A quick benchmark of two types in Haskell's primitve library.
Have two implemenations of a WordArray. Once builds on top of ByteArray and
the other uses Array Word. I expected the performance of these two
implementations to be identical, but for some reason the Array Word one
seems to be roughly 4 times slower than the ByteArray version.
I am open to the possibilty of this being a bug in my code. If its not, then
Data.Primitve needs fixing.
To build and run the benchmark:
cabal sandbox init
cabal install --dependencies-only
cabal build
dist/build/benchmark-primitive/benchmark-primitive
Running this test with ghc-8.0.1 on a recent Macbook Pro running Debian Linux, I'm getting results like this:
benchmarking copyWordArray 1000/Data.Primitive,Array Word
time 3.438 μs (3.432 μs .. 3.444 μs)
1.000 R² (1.000 R² .. 1.000 R²)
mean 3.437 μs (3.434 μs .. 3.448 μs)
std dev 20.61 ns (10.54 ns .. 40.85 ns)
benchmarking copyWordArray 1000/Data.Primitive.ByteArray
time 612.2 ns (611.3 ns .. 613.2 ns)
1.000 R² (1.000 R² .. 1.000 R²)
mean 611.9 ns (611.2 ns .. 613.1 ns)
std dev 2.910 ns (1.635 ns .. 4.961 ns)
benchmarking copyWordArray 10000/Data.Primitive,Array Word
time 28.71 μs (28.63 μs .. 28.79 μs)
1.000 R² (1.000 R² .. 1.000 R²)
mean 28.67 μs (28.63 μs .. 28.75 μs)
std dev 193.3 ns (117.3 ns .. 346.1 ns)
benchmarking copyWordArray 10000/Data.Primitive.ByteArray
time 6.001 μs (5.992 μs .. 6.012 μs)
1.000 R² (1.000 R² .. 1.000 R²)
mean 6.000 μs (5.994 μs .. 6.014 μs)
std dev 30.21 ns (18.17 ns .. 54.13 ns)
benchmarking copyWordArray 100000/Data.Primitive,Array Word
time 291.5 μs (291.0 μs .. 292.2 μs)
1.000 R² (1.000 R² .. 1.000 R²)
mean 291.8 μs (291.3 μs .. 292.8 μs)
std dev 2.070 μs (1.146 μs .. 3.810 μs)
benchmarking copyWordArray 100000/Data.Primitive.ByteArray
time 60.88 μs (60.86 μs .. 60.90 μs)
1.000 R² (1.000 R² .. 1.000 R²)
mean 60.93 μs (60.91 μs .. 61.00 μs)
std dev 110.9 ns (42.01 ns .. 212.5 ns)