我为Haskell的内置exists列表数据类型编写了量化函数forallnone[]。在很多情况下,这些方法似乎比Prelude / Data.Listanyall效率更高。我天真地怀疑这种性能是由于anyall是使用Θ(n)折叠实现的。由于我是Haskell的新手,所以我认为我一定会误会,否则会有这种现象的充分原因。

Data.Foldable:

-- | Determines whether any element of the structure satisfies the predicate.
any :: Foldable t => (a -> Bool) -> t a -> Bool
any p = getAny #. foldMap (Any #. p)

-- | Determines whether all elements of the structure satisfy the predicate.
all :: Foldable t => (a -> Bool) -> t a -> Bool
all p = getAll #. foldMap (All #. p)

我的实现:
exists :: (a -> Bool) -> [a] -> Bool
exists _    []                   = False
exists pred (x : xs) | pred x    = True
                     | otherwise = exists pred xs


forall pred  =  not . exists (not . pred)
none pred  =  not . exists pred  =  forall (not . pred)

消除布尔反演:
forall, none :: (a -> Bool) -> [a] -> Bool

forall _    []                   = True
forall pred (x : xs) | pred x    = forall pred xs
                     | otherwise = False

none _    []                   = True
none pred (x : xs) | pred x    = False
                   | otherwise = none pred xs
all:
time                 327.8 μs   (322.4 μs .. 333.0 μs)
                     0.997 R²   (0.996 R² .. 0.998 R²)
mean                 328.7 μs   (324.1 μs .. 334.2 μs)
std dev              16.95 μs   (14.63 μs .. 22.02 μs)

forall:
time                 113.2 μs   (111.2 μs .. 115.0 μs)
                     0.997 R²   (0.996 R² .. 0.998 R²)
mean                 112.0 μs   (110.0 μs .. 113.9 μs)
std dev              6.333 μs   (5.127 μs .. 7.896 μs)

使用标准的nf衡量的性能。

不出所料,我并没有重蹈覆辙,而是低估了编译器标志,并且天真地希望-O2与默认优化级别的性能相比,不会产生如此大的差异,也不会期望单个定制编写方法与库公式之间的优化效果差异。许多高效的专业标准功能优化显然只有在明确启用后才能启动。

Haskell标记信息的“性能”部分强调测试代码效率时优化级别编译器标志的重要性。通常建议您信任库函数实现的复杂性,而不是重新布线RULES编译指示或重新编写基本形式,而是尝试利用已经培养的优化潜力。

最佳答案

我发现以各种方式重新实现any具有指导意义:

import Prelude hiding (any)
import Criterion.Main
import Data.Foldable (foldMap)
import Data.Monoid

您的exists:
exists :: (a -> Bool) -> [a] -> Bool
exists _ [] = False
exists pred (x : xs)
    = if pred x
      then True
      else exists pred xs

使用(||)的版本:
existsOr :: (a -> Bool) -> [a] -> Bool
existsOr _ [] = False
existsOr pred (x : xs) = pred x || existsOr pred xs

使用foldr:
any :: (a -> Bool) -> [a] -> Bool
any pred = foldr ((||) . pred) False

使用foldrAny:
anyF :: (a -> Bool) -> [a] -> Bool
anyF pred = getAny . foldr (mappend . (Any . pred)) mempty

使用foldMapAny:
anyFM :: (a -> Bool) -> [a] -> Bool
anyFM pred = getAny . foldMap (Any . pred)

带有ghc -O0的基准:
benchmarking exists
time                 1.552 μs   (1.504 μs .. 1.593 μs)
                     0.989 R²   (0.983 R² .. 0.993 R²)
mean                 1.482 μs   (1.427 μs .. 1.545 μs)
std dev              196.1 ns   (168.8 ns .. 229.2 ns)
variance introduced by outliers: 93% (severely inflated)

benchmarking existsOr
time                 2.699 μs   (2.616 μs .. 2.768 μs)
                     0.992 R²   (0.988 R² .. 0.995 R²)
mean                 2.629 μs   (2.554 μs .. 2.704 μs)
std dev              277.8 ns   (235.8 ns .. 351.1 ns)
variance introduced by outliers: 89% (severely inflated)

benchmarking any
time                 5.551 μs   (5.354 μs .. 5.777 μs)
                     0.990 R²   (0.986 R² .. 0.995 R²)
mean                 5.553 μs   (5.395 μs .. 5.750 μs)
std dev              584.2 ns   (447.5 ns .. 835.5 ns)
variance introduced by outliers: 88% (severely inflated)

benchmarking anyF
time                 7.330 μs   (7.081 μs .. 7.612 μs)
                     0.988 R²   (0.982 R² .. 0.994 R²)
mean                 7.502 μs   (7.272 μs .. 7.762 μs)
std dev              848.2 ns   (712.6 ns .. 1.022 μs)
variance introduced by outliers: 89% (severely inflated)

benchmarking anyFM
time                 5.668 μs   (5.451 μs .. 6.008 μs)
                     0.987 R²   (0.975 R² .. 0.996 R²)
mean                 5.807 μs   (5.659 μs .. 5.975 μs)
std dev              542.5 ns   (446.4 ns .. 721.8 ns)
variance introduced by outliers: 86% (severely inflated)

您的版本(exists)确实是最快的,而foldr版本却相当慢。

使用ghc -O2,您的版本(exists)是最慢的,而所有其他功能彼此之间几乎同样快:
benchmarking exists
time                 753.5 ns   (725.4 ns .. 779.9 ns)
                     0.990 R²   (0.986 R² .. 0.995 R²)
mean                 762.4 ns   (737.0 ns .. 787.0 ns)
std dev              82.47 ns   (66.79 ns .. 105.1 ns)
variance introduced by outliers: 91% (severely inflated)

benchmarking existsOr
time                 491.5 ns   (478.2 ns .. 503.2 ns)
                     0.994 R²   (0.992 R² .. 0.996 R²)
mean                 494.5 ns   (481.1 ns .. 512.9 ns)
std dev              54.97 ns   (42.54 ns .. 80.34 ns)
variance introduced by outliers: 92% (severely inflated)

benchmarking any
time                 461.2 ns   (442.0 ns .. 479.7 ns)
                     0.989 R²   (0.985 R² .. 0.993 R²)
mean                 456.0 ns   (439.3 ns .. 476.3 ns)
std dev              60.04 ns   (47.27 ns .. 89.47 ns)
variance introduced by outliers: 94% (severely inflated)

benchmarking anyF
time                 436.9 ns   (415.8 ns .. 461.0 ns)
                     0.978 R²   (0.967 R² .. 0.988 R²)
mean                 450.8 ns   (430.1 ns .. 472.6 ns)
std dev              70.64 ns   (57.04 ns .. 85.92 ns)
variance introduced by outliers: 96% (severely inflated)

benchmarking anyFM
time                 438.9 ns   (426.9 ns .. 449.5 ns)
                     0.993 R²   (0.989 R² .. 0.996 R²)
mean                 435.8 ns   (421.4 ns .. 447.6 ns)
std dev              45.32 ns   (36.73 ns .. 58.74 ns)
variance introduced by outliers: 90% (severely inflated)

如果人们查看简化的Core代码(ghc -O2 -ddump-simpl),就会发现不再有foldr了(有了-O0,一切仍然在那里,包括fold)。

因此,我敢说您的代码(在未优化的版本中,-O0)比库代码更快,因为它更简单(因为潜在的价格降低了通用性)。经过优化的库代码比您的版本要快,因为它是通过编译器可以识别其优化潜力的方式编写的。 (诚​​然,这是一些猜测工作)

10-04 18:12
查看更多