packages feed

half-0.3.2: src/Numeric/Half/Internal.hs

{-# LANGUAGE BangPatterns             #-}
{-# LANGUAGE CPP                      #-}
{-# LANGUAGE DeriveGeneric            #-}
{-# LANGUAGE ForeignFunctionInterface #-}
{-# LANGUAGE TemplateHaskellQuotes    #-}
{-# LANGUAGE PatternSynonyms          #-}
{-# LANGUAGE Trustworthy              #-}

-----------------------------------------------------------------------------
-- |
-- Copyright   :  (C) 2014 Edward Kmett
-- License     :  BSD-style (see the file LICENSE)
-- Maintainer  :  Edward Kmett <ekmett@gmail.com>
-- Stability   :  experimental
-- Portability :  PatternSynonyms
--
-- Half-precision floating-point values. These arise commonly in GPU work
-- and it is useful to be able to compute them and compute with them on the
-- CPU as well.
----------------------------------------------------------------------------

module Numeric.Half.Internal
  ( Half(..)
  , isZero
  , fromHalf
  , toHalf
  -- * Patterns
  , pattern POS_INF
  , pattern NEG_INF
  , pattern QNaN
  , pattern SNaN
  , pattern HALF_MIN
  , pattern HALF_NRM_MIN
  , pattern HALF_MAX
  , pattern HALF_EPSILON
  , pattern HALF_DIG
  , pattern HALF_MIN_10_EXP
  , pattern HALF_MAX_10_EXP
  -- * Pure conversions
  , pure_floatToHalf
  , pure_halfToFloat
  ) where

import Control.DeepSeq (NFData (..))
import Data.Bits
import Data.Function (on)
import Data.Int
import Foreign.C.Types (CUShort (..))
import Foreign.Ptr (castPtr)
import Foreign.Storable
import GHC.Generics
import Text.Read (Read (..))

import Language.Haskell.TH.Syntax (Lift (..))

import Data.Binary (Binary (..))

#ifdef __GHCJS__
toHalf :: Float -> Half
toHalf = pure_floatToHalf

fromHalf :: Half -> Float
fromHalf = pure_halfToFloat
#else
-- | Convert a 'Float' to a 'Half' with proper rounding, while preserving NaN and dealing appropriately with infinity
foreign import ccall unsafe "hs_floatToHalf" toHalf :: Float -> Half
-- {-# RULES "toHalf"  realToFrac = toHalf #-}

-- | Convert a 'Half' to a 'Float' while preserving NaN
foreign import ccall unsafe "hs_halfToFloat" fromHalf :: Half -> Float
-- {-# RULES "fromHalf" realToFrac = fromHalf #-}
#endif

-- | A half-precision floating point value
newtype
  {-# CTYPE "unsigned short" #-}
  Half = Half { getHalf :: CUShort } deriving Generic

instance NFData Half where
  rnf (Half f) = rnf f

instance Binary Half where
  put (Half (CUShort w)) = put w
  get = fmap (Half . CUShort)  get

instance Storable Half where
  sizeOf = sizeOf . getHalf
  alignment = alignment . getHalf
  peek p = fmap Half (peek (castPtr p))
  poke p = poke (castPtr p) . getHalf

instance Show Half where
  showsPrec d h = showsPrec d (fromHalf h)

instance Read Half where
  readPrec = fmap toHalf readPrec

instance Eq Half where
  (==) = (==) `on` fromHalf

instance Ord Half where
  compare = compare `on` fromHalf
  (<) = (<) `on` fromHalf
  (<=) = (<=) `on` fromHalf
  (>) = (>) `on` fromHalf
  (>=) = (>=) `on` fromHalf

instance Real Half where
  toRational = toRational . fromHalf

instance Fractional Half where
  fromRational = toHalf . fromRational
  recip = toHalf . recip . fromHalf
  a / b = toHalf $ fromHalf a / fromHalf b

instance RealFrac Half where
  properFraction a = case properFraction (fromHalf a) of
    (b, c) -> (b, toHalf c)
  truncate = truncate . fromHalf
  round = round . fromHalf
  ceiling = ceiling . fromHalf
  floor = floor . fromHalf

instance Floating Half where
  pi = toHalf pi
  exp = toHalf . exp . fromHalf
  sqrt = toHalf . sqrt . fromHalf
  log = toHalf . log . fromHalf
  a ** b = toHalf $ fromHalf a ** fromHalf b
  logBase a b = toHalf $ logBase (fromHalf a) (fromHalf b)
  sin = toHalf . sin . fromHalf
  tan = toHalf . tan . fromHalf
  cos = toHalf . cos . fromHalf
  asin = toHalf . asin . fromHalf
  atan = toHalf . atan . fromHalf
  acos = toHalf . acos . fromHalf
  sinh = toHalf . sinh . fromHalf
  tanh = toHalf . tanh . fromHalf
  cosh = toHalf . cosh . fromHalf
  asinh = toHalf . asinh . fromHalf
  atanh = toHalf . atanh . fromHalf
  acosh = toHalf . acosh . fromHalf

instance RealFloat Half where
  floatRadix  _ = 2
  floatDigits _ = 11
  decodeFloat = ieee754_f16_decode
  isIEEE _ = isIEEE (undefined :: Float)
  atan2 a b = toHalf $ atan2 (fromHalf a) (fromHalf b)

  isInfinite (Half h) = unsafeShiftR h 10 .&. 0x1f >= 31 && h .&. 0x3ff == 0
  isDenormalized (Half h) = unsafeShiftR h 10 .&. 0x1f == 0 && h .&. 0x3ff /= 0
  isNaN (Half h) = unsafeShiftR h 10 .&. 0x1f == 0x1f && h .&. 0x3ff /= 0

  isNegativeZero (Half h) = h == 0x8000
  floatRange _ = (-13,16)
  encodeFloat i j = toHalf $ encodeFloat i j
  exponent = exponent . fromHalf
  significand = toHalf . significand . fromHalf
  scaleFloat n = toHalf . scaleFloat n . fromHalf

-- | Is this 'Half' equal to 0?
isZero :: Half -> Bool
isZero (Half h) = h .&. 0x7fff == 0

-- | Positive infinity
pattern POS_INF = Half 0x7c00

-- | Negative infinity
pattern NEG_INF = Half 0xfc00

-- | Quiet NaN
pattern QNaN    = Half 0x7fff

-- | Signalling NaN
pattern SNaN    = Half 0x7dff

-- | Smallest positive half
pattern HALF_MIN = Half 0x0001  -- 5.96046448e-08

-- | Smallest positive normalized half
pattern HALF_NRM_MIN = Half 0x0400  -- 6.10351562e-05

-- | Largest positive half
pattern HALF_MAX = Half 0x7bff  -- 65504.0

-- | Smallest positive e for which half (1.0 + e) != half (1.0)
pattern HALF_EPSILON = Half 0x1400  -- 0.00097656

-- | Number of base 10 digits that can be represented without change
pattern HALF_DIG = 2

-- | Minimum positive integer such that 10 raised to that power is a normalized half
pattern HALF_MIN_10_EXP = -4

-- | Maximum positive integer such that 10 raised to that power is a normalized half
pattern HALF_MAX_10_EXP = 4

instance Num Half where
  a * b = toHalf (fromHalf a * fromHalf b)
  a - b = toHalf (fromHalf a - fromHalf b)
  a + b = toHalf (fromHalf a + fromHalf b)
  negate (Half a) = Half (xor 0x8000 a)
  abs = toHalf . abs . fromHalf
  signum = toHalf . signum . fromHalf
  fromInteger a = toHalf (fromInteger a)

instance Lift Half where
  lift (Half (CUShort w)) = [| Half (CUShort w) |]
#if MIN_VERSION_template_haskell(2,16,0)
  liftTyped (Half (CUShort w)) = [|| Half (CUShort w) ||]
#endif

-- Adapted from ghc/rts/StgPrimFloat.c
--
ieee754_f16_decode :: Half -> (Integer, Int)
ieee754_f16_decode (Half (CUShort i)) =
  let
      _HHIGHBIT                       = 0x0400
      _HMSBIT                         = 0x8000
      _HMINEXP                        = ((_HALF_MIN_EXP) - (_HALF_MANT_DIG) - 1)
      _HALF_MANT_DIG                  = floatDigits (undefined::Half)
      (_HALF_MIN_EXP, _HALF_MAX_EXP)  = floatRange  (undefined::Half)

      high1 = fromIntegral i
      high2 = high1 .&. (_HHIGHBIT - 1)

      exp1  = ((fromIntegral high1 `unsafeShiftR` 10) .&. 0x1F) + _HMINEXP
      exp2  = exp1 + 1

      (high3, exp3)
            = if exp1 /= _HMINEXP
                then (high2 .|. _HHIGHBIT, exp1)
                else
                      let go (!h, !e) =
                            if h .&. _HHIGHBIT /= 0
                              then go (h `unsafeShiftL` 1, e-1)
                              else (h, e)
                      in
                      go (high2, exp2)

      high4 = if fromIntegral i < (0 :: Int16)
                then -high3
                else  high3
  in
  if high1 .&. complement _HMSBIT == 0
    then (0,0)
    else (high4, exp3)

-- | Naive pure-Haskell implementation of 'toHalf'.
--
pure_floatToHalf :: Float -> Half
pure_floatToHalf = Half . pure_floatToHalf'

pure_floatToHalf' :: Float -> CUShort
pure_floatToHalf' x | isInfinite x = if x < 0 then 0xfc00 else 0x7c00
pure_floatToHalf' x | isNaN x = 0xfe00
-- for some reason, comparing with 0 and then deciding sign fails with GHC-7.8
pure_floatToHalf' x | isNegativeZero x = 0x8000
pure_floatToHalf' 0 = 0
pure_floatToHalf' x = let
  (m, n) = decodeFloat x
  -- sign bit
  s = if signum m < 0 then 0x8000 else 0
  m1 = fromIntegral $ abs m :: Int
  -- bit len of m1, here m1 /= 0
  len = 1 + snd (foldl (\(acc, res) y -> if acc .&. y == 0
                                         then (acc,       2*res)
                                         else (acc .&. y, 2*res + 1))
                       (m1, 0)
                       [ 0xffff0000, 0xff00ff00ff00, 0xf0f0f0f0
                       , 0xcccccccc, 0xaaaaaaaa]
                )
  -- scale to at least 12bit
  (len', m', n') = if len > 11 then (len, m1, n)
                   else (12, shiftL m1 (11 - len), n - (11 - len))
  e = n' + len' - 1
  in
  if e > 15 then fromIntegral (s .|. 0x7c00)
  else if e >= -14 then let t' = len' - 11
                            m'' = m' + (2 ^ (t' - 1) - 1) +
                                  (shiftR m' t' .&. 1)
                            len'' = if testBit m'' len then len' + 1 else len'
                            t'' = len'' - 11
                            e'' = n' + len'' - 1
                            res = (shiftR m'' t'' .&. 0x3ff) .|.
                                  shiftL ((e'' + 15) .&. 0x1f) 10 .|.
                                  s
                            in if e'' > 15
                               then fromIntegral (s .|. 0x7c00)
                               else fromIntegral res
  -- subnormal
  else if e >= -25 then let t = -n' + 1 -11 - 14
                            m'' = m' + (2 ^ (t - 1) - 1) +
                                  (shiftR m' t .&. 1)
                            res = shiftR m'' t .|. s
                            in if e == -15 && testBit m'' (10 + t)
                               then fromIntegral $ (shiftR m'' t .&. 0x3ff) .|.
                                                   0x400 .|. s
                               else fromIntegral res
  else fromIntegral s

-- | Naive pure-Haskell implementation of 'fromHalf'.
pure_halfToFloat :: Half -> Float
pure_halfToFloat = pure_halfToFloat' . getHalf

pure_halfToFloat' :: CUShort -> Float
pure_halfToFloat' 0xfc00 = -1/0
pure_halfToFloat' 0x7c00 =  1/0
pure_halfToFloat' 0x0000 =  0
pure_halfToFloat' 0x8000 = -0
pure_halfToFloat' x | (x .&. 0x7c00 == 0x7c00) && (x .&. 0x3ff /= 0) = 0/0
pure_halfToFloat' x = let
  s = if x .&. 0x8000 /= 0 then -1 else 1
  e = fromIntegral (shiftR x 10) .&. 0x1f :: Int
  m = x .&. 0x3ff
  (a, b) = if e > 0 then (e - 15 - 10, m .|. 0x400)
           else (-15 - 10 + 1, m)
  in encodeFloat (s * fromIntegral b) a