Better shape reuse identification

src/picosvg/geometric_types.py

@@ -15,10 +15,15 @@
 import math
 from typing import NamedTuple, Optional, Union
 
-_DEFAULT_ALMOST_EQUAL_TOLERENCE = 1e-9
+
+_DEFAULT_ALMOST_EQUAL_TOLERANCE = 1e-9
 _PointOrVec = Union["Point", "Vector"]
 
 
+def almost_equal(c1, c2, tolerance=_DEFAULT_ALMOST_EQUAL_TOLERANCE) -> bool:
+    return abs(c1 - c2) <= tolerance
+
+
 class Point(NamedTuple):
     x: float = 0
     y: float = 0
@@ -51,9 +56,11 @@ def round(self, digits: int) -> "Point":
         return Point(round(self.x, digits), round(self.y, digits))
 
     def almost_equals(
-        self, other: "Point", tolerence=_DEFAULT_ALMOST_EQUAL_TOLERENCE
+        self, other: "Point", tolerance=_DEFAULT_ALMOST_EQUAL_TOLERANCE
     ) -> bool:
-        return abs(self.x - other.x) <= tolerence and abs(self.y - other.y) <= tolerence
+        return almost_equal(self.x, other.x, tolerance) and almost_equal(
+            self.y, other.y, tolerance
+        )
 
 
 class Vector(NamedTuple):

src/picosvg/svg.py

@@ -765,9 +765,10 @@ def _apply_gradient_translation(self, inplace=False):
                 x_prime = (r1 - c * y_prime) / a
 
                 # sanity check: a`(x`, y`) should be a(x, y)
-                p = affine.map_point((x, y))
+                # all our float brutality damages points; low tolerance sanity checks!
+                p = Point(r1, r2)
                 p_prime = affine_prime.map_point((x_prime, y_prime))
-                assert p.almost_equals(p_prime)
+                assert p.almost_equals(p_prime, tolerance=1e-1), f"{p} != {p_prime}"
 
                 el.attrib[x_attr] = ntos(round(x_prime, _GRADIENT_TRANSFORM_NDIGITS))
                 el.attrib[y_attr] = ntos(round(y_prime, _GRADIENT_TRANSFORM_NDIGITS))

src/picosvg/svg_meta.py

@@ -135,7 +135,7 @@ def parse_css_declarations(
     output: MutableMapping[str, Any],
     property_names: Optional[Container[str]] = None,
 ) -> str:
-    """ Parse CSS declaration list into {property: value} dict.
+    """Parse CSS declaration list into {property: value} dict.
 
     Args:
         style: CSS declaration list without the enclosing braces,

src/picosvg/svg_path_iter.py

@@ -35,11 +35,11 @@ def _explode_cmd(args_per_cmd, cmd, args):
 def parse_svg_path(svg_path: str, exploded=False):
     """Parses an svg path.
 
-  Exploded means when params repeat each the command is reported as
-  if multiplied. For example "M1,1 2,2 3,3" would report as three
-  separate steps when exploded.
+    Exploded means when params repeat each the command is reported as
+    if multiplied. For example "M1,1 2,2 3,3" would report as three
+    separate steps when exploded.
 
-  Yields tuples of (cmd, (args))."""
+    Yields tuples of (cmd, (args))."""
     command_tuples = []
     parts = _CMD_RE.split(svg_path)[1:]
     for i in range(0, len(parts), 2):

src/picosvg/svg_reuse.py

@@ -14,59 +14,233 @@
 
 # Functions meant to help discern if shapes are the same or not.
 
+import copy
 import dataclasses
+from itertools import islice
+from math import atan2
+from picosvg.geometric_types import Vector, almost_equal
 from picosvg.svg_types import SVGShape, SVGPath
-from typing import Optional, Tuple
+from typing import Generator, Iterable, Optional, Tuple
+from picosvg import svg_meta
 from picosvg.svg_transform import Affine2D
 
 
-# Number of decimal digits to round floats when normalizing or comparing
-DEFAULT_TOLERANCE = 9
+_SIGNIFICANCE_FACTOR = 5  # Must be at least N x tolerance to be significant
+_ROUND_RANGE = range(3, 13)  # range of rounds to try
 
 
-def _first_move(shape: SVGShape) -> Tuple[float, float]:
-    cmd, args = next(iter(shape.as_path()))
+def _first_move(path: SVGPath) -> Tuple[float, float]:
+    cmd, args = next(iter(path))
     if cmd.upper() != "M":
-        raise ValueError(f"Path for {shape} should start with a move")
+        raise ValueError(f"Path for {path} should start with a move")
     return args
 
 
-def normalize(shape: SVGShape, tolerance: int = DEFAULT_TOLERANCE) -> SVGShape:
-    """Build a version of shape that will compare == to other shapes even if offset.
+def _vectors(path: SVGPath) -> Generator[Vector, None, None]:
+    for cmd, args in path:
+        x_coord_idxs, y_coord_idxs = svg_meta.cmd_coords(cmd)
+        if cmd.lower() == "z":
+            yield Vector(0.0, 0.0)
+        else:
+            yield Vector(args[x_coord_idxs[-1]], args[y_coord_idxs[-1]])
 
-    Intended use is to normalize multiple shapes to identify opportunity for reuse."""
-    shape = dataclasses.replace(shape, id="")
-    x, y = _first_move(shape)
-    shape = (
-        shape.as_path().move(-x, -y, inplace=True).round_floats(tolerance, inplace=True)
+
+def _nth_vector(path: SVGPath, n: int) -> Vector:
+    return next(islice(_vectors(path), n, n + 1))
+
+
+def _angle(v: Vector) -> float:
+    # gives the directional angle of vector (unlike acos)
+    return atan2(v.y, v.x)
+
+
+def _affine_vec2vec(initial: Vector, target: Vector) -> Affine2D:
+    affine = Affine2D.identity()
+
+    # rotate initial to have the same angle as target (may have different magnitude)
+    angle = _angle(target) - _angle(initial)
+    affine = Affine2D.identity().rotate(angle)
+    vec = affine.map_vector(initial)
+
+    # scale to target magnitude
+    s = 0
+    if vec.norm() != 0:
+        s = target.norm() / vec.norm()
+
+    affine = Affine2D.compose_ltr((affine, Affine2D.identity().scale(s, s)))
+
+    return affine
+
+
+def _first_y(vectors: Iterable[Vector], tolerance: float) -> Optional[Vector]:
+    tolerance = _SIGNIFICANCE_FACTOR * tolerance
+    for idx, vec in enumerate(vectors):
+        if idx > 0 and abs(vec.y) > tolerance:
+            return vec
+    return None
+
+
+# Makes a shape safe for a walk with _affine_callback
+def _affine_friendly(shape: SVGShape) -> SVGPath:
+    path = shape.as_path()
+    if shape is path:
+        path = copy.deepcopy(path)
+    return (
+        path.relative(inplace=True)
+        .explicit_lines(inplace=True)
+        .expand_shorthand(inplace=True)
     )
-    return shape
 
 
-def affine_between(
-    s1: SVGShape, s2: SVGShape, tolerance: int = DEFAULT_TOLERANCE
-) -> Optional[Affine2D]:
+# Transform all coords in an affine-friendly path
+def _affine_callback(affine, subpath_start, curr_pos, cmd, args, *_unused):
+    x_coord_idxs, y_coord_idxs = svg_meta.cmd_coords(cmd)
+    # hard to do things like rotate if we have 1d coords
+    assert len(x_coord_idxs) == len(y_coord_idxs), f"{cmd}, {args}"
+
+    args = list(args)  # we'd like to mutate 'em
+    for x_coord_idx, y_coord_idx in zip(x_coord_idxs, y_coord_idxs):
+        if cmd == cmd.upper():
+            # for an absolute cmd allow translation: map_point
+            new_x, new_y = affine.map_point((args[x_coord_idx], args[y_coord_idx]))
+        else:
+            # for a relative coord no translate: map_vector
+            new_x, new_y = affine.map_vector((args[x_coord_idx], args[y_coord_idx]))
+
+        if almost_equal(new_x, 0):
+            new_x = 0
+        if almost_equal(new_y, 0):
+            new_y = 0
+        args[x_coord_idx] = new_x
+        args[y_coord_idx] = new_y
+    return ((cmd, args),)
+
+
+def normalize(shape: SVGShape, tolerance: float, ndigits: int) -> SVGShape:
+    """Build a version of shape that will compare == to other shapes even if offset,
+    scaled, rotated, etc.
+
+    Intended use is to normalize multiple shapes to identify opportunity for reuse."""
+
+    path = _affine_friendly(dataclasses.replace(shape, id=""))
+
+    # Make path relative, with first coord at 0,0
+    x, y = _first_move(path)
+    path.move(-x, -y, inplace=True)
+
+    # Normlize vector 1 to [1 0]; eliminates rotation and uniform scaling
+    vec1 = _nth_vector(path, 1)  # ignore M 0,0
+    affine1 = _affine_vec2vec(vec1, Vector(1, 0))
+    path.walk(lambda *args: _affine_callback(affine1, *args))
+
+    # Scale first y movement to 1.0
+    vecy = _first_y(_vectors(path), tolerance)
+    if vecy and not almost_equal(vecy.y, 1.0):
+        affine2 = Affine2D.identity().scale(1, 1 / vecy.y)
+        path.walk(lambda *args: _affine_callback(affine2, *args))
+
+    # TODO: what if shapes are the same but different start point
+    # TODO: what if shapes are the same but different drawing cmds
+    # This DOES happen in Noto; extent unclear
+
+    path.round_floats(ndigits, inplace=True)
+    return path
+
+
+def _apply_affine(affine: Affine2D, s: SVGPath) -> SVGPath:
+    s_prime = copy.deepcopy(s)
+    s_prime.walk(lambda *args: _affine_callback(affine, *args))
+    return s_prime
+
+
+def _try_affine(affine: Affine2D, s1: SVGPath, s2: SVGPath, tolerance: float):
+    return _apply_affine(affine, s1).almost_equals(s2, tolerance)
+
+
+def _round(affine, s1, s2, tolerance):
+    # TODO bsearch?
+    for i in _ROUND_RANGE:
+        rounded = affine.round(i)
+        if _try_affine(rounded, s1, s2, tolerance):
+            return rounded
+    return affine  # give up
+
+
+def affine_between(s1: SVGShape, s2: SVGShape, tolerance: float) -> Optional[Affine2D]:
     """Returns the Affine2D to change s1 into s2 or None if no solution was found.
 
-    Implementation starting *very* basic, can improve over time.
+    Intended use is to call this only when the normalized versions of the shapes
+    are the same, in which case finding a solution is typical
+
     """
     s1 = dataclasses.replace(s1, id="")
     s2 = dataclasses.replace(s2, id="")
 
     if s1.almost_equals(s2, tolerance):
         return Affine2D.identity()
 
-    s1 = s1.as_path()
-    s2 = s2.as_path()
+    s1 = _affine_friendly(s1)
+    s2 = _affine_friendly(s2)
 
     s1x, s1y = _first_move(s1)
     s2x, s2y = _first_move(s2)
-    dx = s2x - s1x
-    dy = s2y - s1y
 
-    s1.move(dx, dy, inplace=True)
+    affine = Affine2D.identity().translate(s2x - s1x, s2y - s1y)
+    if _try_affine(affine, s1, s2, tolerance):
+        return affine
 
-    if s1.almost_equals(s2, tolerance):
-        return Affine2D.identity().translate(dx, dy)
+    # Normalize first edge.
+    # Fixes rotation, x-scale, and uniform scaling.
+    s1_vec1 = _nth_vector(s1, 1)
+    s2_vec1 = _nth_vector(s2, 1)
+
+    s1_to_origin = Affine2D.identity().translate(-s1x, -s1y)
+    s2_to_origin = Affine2D.identity().translate(-s2x, -s2y)
+    s1_vec1_to_s2_vec1 = _affine_vec2vec(s1_vec1, s2_vec1)
 
+    # Move to s2 start
+    origin_to_s2 = Affine2D.identity().translate(s2x, s2y)
+
+    affine = Affine2D.compose_ltr((s1_to_origin, s1_vec1_to_s2_vec1, origin_to_s2))
+    if _try_affine(affine, s1, s2, tolerance):
+        return _round(affine, s1, s2, tolerance)
+
+    # Could be non-uniform scaling and/or mirroring
+    # Scale first y movement (after matching up vec1) to match
+
+    # Rotate first edge to lie on x axis
+    s2_vec1_angle = _angle(s2_vec1)
+    rotate_s2vec1_onto_x = Affine2D.identity().rotate(-s2_vec1_angle)
+    rotate_s2vec1_off_x = Affine2D.identity().rotate(s2_vec1_angle)
+
+    affine = Affine2D.compose_ltr(
+        (s1_to_origin, s1_vec1_to_s2_vec1, rotate_s2vec1_onto_x)
+    )
+    s1_prime = _apply_affine(affine, s1)
+
+    affine = Affine2D.compose_ltr((s2_to_origin, rotate_s2vec1_onto_x))
+    s2_prime = _apply_affine(affine, s2)
+
+    s1_vecy = _first_y(_vectors(s1_prime), tolerance)
+    s2_vecy = _first_y(_vectors(s2_prime), tolerance)
+
+    if s1_vecy and s2_vecy:
+        affine = Affine2D.compose_ltr(
+            (
+                s1_to_origin,
+                s1_vec1_to_s2_vec1,
+                # lie vec1 along x axis
+                rotate_s2vec1_onto_x,
+                # scale first y-vectors to match; x-parts should already match
+                Affine2D.identity().scale(1.0, s2_vecy.y / s1_vecy.y),
+                # restore the rotation we removed
+                rotate_s2vec1_off_x,
+                # drop into final position
+                origin_to_s2,
+            )
+        )
+        if _try_affine(affine, s1, s2, tolerance):
+            return _round(affine, s1, s2, tolerance)
+
+    # If we still aren't the same give up
     return None

src/picosvg/svg_transform.py

@@ -17,9 +17,10 @@
 Focuses on converting to a sequence of affine matrices.
 """
 import collections
+from functools import reduce
 from math import cos, sin, radians, tan
 import re
-from typing import NamedTuple, Tuple
+from typing import NamedTuple, Sequence, Tuple
 from sys import float_info
 from picosvg.geometric_types import Point, Rect, Vector
 
@@ -48,6 +49,10 @@ class Affine2D(NamedTuple):
     e: float
     f: float
 
+    @staticmethod
+    def flip_y():
+        return Affine2D._flip_y
+
     @staticmethod
     def identity():
         return Affine2D._identity
@@ -145,9 +150,22 @@ def map_vector(self, vec: Tuple[float, float]) -> Vector:
         x, y = vec
         return Vector(self.a * x + self.c * y, self.b * x + self.d * y)
 
+    @classmethod
+    def compose_ltr(cls, affines: Sequence["Affine2D"]) -> "Affine2D":
+        """Creates merged transform equivalent to applying transforms left-to-right order.
+
+        Affines apply like functions - f(g(x)) - so we merge them in reverse order.
+        """
+        return reduce(
+            lambda acc, a: cls.product(a, acc), reversed(affines), cls.identity()
+        )
+
+    def round(self, digits: int) -> "Affine2D":
+        return Affine2D(*(round(v, digits) for v in self))
+
     @classmethod
     def rect_to_rect(cls, src: Rect, dst: Rect) -> "Affine2D":
-        """ Return Affine2D set to scale and translate src Rect to dst Rect.
+        """Return Affine2D set to scale and translate src Rect to dst Rect.
         The mapping completely fills dst, it does not preserve aspect ratio.
         """
         if src.empty():
@@ -163,6 +181,7 @@ def rect_to_rect(cls, src: Rect, dst: Rect) -> "Affine2D":
 
 Affine2D._identity = Affine2D(1, 0, 0, 1, 0, 0)
 Affine2D._degnerate = Affine2D(0, 0, 0, 0, 0, 0)
+Affine2D._flip_y = Affine2D(1, 0, 0, -1, 0, 0)
 
 
 def _fix_rotate(args):