diff --git a/hw-simd.cabal b/hw-simd.cabal --- a/hw-simd.cabal +++ b/hw-simd.cabal @@ -1,11 +1,9 @@ --- This file has been generated from package.yaml by hpack version 0.20.0. +-- This file has been generated from package.yaml by hpack version 0.18.1. -- -- see: https://github.com/sol/hpack --- --- hash: d4e6b647bff1de25a793f93a5adb597bbbc16d00285948adf0c1aa4f8e3c542f name: hw-simd -version: 0.0.0.2 +version: 0.0.0.3 synopsis: SIMD library description: Please see the README on Github at category: Text, Web, CSV @@ -53,15 +51,15 @@ c-sources: cbits/simd.c build-depends: - base >=4.7 && <5 - , bits-extra >=0.0.1.2 && <0.1 - , bytestring >=0.10 && <0.11 - , deepseq >=1.4 && <1.5 - , hw-bits >=0.7.0.2 && <0.8 - , hw-prim >=0.6.2.0 && <0.7 - , hw-rankselect >=0.12.0.2 && <0.13 - , hw-rankselect-base >=0.3.2.0 && <0.4 - , vector >=0.12.0.1 && <0.13 + base >= 4.7 && < 5 + , bits-extra >= 0.0.1.2 && < 0.1 + , bytestring >= 0.10 && < 0.11 + , deepseq >= 1.4 && < 1.5 + , hw-bits >= 0.7.0.2 && < 0.8 + , hw-rankselect >= 0.12.0.2 && < 0.13 + , hw-rankselect-base >= 0.3.2.0 && < 0.4 + , hw-prim >= 0.6.2.0 && < 0.7 + , vector >= 0.12.0.1 && < 0.13 build-tools: c2hs if flag(sse42) @@ -80,8 +78,8 @@ ghc-options: -Wcompat -Wincomplete-record-updates -Wincomplete-uni-patterns -Wredundant-constraints if (!impl(ghc >=8.0.1)) build-depends: - semigroups >=0.8.4 && <0.19 - , transformers >=0.4 && <0.6 + transformers >= 0.4 && < 0.6 + , semigroups >= 0.8.4 && < 0.19 exposed-modules: HaskellWorks.Data.Simd.Capabilities HaskellWorks.Data.Simd.Comparison @@ -103,21 +101,21 @@ test ghc-options: -O2 -Wall -threaded -rtsopts -with-rtsopts=-N build-depends: - base >=4.7 && <5 - , bits-extra >=0.0.1.2 && <0.1 - , bytestring >=0.10 && <0.11 - , deepseq >=1.4 && <1.5 - , directory >=1.2.2 && <1.4 - , hedgehog >=0.5 && <0.7 - , hspec >=2.4 && <3 - , hw-bits >=0.7.0.2 && <0.8 - , hw-hspec-hedgehog >=0.1.0.4 && <0.2 - , hw-prim >=0.6.2.0 && <0.7 - , hw-rankselect >=0.12.0.2 && <0.13 - , hw-rankselect-base >=0.3.2.0 && <0.4 + base >= 4.7 && < 5 + , bits-extra >= 0.0.1.2 && < 0.1 + , bytestring >= 0.10 && < 0.11 + , deepseq >= 1.4 && < 1.5 + , hw-bits >= 0.7.0.2 && < 0.8 + , hw-rankselect >= 0.12.0.2 && < 0.13 + , hw-rankselect-base >= 0.3.2.0 && < 0.4 + , hw-prim >= 0.6.2.0 && < 0.7 + , vector >= 0.12.0.1 && < 0.13 + , directory >= 1.2.2 && < 1.4 , hw-simd - , text >=1.2.2 && <2.0 - , vector >=0.12.0.1 && <0.13 + , hedgehog >= 0.5 && < 0.7 + , hspec >= 2.4 && < 3 + , hw-hspec-hedgehog >= 0.1.0.4 && < 0.2 + , text >= 1.2.2 && < 2.0 if flag(sse42) ghc-options: -msse4.2 cc-options: -msse4.2 @@ -134,11 +132,10 @@ ghc-options: -Wcompat -Wincomplete-record-updates -Wincomplete-uni-patterns -Wredundant-constraints if (!impl(ghc >=8.0.1)) build-depends: - semigroups >=0.8.4 && <0.19 - , transformers >=0.4 && <0.6 + transformers >= 0.4 && < 0.6 + , semigroups >= 0.8.4 && < 0.19 other-modules: HaskellWorks.Data.Simd.BroadwordSpec - Paths_hw_simd default-language: Haskell2010 benchmark bench @@ -148,21 +145,21 @@ bench ghc-options: -O2 -Wall -msse4.2 build-depends: - base >=4.7 && <5 - , bits-extra >=0.0.1.2 && <0.1 - , bytestring >=0.10 && <0.11 - , cassava >=0.5.1.0 && <0.6 + base >= 4.7 && < 5 + , bits-extra >= 0.0.1.2 && < 0.1 + , bytestring >= 0.10 && < 0.11 + , deepseq >= 1.4 && < 1.5 + , hw-bits >= 0.7.0.2 && < 0.8 + , hw-rankselect >= 0.12.0.2 && < 0.13 + , hw-rankselect-base >= 0.3.2.0 && < 0.4 + , hw-prim >= 0.6.2.0 && < 0.7 + , vector >= 0.12.0.1 && < 0.13 + , cassava >= 0.5.1.0 && < 0.6 , containers - , criterion >=1.4.1.0 && <1.5 - , deepseq >=1.4 && <1.5 - , directory >=1.3.1.5 && <1.4 - , hw-bits >=0.7.0.2 && <0.8 - , hw-prim >=0.6.2.0 && <0.7 - , hw-rankselect >=0.12.0.2 && <0.13 - , hw-rankselect-base >=0.3.2.0 && <0.4 + , criterion >= 1.4.1.0 && < 1.5 + , directory >= 1.3.1.5 && < 1.4 , hw-simd - , mmap >=0.5.9 && <0.6 - , vector >=0.12.0.1 && <0.13 + , mmap >= 0.5.9 && < 0.6 if flag(sse42) ghc-options: -msse4.2 cc-options: -msse4.2 @@ -179,8 +176,6 @@ ghc-options: -Wcompat -Wincomplete-record-updates -Wincomplete-uni-patterns -Wredundant-constraints if (!impl(ghc >=8.0.1)) build-depends: - semigroups >=0.8.4 && <0.19 - , transformers >=0.4 && <0.6 - other-modules: - Paths_hw_simd + transformers >= 0.4 && < 0.6 + , semigroups >= 0.8.4 && < 0.19 default-language: Haskell2010 diff --git a/src/HaskellWorks/Data/Simd/Comparison/Stock.hs b/src/HaskellWorks/Data/Simd/Comparison/Stock.hs --- a/src/HaskellWorks/Data/Simd/Comparison/Stock.hs +++ b/src/HaskellWorks/Data/Simd/Comparison/Stock.hs @@ -2,38 +2,54 @@ ( cmpeq8s ) where -import Data.Monoid ((<>)) import Data.Word import HaskellWorks.Data.AtIndex import HaskellWorks.Data.Bits.BitWise import HaskellWorks.Data.Simd.Internal.Bits +import HaskellWorks.Data.Simd.Internal.Broadword import qualified Data.Vector.Storable as DVS cmpeq8s :: Word8 -> DVS.Vector Word64 -> DVS.Vector Word64 -cmpeq8s w8 v = if disalignment == 0 - then DVS.constructN (DVS.length v) go - else error $ "Unaligned byte string: " <> show disalignment <> ", vLen: " <> show vLen - where vLen = DVS.length v - disalignment = vLen - (vLen `div` 8) * 8 - w64 = fromIntegral w8 * 0x0101010101010101 :: Word64 +cmpeq8s w8 v = DVS.constructN ((DVS.length v + 7) `div` 8) go + where iw = fillWord64 w8 go :: DVS.Vector Word64 -> Word64 - go u = comp w - where ui = fromIntegral $ DVS.length u - w0 = testWord8s ((u !!! (ui + 0)) .^. w64) - w1 = testWord8s ((u !!! (ui + 1)) .^. w64) - w2 = testWord8s ((u !!! (ui + 2)) .^. w64) - w3 = testWord8s ((u !!! (ui + 3)) .^. w64) - w4 = testWord8s ((u !!! (ui + 4)) .^. w64) - w5 = testWord8s ((u !!! (ui + 5)) .^. w64) - w6 = testWord8s ((u !!! (ui + 6)) .^. w64) - w7 = testWord8s ((u !!! (ui + 7)) .^. w64) - w = (w7 .<. 56) .|. - (w6 .<. 48) .|. - (w5 .<. 40) .|. - (w4 .<. 32) .|. - (w3 .<. 24) .|. - (w2 .<. 16) .|. - (w1 .<. 8) .|. - w0 + go u = let ui = end u in + if ui * 8 + 8 < end v + then let vi = ui * 8 + w0 = testWord8s ((v !!! (vi + 0)) .^. iw) + w1 = testWord8s ((v !!! (vi + 1)) .^. iw) + w2 = testWord8s ((v !!! (vi + 2)) .^. iw) + w3 = testWord8s ((v !!! (vi + 3)) .^. iw) + w4 = testWord8s ((v !!! (vi + 4)) .^. iw) + w5 = testWord8s ((v !!! (vi + 5)) .^. iw) + w6 = testWord8s ((v !!! (vi + 6)) .^. iw) + w7 = testWord8s ((v !!! (vi + 7)) .^. iw) + w = (w7 .<. 56) .|. + (w6 .<. 48) .|. + (w5 .<. 40) .|. + (w4 .<. 32) .|. + (w3 .<. 24) .|. + (w2 .<. 16) .|. + (w1 .<. 8) .|. + w0 + in comp w + else let vi = ui * 8 + w0 = testWord8s (atIndexOr 0 v (vi + 0) .^. iw) + w1 = testWord8s (atIndexOr 0 v (vi + 1) .^. iw) + w2 = testWord8s (atIndexOr 0 v (vi + 2) .^. iw) + w3 = testWord8s (atIndexOr 0 v (vi + 3) .^. iw) + w4 = testWord8s (atIndexOr 0 v (vi + 4) .^. iw) + w5 = testWord8s (atIndexOr 0 v (vi + 5) .^. iw) + w6 = testWord8s (atIndexOr 0 v (vi + 6) .^. iw) + w7 = testWord8s (atIndexOr 0 v (vi + 7) .^. iw) + w = (w7 .<. 56) .|. + (w6 .<. 48) .|. + (w5 .<. 40) .|. + (w4 .<. 32) .|. + (w3 .<. 24) .|. + (w2 .<. 16) .|. + (w1 .<. 8) .|. + w0 + in comp w {-# INLINE cmpeq8s #-}