[Haskell-cafe] Re: RealFrac methods for Double and Float

[Antoine Latter]

On Sun, Oct 10, 2010 at 7:02 PM, Daniel Fischer
<daniel.is.fischer at web.de> wrote:
> Greetings,
>
> I have put together a package to test possible implementations of the
> RealFrac methods for Double and Float (base-2 IEEE754) and uploaded a
> .tar.gz bundle to http://hackage.haskell.org/trac/ghc/ticket/2271 .
>
> On the one hand, pure Haskell implementations, on the other hand
> implementations calling out to rint[f], trunc[f], floor[f] and ceil[f] from
> math.h.
>
> Both ways go via Integer by default, with a specialised faster
> implementation for Int (and narrower types, but those RULES haven't yet
> been written) enabled by a rewrite rule.
>
> Overall, the pure Haskell implementations don't fare badly on my computer.
> All give a speedup compared to the current implementation, for most
> conversions, pure Haskell is on par with or faster than the C-call
> (although that would probably change if the C functions were made primops).
>
> The FFI calls are significantly faster for
> properFraction :: Double -> (Integer, Double)
> and for round (except round :: Integral a => Float -> a when compiled
> via C, then native and FFI are on par).
>
> Sample results for the speedups against the current implementation (note:
> for truncate :: x -> Int, the Prelude value is fst . properFraction, not
> the rewritten float2Int or double2Int) are included in the tarball.
>
> I would appreciate feedback from your tests/benchmarks on other platforms,
> especially 64-bit platforms (mine is x86 linux, 32 bit).
>
> To run the QuickCheck tests, you need QuickCheck-2.*, to run the
> benchmarks, criterion.
>
> More instructions in the README.
>
> Thanks,
> Daniel
>

I have results from my Intel-based MacBook, 64-bits, GHC 7 rc. The
quickchecks failed to run:

QuickChecking Double
properFraction/Int
qcDouble: qcDouble.hs:(10,10)-(15,5): Missing field in record
construction Test.QuickCheck.Test.chatty

QuickChecking Float
properFraction/Int
qcFloat: qcFloat.hs:(10,10)-(15,5): Missing field in record
construction Test.QuickCheck.Test.chatty

This is with QuickCheck 2.3.0.2.

Take care,
Antoine

$ sh bench.sh
Results from ncgDouble:

Relations for properFraction:
  Prelude          1.000000
  C via Integer    2.429916
  Hs via Integer   1.904705
  C Int           17.251877
  Hs Int          28.596726

Relations for truncate:
  Prelude          1.000000
  C via Integer    3.743818
  Hs via Integer   3.535077
  C Int           15.390511
  Hs Int          25.461061

Relations for floor:
  Prelude          1.000000
  C via Integer    3.853145
  Hs via Integer   4.790661
  C Int           14.547410
  Hs Int          15.272021

Relations for ceiling:
  Prelude          1.000000
  C via Integer    4.234731
  Hs via Integer   2.934738
  C Int           14.128199
  Hs Int          15.183423

Relations for round:
  Prelude          1.000000
  C via Integer    4.380557
  Hs via Integer   1.836171
  C Int           27.479207
  Hs Int           9.352543

Results from viaCDouble:

Relations for properFraction:
  Prelude          1.000000
  C via Integer    2.424372
  Hs via Integer   1.899951
  C Int           17.387781
  Hs Int          28.502385

Relations for truncate:
  Prelude          1.000000
  C via Integer    3.735578
  Hs via Integer   3.533399
  C Int           15.513855
  Hs Int          25.389048

Relations for floor:
  Prelude          1.000000
  C via Integer    3.844556
  Hs via Integer   4.793321
  C Int           14.418142
  Hs Int          15.235541

Relations for ceiling:
  Prelude          1.000000
  C via Integer    4.238978
  Hs via Integer   2.939887
  C Int           14.255595
  Hs Int          15.012826

Relations for round:
  Prelude          1.000000
  C via Integer    4.375220
  Hs via Integer   1.832638
  C Int           27.413318
  Hs Int           9.386865

Results from ncgFloat:

Relations for properFraction:
  Prelude          1.000000
  C via Integer    0.333999
  Hs via Integer   0.499852
  C Int            4.417461
  Hs Int           4.558648

Relations for truncate:
  Prelude          1.000000
  C via Integer    0.496294
  Hs via Integer   0.525865
  C Int            4.311807
  Hs Int           4.556461

Relations for floor:
  Prelude          1.000000
  C via Integer    0.519939
  Hs via Integer   0.558956
  C Int            4.611193
  Hs Int           4.154393

Relations for ceiling:
  Prelude          1.000000
  C via Integer    0.532954
  Hs via Integer   0.572631
  C Int            4.608487
  Hs Int           4.234378

Relations for round:
  Prelude          1.000000
  C via Integer    0.618715
  Hs via Integer   0.575423
  C Int            5.438467
  Hs Int           3.705514

Results from viaCFloat:

Relations for properFraction:
  Prelude          1.000000
  C via Integer    0.334372
  Hs via Integer   0.500771
  C Int            4.479688
  Hs Int           4.716925

Relations for truncate:
  Prelude          1.000000
  C via Integer    0.499481
  Hs via Integer   0.529176
  C Int            4.343629
  Hs Int           4.589870

Relations for floor:
  Prelude          1.000000
  C via Integer    0.520805
  Hs via Integer   0.559927
  C Int            4.631558
  Hs Int           4.216547

Relations for ceiling:
  Prelude          1.000000
  C via Integer    0.532622
  Hs via Integer   0.574791
  C Int            4.635416
  Hs Int           4.266135

Relations for round:
  Prelude          1.000000
  C via Integer    0.619787
  Hs via Integer   0.577242
  C Int            5.477356
  Hs Int           3.731998