diff --git a/Data/Number/LogFloat.hs b/Data/Number/LogFloat.hs --- a/Data/Number/LogFloat.hs +++ b/Data/Number/LogFloat.hs @@ -10,12 +10,19 @@ -- http://www.haskell.org/ghc/docs/latest/html/users_guide/rewrite-rules.html -- http://www.randomhacks.net/articles/2007/02/10/map-fusion-and-haskell-performance --- {-# OPTIONS_GHC -ddump-simpl-stats #-} +-- {-# OPTIONS_GHC -ddump-rules -ddump-simpl-stats #-} -{-# OPTIONS_GHC -Wall -Werror #-} -{-# OPTIONS_GHC -O2 -fvia-C -optc-O3 #-} +{-# OPTIONS_GHC -Wall -Werror #-} +-- Unfortunately we need -fglasgow-exts in order to actually pick +-- up on the rules (see -ddump-rules). The -frewrite-rules flag +-- doesn't do what you want. +-- cf +-- cf +{-# OPTIONS_GHC -O2 -fvia-C -optc-O3 -fexcess-precision -fglasgow-exts #-} + -- Version History +-- (v0.8.6) Removed buggy RULES -- (v0.8.5) Gave up and converted from lhs to hs so Hackage docs work -- (v0.8.4) Broke out Transfinite -- (v0.8.3) Documentation updates @@ -30,7 +37,7 @@ -- (v0.1) Initial version created for hw5 for NLP with Jason Eisner. -- ---------------------------------------------------------------- --- ~ 2008.08.16 +-- ~ 2008.08.17 -- | -- Module : Data.Number.LogFloat -- Copyright : Copyright (c) 2007--2008 wren ng thornton @@ -78,11 +85,11 @@ ---------------------------------------------------------------- -- --- Try to add in some optimizations. Why the first few need to be +-- Try to add in some optimizations. Why these need to be -- down here and localized to the module, I don't know. We don't -- do anything foolish like this, but our clients might, or they -- might be generated by other code transformations. Note that due --- to the fuzz, these equations are not actually true, even though +-- to the fuzz, these equations are not strictly true, even though -- they are mathematically correct. {-# RULES @@ -93,14 +100,19 @@ "exp.log" exp . log = id #-} --- These are general rule versions of our operators for 'LogFloat'. --- I had some issues inducing 'Ord' on @x@ and @y@, even though --- they're 'Num' so I can't do "(+)\/log" and "(-)\/log" so easily. - -{-# RULES -"(*)/log" forall x y. log x * log y = log (x + y) -"(/)/log" forall x y. log x / log y = log (x - y) - #-} +-- We'd like to be able to take advantage of general rule versions +-- of our operators for 'LogFloat', with rules like @log x + log y +-- = log (x * y)@ and @log x - log y = log (x / y)@. However the +-- problem is that those equations could be driven in either direction +-- depending on whether we think time performance or non-underflow +-- performance is more important, and the answers may be different +-- at every call site. +-- +-- Since we implore users to do normal-domain computations whenever +-- it would not degenerate accuracy, we should not rewrite their +-- decisions in any way. The log\/exp fusion strictly improves both +-- time and accuracy, so those are safe. But the buck stops with +-- them. ---------------------------------------------------------------- @@ -140,23 +152,52 @@ -- into @Ratio@ types is error-prone and non-portable, as discussed -- in "Data.Number.Transfinite". +{-# INLINE [1] toFractional #-} +{-# SPECIALIZE toFractional :: (Real a) => a -> Float #-} {-# SPECIALIZE toFractional :: (Real a) => a -> Double #-} {-# SPECIALIZE toFractional :: (Fractional b) => Double -> b #-} toFractional :: (Real a, Fractional b) => a -> b toFractional = fromRational . toRational +-- The INLINE pragma is to /delay/ inlining, so the rules below can +-- have their way + -- This should only fire when it's type-safe {-# RULES "toFractional/id" toFractional = id #-} --- This should happen already, but who knows --- TODO: see if it ever fires +-- These too should only fire when it's type-safe +-- This should already happen, but... +-- TODO: Check the logs to see if it ever fires +-- BUG: why does -ddump-rules call these two orphaned? {-# RULES -"toFractional/toFractional" forall x. - toFractional (toFractional x) = toFractional x -"toFractional.toFractional" toFractional . toFractional = toFractional +"toRational/fromRational" forall x. toRational (fromRational x) = x +"toRational.fromRational" toRational . fromRational = id #-} +-- 'toFractional' should be inlined and the above rules should +-- obviate this, but... +-- TODO: We should check the logs to see if it ever fires before +-- removing them +{-# RULES +"toFractional/toFractional" forall x. + toFractional (toFractional x) = toFractional x +"toFractional.toFractional" toFractional . toFractional = toFractional + #-} + +{- It looks like we need these for vast performance improvement. + Is there some way to include them without resorting to CPP or + other non-portability? + + +import GHC.Prim +{-# RULES "toFractional::Int->Double" toFractional = i2d #-} +i2d (I# i) = D# (int2Double# i) +{-# RULES "toFractional::Int->Float" toFractional = i2f #-} +i2f (I# i) = F# (int2Float# i) +-} + + ---------------------------------------------------------------- -- -- | Reduce the number of constant string literals we need to store. @@ -173,7 +214,7 @@ | otherwise = errorOutOfRange fun --- | It's unfortunate that notANumber is not equal to itself, but +-- | It's unfortunate that 'notANumber' is not equal to itself, but -- we can hack around that. GHC gives NaN for the log of negatives -- and so we could ideally take advantage of @log . guardNonNegative -- fun = guardIsANumber fun . log@ to simplify things, but Hugs @@ -255,6 +296,14 @@ -- These are our module-specific versions of "log\/exp" and "exp\/log"; -- They do the same things but also have a @LogFloat@ in between -- the logarithm and exponentiation. +-- +-- In order to ensure these rules fire we may need to delay inlining +-- of the four con-\/destructors, like we do for 'toFractional'. +-- Unfortunately, I'm not entirely sure whether they will be inlined +-- already or not (and whether they are may be fragile) and I don't +-- want to inline them excessively and lead to code bloat in the +-- off chance that we could prune some of it away. +-- TODO: thoroughly investigate this. {-# RULES -- Out of log-domain and back in diff --git a/logfloat.cabal b/logfloat.cabal --- a/logfloat.cabal +++ b/logfloat.cabal @@ -1,6 +1,9 @@ ---------------------------------------------------------------- +-- wren ng thornton ~ 2008.08.17 +---------------------------------------------------------------- + Name: logfloat -Version: 0.8.5 +Version: 0.8.6 Cabal-Version: >= 1.2 Build-Type: Simple Stability: stable @@ -22,7 +25,6 @@ Exposed-Modules: Data.Number.LogFloat , Data.Number.Transfinite Build-Depends: base - Ghc-Options: -O2 -fvia-C -optc-O3 ---------------------------------------------------------------- ----------------------------------------------------------- fin.