
+                # - Combine with edges that are sharp because they're in more than two faces
+                e_use_sharp_mask = np.logical_or(e_use_sharp_mask, e_more_than_two_faces_mask, out=e_use_sharp_mask)
+
+                # - Combine with edges that are sharp because a polygon they're in has flat shading
+                e_use_sharp_mask[sharp_edge_indices_from_polygons] = True
+
+                # - Convert sharp edges to sharp edge keys (t_pvi)
+                ek_use_sharp_mask = e_use_sharp_mask[t_pvi_edge_indices]
+
+                # - Sharp edges are indicated in FBX as zero (False), so invert
+                t_ps = np.invert(ek_use_sharp_mask, out=ek_use_sharp_mask)
+                del ek_use_sharp_mask
+                del e_use_sharp_mask
+                del sharp_edge_indices_from_polygons
+                del p_flat_loop_mask
+                del polygon_sides
+                del p_flat_mask
+                del p_use_smooth_mask
+                del mesh_t_lei_view
+                del mesh_t_ls_view
             else:
-                        temp_sharp_edges[k] += 1
-            del temp_sharp_edges
-            for e in me.edges:
-                if e.key not in edges_map:
-                    continue  # Only loose edges, in theory!
-                t_ps[edges_map[e.key]] = not (e.use_edge_sharp or (e.key in sharp_edges))
-            _map = b"ByEdge"
+                t_ps = np.empty(0, dtype=ps_fbx_dtype)
+        t_ps = t_ps.astype(ps_fbx_dtype, copy=False)
         lay_smooth = elem_data_single_int32(geom, b"LayerElementSmoothing", 0)
         elem_data_single_int32(lay_smooth, b"Version", FBX_GEOMETRY_SMOOTHING_VERSION)
         elem_data_single_string(lay_smooth, b"Name", b"")
@@ -1003,16 +1102,29 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
         elem_data_single_string(lay_smooth, b"ReferenceInformationType", b"Direct")
         elem_data_single_int32_array(lay_smooth, b"Smoothing", t_ps)  # Sight, int32 for bool...
         del t_ps
+    del t_ls
+    del t_lei
 
     # Edge crease for subdivision
     if write_crease:
-        t_ec = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * edges_nbr
-        for e in me.edges:
-            if e.key not in edges_map:
-                continue  # Only loose edges, in theory!
+        ec_fbx_dtype = np.float64
+        if t_pvi_edge_indices.size:
+            ec_bl_dtype = np.single
+            t_ec_raw = np.empty(len(me.edges), dtype=ec_bl_dtype)
+            me.edges.foreach_get('crease', t_ec_raw)
+
+            # Convert to t_pvi edge-keys.
+            t_ec_ek_raw = t_ec_raw[t_pvi_edge_indices]
+
             # Blender squares those values before sending them to OpenSubdiv, when other software don't,
             # so we need to compensate that to get similar results through FBX...
-            t_ec[edges_map[e.key]] = e.crease * e.crease
+            # Use the precision of the fbx dtype for the calculation since it's usually higher precision.
+            t_ec_ek_raw = t_ec_ek_raw.astype(ec_fbx_dtype, copy=False)
+            t_ec = np.square(t_ec_ek_raw, out=t_ec_ek_raw)
+            del t_ec_ek_raw
+            del t_ec_raw
+        else:
+            t_ec = np.empty(0, dtype=ec_fbx_dtype)
 
         lay_crease = elem_data_single_int32(geom, b"LayerElementEdgeCrease", 0)
         elem_data_single_int32(lay_crease, b"Version", FBX_GEOMETRY_CREASE_VERSION)
@@ -1023,7 +1135,7 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
         del t_ec
 
     # And we are done with edges!
-    del edges_map
+    del t_pvi_edge_indices
 
     # Loop normals.
     tspacenumber = 0
@@ -1033,28 +1145,35 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
         #     but this does not seem well supported by apps currently...
         me.calc_normals_split()
 
-        t_ln = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(me.loops) * 3
+        ln_bl_dtype = np.single
+        ln_fbx_dtype = np.float64
+        t_ln = np.empty(len(me.loops) * 3,  dtype=ln_bl_dtype)
         me.loops.foreach_get("normal", t_ln)
-        t_ln = nors_transformed_gen(t_ln, geom_mat_no)
+        t_ln = nors_transformed(t_ln, geom_mat_no, ln_fbx_dtype)
         if 0:
-            t_ln = tuple(t_ln)  # No choice... :/
-
+            lnidx_fbx_dtype = np.int32
             lay_nor = elem_data_single_int32(geom, b"LayerElementNormal", 0)
             elem_data_single_int32(lay_nor, b"Version", FBX_GEOMETRY_NORMAL_VERSION)
             elem_data_single_string(lay_nor, b"Name", b"")
             elem_data_single_string(lay_nor, b"MappingInformationType", b"ByPolygonVertex")
             elem_data_single_string(lay_nor, b"ReferenceInformationType", b"IndexToDirect")
 
-            ln2idx = tuple(set(t_ln))
-            elem_data_single_float64_array(lay_nor, b"Normals", chain(*ln2idx))
+            # Tuple of unique sorted normals and then the index in the unique sorted normals of each normal in t_ln.
+            # Since we don't care about how the normals are sorted, only that they're unique, we can use the fast unique
+            # helper function.
+            t_ln, t_lnidx = fast_first_axis_unique(t_ln.reshape(-1, 3), return_inverse=True)
+
+            # Convert to the type for fbx
+            t_lnidx = astype_view_signedness(t_lnidx, lnidx_fbx_dtype)
+
+            elem_data_single_float64_array(lay_nor, b"Normals", t_ln)
             # Normal weights, no idea what it is.
-            # t_lnw = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(ln2idx)
+            # t_lnw = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(t_ln)
             # elem_data_single_float64_array(lay_nor, b"NormalsW", t_lnw)
 
-            ln2idx = {nor: idx for idx, nor in enumerate(ln2idx)}
-            elem_data_single_int32_array(lay_nor, b"NormalsIndex", (ln2idx[n] for n in t_ln))
+            elem_data_single_int32_array(lay_nor, b"NormalsIndex", t_lnidx)
 
-            del ln2idx
+            del t_lnidx
             # del t_lnw
         else:
             lay_nor = elem_data_single_int32(geom, b"LayerElementNormal", 0)
@@ -1062,7 +1181,7 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
             elem_data_single_string(lay_nor, b"Name", b"")
             elem_data_single_string(lay_nor, b"MappingInformationType", b"ByPolygonVertex")
             elem_data_single_string(lay_nor, b"ReferenceInformationType", b"Direct")
-            elem_data_single_float64_array(lay_nor, b"Normals", chain(*t_ln))
+            elem_data_single_float64_array(lay_nor, b"Normals", t_ln)
             # Normal weights, no idea what it is.
             # t_ln = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(me.loops)
             # elem_data_single_float64_array(lay_nor, b"NormalsW", t_ln)
@@ -1073,9 +1192,10 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
             tspacenumber = len(me.uv_layers)
             if tspacenumber:
                 # We can only compute tspace on tessellated meshes, need to check that here...
-                t_lt = [None] * len(me.polygons)
+                lt_bl_dtype = np.uintc
+                t_lt = np.empty(len(me.polygons), dtype=lt_bl_dtype)
                 me.polygons.foreach_get("loop_total", t_lt)
-                if any((lt > 4 for lt in t_lt)):
+                if (t_lt > 4).any():
                     del t_lt
                     scene_data.settings.report(
                         {'WARNING'},
@@ -1084,7 +1204,7 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
                 else:
                     del t_lt
                     num_loops = len(me.loops)
-                    t_ln = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * num_loops * 3
+                    t_ln = np.empty(num_loops * 3, dtype=ln_bl_dtype)
                     # t_lnw = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(me.loops)
                     uv_names = [uvlayer.name for uvlayer in me.uv_layers]
                     # Annoying, `me.calc_tangent` errors in case there is no geometry...
@@ -1102,7 +1222,7 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
                         elem_data_single_string(lay_nor, b"MappingInformationType", b"ByPolygonVertex")
                         elem_data_single_string(lay_nor, b"ReferenceInformationType", b"Direct")
                         elem_data_single_float64_array(lay_nor, b"Binormals",
-                                                       chain(*nors_transformed_gen(t_ln, geom_mat_no)))
+                                                       nors_transformed(t_ln, geom_mat_no, ln_fbx_dtype))
                         # Binormal weights, no idea what it is.
                         # elem_data_single_float64_array(lay_nor, b"BinormalsW", t_lnw)
 
@@ -1115,7 +1235,7 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
                         elem_data_single_string(lay_nor, b"MappingInformationType", b"ByPolygonVertex")
                         elem_data_single_string(lay_nor, b"ReferenceInformationType", b"Direct")
                         elem_data_single_float64_array(lay_nor, b"Tangents",
-                                                       chain(*nors_transformed_gen(t_ln, geom_mat_no)))
+                                                       nors_transformed(t_ln, geom_mat_no, ln_fbx_dtype))
                         # Tangent weights, no idea what it is.
                         # elem_data_single_float64_array(lay_nor, b"TangentsW", t_lnw)
 
@@ -1129,70 +1249,164 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
     colors_type = scene_data.settings.colors_type
     vcolnumber = 0 if colors_type == 'NONE' else len(me.color_attributes)
     if vcolnumber:
-        def _coltuples_gen(raw_cols):
-            return zip(*(iter(raw_cols),) * 4)
-
         color_prop_name = "color_srgb" if colors_type == 'SRGB' else "color"
+        # ByteColorAttribute color also gets returned by the API as single precision float
+        bl_lc_dtype = np.single
+        bl_lvi_dtype = np.uintc
+        fbx_lc_dtype = np.float64
+        fbx_lcidx_dtype = np.int32
+        t_lvi = None
 
-        for colindex, collayer in enumerate(me.color_attributes):
+        color_attributes = me.color_attributes
+        if scene_data.settings.prioritize_active_color:
+            active_color = me.color_attributes.active_color
+            color_attributes = sorted(color_attributes, key=lambda x: x == active_color, reverse=True)            
+
+        for colindex, collayer in enumerate(color_attributes):
             is_point = collayer.domain == "POINT"
             vcollen = len(me.vertices if is_point else me.loops)
-            t_lc = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * vcollen * 4
+            # Each rgba component is flattened in the array
+            t_lc = np.empty(vcollen * 4, dtype=bl_lc_dtype)
             collayer.data.foreach_get(color_prop_name, t_lc)
-
             lay_vcol = elem_data_single_int32(geom, b"LayerElementColor", colindex)
             elem_data_single_int32(lay_vcol, b"Version", FBX_GEOMETRY_VCOLOR_VERSION)
             elem_data_single_string_unicode(lay_vcol, b"Name", collayer.name)
             elem_data_single_string(lay_vcol, b"MappingInformationType", b"ByPolygonVertex")
             elem_data_single_string(lay_vcol, b"ReferenceInformationType", b"IndexToDirect")
 
-            col2idx = tuple(set(_coltuples_gen(t_lc)))
-            elem_data_single_float64_array(lay_vcol, b"Colors", chain(*col2idx))  # Flatten again...
+            # Use the fast uniqueness helper function since we don't care about sorting.
+            t_lc, col_indices = fast_first_axis_unique(t_lc.reshape(-1, 4), return_inverse=True)
 
-            col2idx = {col: idx for idx, col in enumerate(col2idx)}
-            col_indices = list(col2idx[c] for c in _coltuples_gen(t_lc))
             if is_point:
                 # for "point" domain colors, we could directly emit them
                 # with a "ByVertex" mapping type, but some software does not
                 # properly understand that. So expand to full "ByPolygonVertex"
                 # index map.
-                col_indices = list((col_indices[c.vertex_index] for c in me.loops))
+                if t_lvi is None:
+                    t_lvi = np.empty(len(me.loops), dtype=bl_lvi_dtype)
+                    me.loops.foreach_get("vertex_index", t_lvi)
+                col_indices = col_indices[t_lvi]
+
+            t_lc = t_lc.astype(fbx_lc_dtype, copy=False)
+            col_indices = astype_view_signedness(col_indices, fbx_lcidx_dtype)
+
+            elem_data_single_float64_array(lay_vcol, b"Colors", t_lc)
             elem_data_single_int32_array(lay_vcol, b"ColorIndex", col_indices)
-            del col2idx
+
             del t_lc
-        del _coltuples_gen
+            del col_indices
+        del t_lvi
 
     # Write UV layers.
     # Note: LayerElementTexture is deprecated since FBX 2011 - luckily!
     #       Textures are now only related to materials, in FBX!
     uvnumber = len(me.uv_layers)
     if uvnumber:
+        luv_bl_dtype = np.single
+        luv_fbx_dtype = np.float64
+        lv_idx_bl_dtype = np.uintc
+        lv_idx_fbx_dtype = np.int32
+
+        t_luv = np.empty(len(me.loops) * 2, dtype=luv_bl_dtype)
+        # Fast view for sort-based uniqueness of pairs.
+        t_luv_fast_pair_view = fast_first_axis_flat(t_luv.reshape(-1, 2))
+        # It must be a view of t_luv otherwise it won't update when t_luv is updated.
+        assert(t_luv_fast_pair_view.base is t_luv)
+
         # Looks like this mapping is also expected to convey UV islands (arg..... :((((( ).
         # So we need to generate unique triplets (uv, vertex_idx) here, not only just based on UV values.
-        def _uvtuples_gen(raw_uvs, raw_lvidxs):
-            return zip(zip(*(iter(raw_uvs),) * 2), raw_lvidxs)
-
-        t_luv = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(me.loops) * 2
-        t_lvidx = array.array(data_types.ARRAY_INT32, (0,)) * len(me.loops)
+        t_lvidx = np.empty(len(me.loops), dtype=lv_idx_bl_dtype)
         me.loops.foreach_get("vertex_index", t_lvidx)
+
+        # If we were to create a combined array of (uv, vertex_idx) elements, we could find unique triplets by sorting
+        # that array by first sorting by the vertex_idx column and then sorting by the uv column using a stable sorting
+        # algorithm.
+        # This is exactly what we'll do, but without creating the combined array, because only the uv elements are
+        # included in the export and the vertex_idx column is the same for every uv layer.
+
+        # Because the vertex_idx column is the same for every uv layer, the vertex_idx column can be sorted in advance.
+        # argsort gets the indices that sort the array, which are needed to be able to sort the array of uv pairs in the
+        # same way to create the indices that recreate the full uvs from the unique uvs.
+        # Loops and vertices tend to naturally have a partial ordering, which makes sorting with kind='stable' (radix
+        # sort) faster than the default of kind='quicksort' (introsort) in most cases.
+        perm_vidx = t_lvidx.argsort(kind='stable')
+
+        # Mask and uv indices arrays will be modified and re-used by each uv layer.
+        unique_mask = np.empty(len(me.loops), dtype=np.bool_)
+        unique_mask[:1] = True
+        uv_indices = np.empty(len(me.loops), dtype=lv_idx_fbx_dtype)
+
         for uvindex, uvlayer in enumerate(me.uv_layers):
-            uvlayer.data.foreach_get("uv", t_luv)
             lay_uv = elem_data_single_int32(geom, b"LayerElementUV", uvindex)
             elem_data_single_int32(lay_uv, b"Version", FBX_GEOMETRY_UV_VERSION)
             elem_data_single_string_unicode(lay_uv, b"Name", uvlayer.name)
             elem_data_single_string(lay_uv, b"MappingInformationType", b"ByPolygonVertex")
             elem_data_single_string(lay_uv, b"ReferenceInformationType", b"IndexToDirect")
 
-            uv_ids = tuple(set(_uvtuples_gen(t_luv, t_lvidx)))
-            elem_data_single_float64_array(lay_uv, b"UV", chain(*(uv for uv, vidx in uv_ids)))  # Flatten again...
+            uvlayer.data.foreach_get("uv", t_luv)
 
-            uv2idx = {uv_id: idx for idx, uv_id in enumerate(uv_ids)}
-            elem_data_single_int32_array(lay_uv, b"UVIndex", (uv2idx[uv_id] for uv_id in _uvtuples_gen(t_luv, t_lvidx)))
-            del uv2idx
-            del uv_ids
-        del t_luv
+            # t_luv_fast_pair_view is a view in a dtype that compares elements by individual bytes, but float types have
+            # separate byte representations of positive and negative zero. For uniqueness, these should be considered
+            # the same, so replace all -0.0 with 0.0 in advance.
+            t_luv[t_luv == -0.0] = 0.0
+
+            # These steps to create unique_uv_pairs are the same as how np.unique would find unique values by sorting a
+            # structured array where each element is a triplet of (uv, vertex_idx), except uv and vertex_idx are
+            # separate arrays here and vertex_idx has already been sorted in advance.
+
+            # Sort according to the vertex_idx column, using the precalculated indices that sort it.
+            sorted_t_luv_fast = t_luv_fast_pair_view[perm_vidx]
+
+            # Get the indices that would sort the sorted uv pairs. Stable sorting must be used to maintain the sorting
+            # of the vertex indices.
+            perm_uv_pairs = sorted_t_luv_fast.argsort(kind='stable')
+            # Use the indices to sort both the uv pairs and the vertex_idx columns.
+            perm_combined = perm_vidx[perm_uv_pairs]
+            sorted_vidx = t_lvidx[perm_combined]
+            sorted_t_luv_fast = sorted_t_luv_fast[perm_uv_pairs]
+
+            # Create a mask where either the uv pair doesn't equal the previous value in the array, or the vertex index
+            # doesn't equal the previous value, these will be the unique uv-vidx triplets.
+            # For an imaginary triplet array:
+            # ...
+            # [(0.4, 0.2), 0]
+            # [(0.4, 0.2), 1] -> Unique because vertex index different from previous
+            # [(0.4, 0.2), 2] -> Unique because vertex index different from previous
+            # [(0.7, 0.6), 2] -> Unique because uv different from previous
+            # [(0.7, 0.6), 2]
+            # ...
+            # Output the result into unique_mask.
+            np.logical_or(sorted_t_luv_fast[1:] != sorted_t_luv_fast[:-1], sorted_vidx[1:] != sorted_vidx[:-1],
+                          out=unique_mask[1:])
+
+            # Get each uv pair marked as unique by the unique_mask and then view as the original dtype.
+            unique_uvs = sorted_t_luv_fast[unique_mask].view(luv_bl_dtype)
+
+            # NaN values are considered invalid and indicate a bug somewhere else in Blender or in an addon, we want
+            # these bugs to be reported instead of hiding them by allowing the export to continue.
+            if np.isnan(unique_uvs).any():
+                raise RuntimeError("UV layer %s on %r has invalid UVs containing NaN values" % (uvlayer.name, me))
+
+            # Convert to the type needed for fbx
+            unique_uvs = unique_uvs.astype(luv_fbx_dtype, copy=False)
+
+            # Set the indices of pairs in unique_uvs that reconstruct the pairs in t_luv into uv_indices.
+            # uv_indices will then be the same as an inverse array returned by np.unique with return_inverse=True.
+            uv_indices[perm_combined] = np.cumsum(unique_mask, dtype=uv_indices.dtype) - 1
+
+            elem_data_single_float64_array(lay_uv, b"UV", unique_uvs)
+            elem_data_single_int32_array(lay_uv, b"UVIndex", uv_indices)
+            del unique_uvs
+            del sorted_t_luv_fast
+            del sorted_vidx
+            del perm_uv_pairs
+            del perm_combined
+        del uv_indices
+        del unique_mask
+        del perm_vidx
         del t_lvidx
-        del _uvtuples_gen
+        del t_luv
+        del t_luv_fast_pair_view
 
     # Face's materials.
     me_fbxmaterials_idx = scene_data.mesh_material_indices.get(me)
@@ -1205,16 +1419,27 @@ def fbx_data_mesh_elements(root, me_obj, scene_data, done_meshes):
             elem_data_single_string(lay_ma, b"Name", b"")
             nbr_mats = len(me_fbxmaterials_idx)
             if nbr_mats > 1:
-                t_pm = array.array(data_types.ARRAY_INT32, (0,)) * len(me.polygons)
+                bl_pm_dtype = np.uintc
+                fbx_pm_dtype = np.int32
+                t_pm = np.empty(len(me.polygons), dtype=bl_pm_dtype)
                 me.polygons.foreach_get("material_index", t_pm)
 
                 # We have to validate mat indices, and map them to FBX indices.
-                # Note a mat might not be in me_fbxmats_idx (e.g. node mats are ignored).
-                def_ma = next(me_fbxmaterials_idx[m] for m in me_blmaterials if m in me_fbxmaterials_idx)
-                blmaterials_to_fbxmaterials_idxs = [me_fbxmaterials_idx.get(m, def_ma) for m in me_blmaterials]
-                ma_idx_limit = len(blmaterials_to_fbxmaterials_idxs)
-                _gen = (blmaterials_to_fbxmaterials_idxs[m] if m < ma_idx_limit else def_ma for m in t_pm)
-                t_pm = array.array(data_types.ARRAY_INT32, _gen)
+                # Note a mat might not be in me_fbxmaterials_idx (e.g. node mats are ignored).
+
+                # The first valid material will be used for materials out of bounds of me_blmaterials or materials not
+                # in me_fbxmaterials_idx.
+                def_me_blmaterial_idx, def_ma = next(
+                    (i, me_fbxmaterials_idx[m]) for i, m in enumerate(me_blmaterials) if m in me_fbxmaterials_idx)
+
+                # Set material indices that are out of bounds to the default material index
+                mat_idx_limit = len(me_blmaterials)
+                t_pm[t_pm >= mat_idx_limit] = def_me_blmaterial_idx
+
+                # Map to FBX indices. Materials not in me_fbxmaterials_idx will be set to the default material index.
+                blmat_fbx_idx = np.fromiter((me_fbxmaterials_idx.get(m, def_ma) for m in me_blmaterials),
+                                            dtype=fbx_pm_dtype)
+                t_pm = blmat_fbx_idx[t_pm]
 
                 elem_data_single_string(lay_ma, b"MappingInformationType", b"ByPolygon")
                 # XXX Logically, should be "Direct" reference type, since we do not have any index array, and have one
@@ -2375,6 +2600,18 @@ def fbx_data_from_scene(scene, depsgraph, settings):
     # ShapeKeys.
     data_deformers_shape = {}
     geom_mat_co = settings.global_matrix if settings.bake_space_transform else None
+    co_bl_dtype = np.single
+    co_fbx_dtype = np.float64
+    idx_fbx_dtype = np.int32
+
+    def empty_verts_fallbacks():
+        """Create fallback arrays for when there are no verts"""
+        # FBX does not like empty shapes (makes Unity crash e.g.).
+        # To prevent this, we add a vertex that does nothing, but it keeps the shape key intact
+        single_vert_co = np.zeros((1, 3), dtype=co_fbx_dtype)
+        single_vert_idx = np.zeros(1, dtype=idx_fbx_dtype)
+        return single_vert_co, single_vert_idx
+
     for me_key, me, _free in data_meshes.values():
         if not (me.shape_keys and len(me.shape_keys.key_blocks) > 1):  # We do not want basis-only relative skeys...
             continue
@@ -2382,41 +2619,44 @@ def fbx_data_from_scene(scene, depsgraph, settings):
             continue
 
         shapes_key = get_blender_mesh_shape_key(me)
-        # We gather all vcos first, since some skeys may be based on others...
-        _cos = array.array(data_types.ARRAY_FLOAT64, (0.0,)) * len(me.vertices) * 3
-        me.vertices.foreach_get("co", _cos)
-        v_cos = tuple(vcos_transformed_gen(_cos, geom_mat_co))
-        sk_cos = {}
-        for shape in me.shape_keys.key_blocks[1:]:
-            shape.data.foreach_get("co", _cos)
-            sk_cos[shape] = tuple(vcos_transformed_gen(_cos, geom_mat_co))
+
         sk_base = me.shape_keys.key_blocks[0]
 
+        # Get and cache only the cos that we need
+        @cache
+        def sk_cos(shape_key):
+            _cos = np.empty(len(me.vertices) * 3, dtype=co_bl_dtype)
+            if shape_key == sk_base:
+                me.vertices.foreach_get("co", _cos)
+            else:
+                shape_key.data.foreach_get("co", _cos)
+            return vcos_transformed(_cos, geom_mat_co, co_fbx_dtype)
+
         for shape in me.shape_keys.key_blocks[1:]:
-            # Only write vertices really different from org coordinates!
-            shape_verts_co = []
-            shape_verts_idx = []
+            # Only write vertices really different from base coordinates!
+            relative_key = shape.relative_key
+            if shape == relative_key:
+                # Shape is its own relative key, so it does nothing
+                shape_verts_co, shape_verts_idx = empty_verts_fallbacks()
+            else:
+                sv_cos = sk_cos(shape)
+                ref_cos = sk_cos(shape.relative_key)
 
-            sv_cos = sk_cos[shape]
-            ref_cos = v_cos if shape.relative_key == sk_base else sk_cos[shape.relative_key]
-            for idx, (sv_co, ref_co) in enumerate(zip(sv_cos, ref_cos)):
-                if similar_values_iter(sv_co, ref_co):
-                    # Note: Maybe this is a bit too simplistic, should we use real shape base here? Though FBX does not
-                    #       have this at all... Anyway, this should cover most common cases imho.
-                    continue
-                shape_verts_co.extend(Vector(sv_co) - Vector(ref_co))
-                shape_verts_idx.append(idx)
+                # Exclude cos similar to ref_cos and get the indices of the cos that remain
+                shape_verts_co, shape_verts_idx = shape_difference_exclude_similar(sv_cos, ref_cos)
 
-            # FBX does not like empty shapes (makes Unity crash e.g.).
-            # To prevent this, we add a vertex that does nothing, but it keeps the shape key intact
-            if not shape_verts_co:
-                shape_verts_co.extend((0, 0, 0))
-                shape_verts_idx.append(0)
+                if not shape_verts_co.size:
+                    shape_verts_co, shape_verts_idx = empty_verts_fallbacks()
+                else:
+                    # Ensure the indices are of the correct type
+                    shape_verts_idx = astype_view_signedness(shape_verts_idx, idx_fbx_dtype)
 
             channel_key, geom_key = get_blender_mesh_shape_channel_key(me, shape)
             data = (channel_key, geom_key, shape_verts_co, shape_verts_idx)
             data_deformers_shape.setdefault(me, (me_key, shapes_key, {}))[2][shape] = data
 
+        del sk_cos
+
     perfmon.step("FBX export prepare: Wrapping Armatures...")
 
     # Armatures!
@@ -3055,6 +3295,7 @@ def save_single(operator, scene, depsgraph, filepath="",
                 bake_space_transform=False,
                 armature_nodetype='NULL',
                 colors_type='SRGB',
+                prioritize_active_color=False,
                 **kwargs
                 ):
 
@@ -3122,7 +3363,7 @@ def save_single(operator, scene, depsgraph, filepath="",
         add_leaf_bones, bone_correction_matrix, bone_correction_matrix_inv,
         bake_anim, bake_anim_use_all_bones, bake_anim_use_nla_strips, bake_anim_use_all_actions,
         bake_anim_step, bake_anim_simplify_factor, bake_anim_force_startend_keying,
-        False, media_settings, use_custom_props, colors_type,
+        False, media_settings, use_custom_props, colors_type, prioritize_active_color
     )
 
     import bpy_extras.io_utils

io_scene_fbx/fbx_utils.py

@@ -260,18 +260,46 @@ def similar_values_iter(v1, v2, e=1e-6):
             return False
     return True
 
+
+def shape_difference_exclude_similar(sv_cos, ref_cos, e=1e-6):
+    """Return a tuple of:
+        the difference between the vertex cos in sv_cos and ref_cos, excluding any that are nearly the same,
+        and the indices of the vertices that are not nearly the same"""
+    assert(sv_cos.size == ref_cos.size)
+
+    # Create views of 1 co per row of the arrays, only making copies if needed.
+    sv_cos = sv_cos.reshape(-1, 3)
+    ref_cos = ref_cos.reshape(-1, 3)
+
+    # Quick check for equality
+    if np.array_equal(sv_cos, ref_cos):
+        # There's no difference between the two arrays.
+        empty_cos = np.empty((0, 3), dtype=sv_cos.dtype)
+        empty_indices = np.empty(0, dtype=np.int32)
+        return empty_cos, empty_indices
+
+    # Note that unlike math.isclose(a,b), np.isclose(a,b) is not symmetrical and the second argument 'b', is
+    # considered to be the reference value.
+    # Note that atol=0 will mean that if only one co component being compared is zero, they won't be considered close.
+    similar_mask = np.isclose(sv_cos, ref_cos, atol=0, rtol=e)
+
+    # A co is only similar if every component in it is similar.
+    co_similar_mask = np.all(similar_mask, axis=1)
+
+    # Get the indices of cos that are not similar.
+    not_similar_verts_idx = np.flatnonzero(~co_similar_mask)
+
+    # Subtracting first over the entire arrays and then indexing seems faster than indexing both arrays first and then
+    # subtracting, until less than about 3% of the cos are being indexed.
+    difference_cos = (sv_cos - ref_cos)[not_similar_verts_idx]
+    return difference_cos, not_similar_verts_idx
+
+
 def vcos_transformed_gen(raw_cos, m=None):
     # Note: we could most likely get much better performances with numpy, but will leave this as TODO for now.
     gen = zip(*(iter(raw_cos),) * 3)
     return gen if m is None else (m @ Vector(v) for v in gen)
 
-def nors_transformed_gen(raw_nors, m=None):
-    # Great, now normals are also expected 4D!
-    # XXX Back to 3D normals for now!
-    # gen = zip(*(iter(raw_nors),) * 3 + (_infinite_gen(1.0),))
-    gen = zip(*(iter(raw_nors),) * 3)
-    return gen if m is None else (m @ Vector(v) for v in gen)
-
 
 def _mat4_vec3_array_multiply(mat4, vec3_array, dtype=None, return_4d=False):
     """Multiply a 4d matrix by each 3d vector in an array and return as an array of either 3d or 4d vectors.
@@ -1461,7 +1489,7 @@ FBXExportSettings = namedtuple("FBXExportSettings", (
     "bone_correction_matrix", "bone_correction_matrix_inv",
     "bake_anim", "bake_anim_use_all_bones", "bake_anim_use_nla_strips", "bake_anim_use_all_actions",
     "bake_anim_step", "bake_anim_simplify_factor", "bake_anim_force_startend_keying",
-    "use_metadata", "media_settings", "use_custom_props", "colors_type",
+    "use_metadata", "media_settings", "use_custom_props", "colors_type", "prioritize_active_color"
 ))
 
 # Helper container gathering some data we need multiple times:

io_scene_gltf2/__init__.py

@@ -4,7 +4,7 @@
 bl_info = {
     'name': 'glTF 2.0 format',
     'author': 'Julien Duroure, Scurest, Norbert Nopper, Urs Hanselmann, Moritz Becher, Benjamin Schmithüsen, Jim Eckerlein, and many external contributors',
-    "version": (3, 6, 4),
+    "version": (3, 6, 5),
     'blender': (3, 5, 0),
     'location': 'File > Import-Export',
     'description': 'Import-Export as glTF 2.0',

io_scene_gltf2/blender/exp/animation/sampled/object/gltf2_blender_gather_object_sampler.py

@@ -137,5 +137,5 @@ def __convert_keyframes(obj_uuid: str, channel: str, keyframes, action_name: str
     return input, output
 
 def __gather_interpolation(export_settings):
-    # TODO: check if the bone was animated with CONSTANT
+    # TODO: check if the object was animated with CONSTANT
     return 'LINEAR'

io_scene_gltf2/blender/exp/animation/sampled/shapekeys/gltf2_blender_gather_sk_sampler.py

@@ -103,5 +103,5 @@ def __convert_keyframes(obj_uuid, keyframes, action_name: str, export_settings):
     return input, output
 
 def __gather_interpolation(export_settings):
-    # TODO: check if the bone was animated with CONSTANT
+    # TODO: check if the SK was animated with CONSTANT
     return 'LINEAR'

io_scene_gltf2/blender/exp/gltf2_blender_gather_primitive_attributes.py

@@ -5,6 +5,7 @@ import numpy as np
 
 from ...io.com import gltf2_io, gltf2_io_constants, gltf2_io_debug
 from ...io.exp import gltf2_io_binary_data
+from ...io.exp.gltf2_io_user_extensions import export_user_extensions
 
 
 def gather_primitive_attributes(blender_primitive, export_settings):
@@ -151,6 +152,8 @@ def __gather_attribute(blender_primitive, attribute, export_settings):
         data['data'] += 0.5  # bias for rounding
         data['data'] = data['data'].astype(np.uint16)
 
+        export_user_extensions('gather_attribute_change', export_settings, attribute, data, True)
+
         return { attribute : gltf2_io.Accessor(
                 buffer_view=gltf2_io_binary_data.BinaryData(data['data'].tobytes(), gltf2_io_constants.BufferViewTarget.ARRAY_BUFFER),
                 byte_offset=None,
@@ -171,6 +174,9 @@ def __gather_attribute(blender_primitive, attribute, export_settings):
         return __gather_skins(blender_primitive, export_settings)
 
     else:
+
+        export_user_extensions('gather_attribute_change', export_settings, attribute, data, False)
+
         return {
             attribute: array_to_accessor(
                 data['data'],