为什么我的SVG弧转换实现不能通过QuickCheck?

时间:2014-11-16 11:26:19

标签: haskell svg ieee-754 quickcheck

我在Haskell中实现了W3s recommended algorithm for converting SVG-path arcs from endpoint-arcs to center-arcs and back

type EndpointArc = ( Double, Double, Double, Double
                   , Bool, Bool, Double, Double, Double )

type CenterArc = ( Double, Double, Double, Double
                 , Double, Double, Double )

endpointToCenter :: EndpointArc -> CenterArc

centerToEndpoint :: CenterArc -> EndpointArc

See full implementation and test-code here

但是我无法通过这个属性:

import Test.QuickCheck
import Data.AEq ((~==))

instance Arbitrary EndpointArc where
    arbitrary = do
        ((x1,y1),(x2,y2)) <- arbitrary `suchThat` (\(u,v) -> u /= v)
        rx                <- arbitrary `suchThat` (>0)
        ry                <- arbitrary `suchThat` (>0)
        phi               <- choose (0,2*pi)
        (fA,fS)           <- arbitrary
        return $ correctRadiiSize (x1, y1, x2, y2, fA, fS, rx, ry, phi)

prop_conversionRetains :: EndpointArc -> Bool
prop_conversionRetains earc =
    let result = centerToEndpoint (endpointToCenter earc)
    in earc ~== result

有时这是由于浮点错误(似乎超过ieee754),但有时结果中有NaN。

(NaN,NaN,NaN,NaN,False,False,1.0314334509082723,2.732814841776921,1.2776112657142984)

这表明没有解决方案,尽管我认为我按F.6.6.2 in W3's document中所述缩放rx,ry。

import Numeric.Matrix

m :: [[Double]] -> Matrix Double
m = fromList

toTuple :: Matrix Double -> (Double, Double)
toTuple = (\[[x],[y]] -> (x,y)) . toList

primed :: Double -> Double -> Double -> Double -> Double
       -> (Double, Double)
primed x1 y1 x2 y2 phi = toTuple $
    m [[ cos phi, sin phi]
      ,[-sin phi, cos phi]
      ]
    * m [[(x1 - x2)/2]
        ,[(y1 - y2)/2]
        ]

correctRadiiSize :: EndpointArc -> EndpointArc
correctRadiiSize (x1, y1, x2, y2, fA, fS, rx, ry, phi) =
    let (x1',y1') = primed x1 y1 x2 y2 phi
        lambda    = (x1'^2/rx^2) + (y1'^2/ry^2)
        (rx',ry') | lambda <= 1 = (rx, ry)
                  | otherwise   = ((sqrt lambda) * rx, (sqrt lambda) * ry)
    in (x1, y1, x2, y2, fA, fS, rx', ry', phi)

1 个答案:

答案 0 :(得分:13)

好吧,我自己想出来了。线索当然是在W3s文件中:

  

在使用公式(F.6.6.3)放大半径的情况下,(F.6.5.2)的基数为零,并且正好有一个椭圆中心的解。

我的代码中的F.6.5.2是

(cx',cy') = (sq * rx * y1' / ry, sq * (-ry) * x1' / rx)
                where sq = negateIf (fA == fS) $ sqrt
                         $ ( rx^2 * ry^2 - rx^2 * y1'^2 - ry^2 * x1'^2 )
                         / ( rx^2 * y1'^2 + ry^2 * x1'^2 )

它指的是radicand

( rx^2 * ry^2 - rx^2 * y1'^2 - ry^2 * x1'^2 )
 / ( rx^2 * y1'^2 + ry^2 * x1'^2 )

但是,当然,因为我们正在使用花车,它不是完全零但是大约有时候它可能像-6.99496644301622e-17这样是负面的!负数的平方根是一个复数,因此计算返回NaN。

诀窍真的是传播rx和ry已经调整大小以返回零并使sq为零而不是不必要地完成整个计算的事实,但快速修复只是为了取绝对值radicand。

(cx',cy') = (sq * rx * y1' / ry, sq * (-ry) * x1' / rx)
                where sq = negateIf (fA == fS) $ sqrt $ abs
                         $ ( rx^2 * ry^2 - rx^2 * y1'^2 - ry^2 * x1'^2 )
                         / ( rx^2 * y1'^2 + ry^2 * x1'^2 )

之后还有一些剩余的浮点问题。首先,错误超出了ieee754的~==运营商所允许的范围,因此我自己制作了approxEq

approxEq (x1a, y1a, x2a, y2a, fAa, fSa, rxa, rya, phia) (x1b, y1b, x2b, y2b, fAb, fSb, rxb, ryb, phib) =
       abs (x1a - x1b  ) < 0.001
    && abs (y1a - y1b  ) < 0.001
    && abs (x2a - x2b  ) < 0.001
    && abs (y2a - y2b  ) < 0.001
    && abs (y2a - y2b  ) < 0.001
    && abs (rxa - rxb  ) < 0.001
    && abs (rya - ryb  ) < 0.001
    && abs (phia - phib) < 0.001
    && fAa == fAb
    && fSa == fSb

prop_conversionRetains :: EndpointArc -> Bool
prop_conversionRetains earc =
    let result = centerToEndpoint (trace ("FIRST:" ++ show (endpointToCenter earc)) (endpointToCenter earc))
    in earc `approxEq` trace ("SECOND:" ++ show result) result

哪个开始带来fA被翻转的案例。发现幻数:

  

FIRST:( - 5.988957688551294,-39.5430169665332,64.95929681921707,29.661347617532357,5.939852349879405,-1.2436798376040206,的 3.141592653589793

     

SECOND:(4.209851895761209,-73.01839718538467,-16.18776727286379,-6.067636747681732,的下,的确,64.95929681921707,29.661347617532357,5.939852349879405)

     

***失败了!可证伪(经过20次测试):
  (4.209851895761204,-73.01839718538467,-16.18776781572145,-6.0676366434916655,的下,的确,64.95929681921707,29.661347617532357,5.939852349879405)

你明白了! fA = abs dtheta > pi位于centerToEndpoint,所以如果它是这些,那么它可以采用任何一种方式。

所以我拿出了fA条件,并在quickcheck中增加了测试次数

approxEq (x1a, y1a, x2a, y2a, fAa, fSa, rxa, rya, phia) (x1b, y1b, x2b, y2b, fAb, fSb, rxb, ryb, phib) =
       abs (x1a - x1b  ) < 0.001
    && abs (y1a - y1b  ) < 0.001
    && abs (x2a - x2b  ) < 0.001
    && abs (y2a - y2b  ) < 0.001
    && abs (y2a - y2b  ) < 0.001
    && abs (rxa - rxb  ) < 0.001
    && abs (rya - ryb  ) < 0.001
    && abs (phia - phib) < 0.001
    -- && fAa == fAb
    && fSa == fSb

main = quickCheckWith stdArgs {maxSuccess = 50000} prop_conversionRetains

这表明阈值approxEq仍然不够宽松。

approxEq (x1a, y1a, x2a, y2a, fAa, fSa, rxa, rya, phia) (x1b, y1b, x2b, y2b, fAb, fSb, rxb, ryb, phib) =
       abs (x1a - x1b  ) < 1
    && abs (y1a - y1b  ) < 1
    && abs (x2a - x2b  ) < 1
    && abs (y2a - y2b  ) < 1
    && abs (y2a - y2b  ) < 1
    && abs (rxa - rxb  ) < 1
    && abs (rya - ryb  ) < 1
    && abs (phia - phib) < 1
    -- && fAa == fAb
    && fSa == fSb

我最终可以通过大量测试可靠地通过。好吧,这一切只是为了制作一些有趣的图形...我相信它足够准确:)