Performance/Floating point: Difference between revisions
(+cat) |
(Haskell code markup) |
||
| Line 1: | Line 1: | ||
{{Performance infobox}} | {{Performance infobox}} | ||
[[Category:Performance|Floating point]] | [[Category:Performance|Floating point]] | ||
== Don't use < | == Don't use <hask>Float</hask> == | ||
< | <hask>Float</hask>s (probably 32-bits) are almost always a bad idea, unless you Really Know What You Are Doing. Use <hask>Double</hask>s. There's rarely a speed disadvantage—modern machines will use the same floating-point unit for both. With <hask>Double</hask>s, you are much less likely to hang yourself with numerical errors. | ||
One time when < | One time when <hask>Float</hask> might be a good idea is if you have a ''lot'' of them, say a giant array of <hask>Float</hask>s. An unboxed array of <hask>Float</hask> (see [[Performance/Arrays]]) takes up half the space in the heap compared to an unboxed array of <hask>Double</hask>. However, boxed <hask>Float</hask>s will only take up less space than boxed <hask>Double</hask>s if you are on a 32-bit machine (on a 64-bit machine, a <hask>Float</hask> still takes up 64 bits). | ||
The speed claims may not be true due to Doubles not necessarily being | The speed claims may not be true due to Doubles not necessarily being | ||
| Line 12: | Line 12: | ||
== GHC-specific advice == | == GHC-specific advice == | ||
On x86 (and other platforms with GHC prior to version 6.4.2), use the <tt>-fexcess-precision</tt> flag to improve performance of floating-point intensive code (up to 2x speedups have been seen). This will keep more intermediates in registers instead of memory, at the expense of occasional differences in results due to unpredictable rounding. See the [http://www.haskell.org/ghc/docs/latest/html/users_guide/options-optimise.html#options-f GHC documentation] for more details. Switching on GCCs | On x86 (and other platforms with GHC prior to version 6.4.2), use the <tt>-fexcess-precision</tt> flag to improve performance of floating-point intensive code (up to 2x speedups have been seen). This will keep more intermediates in registers instead of memory, at the expense of occasional differences in results due to unpredictable rounding. See the [http://www.haskell.org/ghc/docs/latest/html/users_guide/options-optimise.html#options-f GHC documentation] for more details. Switching on GCCs <tt>-ffast-math</tt> and <tt>-O3</tt> can also help (use <tt>-optc-ffast-math</tt> and <tt>-optc-O3</tt>). | ||
Where available, the <tt>-optc-march=pentium4 -optc-mfpmath=sse</tt> flags may also help. | Where available, the <tt>-optc-march=pentium4 -optc-mfpmath=sse</tt> flags may also help. | ||
Note that the <tt>-fexcess-precision</tt> flag may make programs behave oddly, | |||
e.g. after falling an < | e.g. after falling an <hask>if x < 0</hask> branch you may find that <hask>x</hask> is now not less than zero, as it has been written out to memory and thus some precision lost in the mean time. | ||
Latest revision as of 08:34, 15 June 2007
| Haskell Performance Resource
Constructs: Techniques: |
Don't use Float
FloatFloats (probably 32-bits) are almost always a bad idea, unless you Really Know What You Are Doing. Use Doubles. There's rarely a speed disadvantage—modern machines will use the same floating-point unit for both. With Doubles, you are much less likely to hang yourself with numerical errors.
One time when Float might be a good idea is if you have a lot of them, say a giant array of Floats. An unboxed array of Float (see Performance/Arrays) takes up half the space in the heap compared to an unboxed array of Double. However, boxed Floats will only take up less space than boxed Doubles if you are on a 32-bit machine (on a 64-bit machine, a Float still takes up 64 bits).
The speed claims may not be true due to Doubles not necessarily being aligned as the machine wishes. We could do with some benchmarking on various platforms to see what's what.
GHC-specific advice
On x86 (and other platforms with GHC prior to version 6.4.2), use the -fexcess-precision flag to improve performance of floating-point intensive code (up to 2x speedups have been seen). This will keep more intermediates in registers instead of memory, at the expense of occasional differences in results due to unpredictable rounding. See the GHC documentation for more details. Switching on GCCs -ffast-math and -O3 can also help (use -optc-ffast-math and -optc-O3).
Where available, the -optc-march=pentium4 -optc-mfpmath=sse flags may also help.
Note that the -fexcess-precision flag may make programs behave oddly,
e.g. after falling an if x < 0 branch you may find that x is now not less than zero, as it has been written out to memory and thus some precision lost in the mean time.