Advent of code 2024 day 20 numpy fully vectorized

def day20(s, *, part2=False, min_savings=100):  # Fully vectorized numpy solution.  Not fastest.
  grid = np.array([list(line) for line in s.splitlines()])
  ((ys, xs),) = np.argwhere(grid == 'S')
  ((ye, xe),) = np.argwhere(grid == 'E')
  distance = np.where(grid == '#', -2, -1)
  neighbors = (0, -1), (0, 1), (-1, 0), (1, 0)

  d, y, x = 0, ys, xs
  distance[y, x] = d
  while (y, x) != (ye, xe):
    ((y, x),) = (yx2 for dy, dx in neighbors if distance[yx2 := (y + dy, x + dx)] == -1)
    distance[y, x] = d = d + 1

  radius = 20 if part2 else 2
  y, x = np.indices((radius + 1, 2 * radius + 1))
  x -= radius
  valid = (abs(y) + abs(x) <= radius) & ((y > 0) | (y == 0) & (x > 0))
  shifts = np.argwhere(valid)
  y_shifts = shifts[:, 0]
  x_shifts = shifts[:, 1] - radius
  manhattan = abs(y_shifts) + abs(x_shifts)

  padded = np.pad(distance, radius, constant_values=-1)

  # Create a view for vectorized shifting.
  ny, nx = distance.shape
  shifted_views = padded[
      radius + y_shifts[:, None, None] + np.arange(ny)[None, :, None],
      radius + x_shifts[:, None, None] + np.arange(nx)[None, None, :],
  ]

  # Create mask for valid (non-negative) values.
  valid_mask = distance >= 0
  shifted_valid = shifted_views >= 0

  # Calculate value differences and adjust by Manhattan distance.
  value_diff = np.abs(shifted_views - distance)
  adjusted_diff = value_diff - manhattan[:, None, None]

  # Count pairs satisfying all conditions.
  valid_pairs = ((adjusted_diff >= min_savings) & valid_mask & shifted_valid).sum(axis=(1, 2))
  return valid_pairs.sum()