Prolog是否有像Haskell这样的别名“运算符”?

时间:2015-12-24 22:08:40

标签: performance haskell prolog alias

在Haskell中,有一种称为“as”-operator的语言特性(有时称为别名)。这个想法如下:假设您有一个函数,例如将列表作为输入并希望返回所有尾部,您可以将其实现为:

tails a@(_:xs) = a : tails xs
tails [] = [[]]

@确保您既可以引用整个参数,也可以引用参数结构的某些部分。这是智能的性能(它更像是性能破解,因为在第一行的主体中重构数组(x:xs)),如果没有被编译器优化,将导致分配新对象,修改字段等。有关详细信息,请参阅here

我想知道Prolog是否有相同的东西:例如,如果你想在Prolog中实现尾部,可以通过以下方式完成:

tails([H|T],[[H|T]|TA]) :-
    tails(T,TA).
tails([],[[]]).

但如果有一个“as”-operator就可以提高效率:

tails(L@[_|T],[L|TA]) :-  %This does not compile
    tails(T,TA).
tails([],[[]]).

是否有任何此类构造或语言扩展?

1 个答案:

答案 0 :(得分:9)

TL; DR:好主意 1 !加速似乎限于约20%(对于大多数列表大小)。

在这个答案中,我们比较了与@不同的三种不同谓词 - 如数据重用:

list_tails([], [[]]).                % (1) like `tails/2` given by the OP ...
list_tails([E|Es], [[E|Es]|Ess]) :-  %     ....... but with a better name :-)
   list_tails(Es, Ess).

list_sfxs1(Es, [Es|Ess]) :-          % (2) "re-use, mutual recursion"
   aux_list_sfxs1(Es, Ess).          %     "sfxs" is short for "suffixes"

aux_list_sfxs1([], []).
aux_list_sfxs1([_|Es], Ess) :-
   list_sfxs1(Es, Ess).

list_sfxs2([], [[]]).                % (3) "re-use, direct recursion"
list_sfxs2(Es0, [Es0|Ess]) :-
   Es0 = [_|Es],
   list_sfxs2(Es, Ess).

要测量运行时间,我们使用以下代码:

:-( dif(D,sicstus), current_prolog_flag(dialect,D)
  ; use_module(library(between))
  ).

run_benchs(P_2s, P_2, L, N, T_ms) :-
   between(1, 6, I),
   L is 10^I,
   N is 10^(8-I),
   length(Xs, L),
   member(P_2, P_2s),
   garbage_collect,
   call_walltime(run_bench_core(P_2,Xs,N), T_ms).

run_bench_core(P_2, Xs, N) :-
   between(1, N, _),
   call(P_2, Xs, _),
   false.
run_bench_core(_, _, _).

要衡量 2 ,我们会使用call_walltime/2 - call_time/2的变体:

call_walltime(G, T_ms) :-
   statistics(walltime, [T0|_]),
   G,
   statistics(walltime, [T1|_]),
   T_ms is T1 - T0.

让我们将代码变体置于测试之上...

  • ...使用不同的列表长度L ...
  • ...并多次运行每次测试N(为了更好的准确性)。

首先,我们使用版本7.3.14(64位):

?- run_benchs([list_sfxs1,list_sfxs2,list_tails], P_2, L, N, T_ms).
   P_2 = list_sfxs1, L*N = 10*10000000, T_ms =  7925
;  P_2 = list_sfxs2, L*N = 10*10000000, T_ms =  7524
;  P_2 = list_tails, L*N = 10*10000000, T_ms =  6936
;
   P_2 = list_sfxs1, L*N = 100*1000000, T_ms =  6502
;  P_2 = list_sfxs2, L*N = 100*1000000, T_ms =  5861
;  P_2 = list_tails, L*N = 100*1000000, T_ms =  5618
;
   P_2 = list_sfxs1, L*N = 1000*100000, T_ms =  6434
;  P_2 = list_sfxs2, L*N = 1000*100000, T_ms =  5817
;  P_2 = list_tails, L*N = 1000*100000, T_ms =  9916
;
   P_2 = list_sfxs1, L*N = 10000*10000, T_ms =  6328
;  P_2 = list_sfxs2, L*N = 10000*10000, T_ms =  5688
;  P_2 = list_tails, L*N = 10000*10000, T_ms =  9442
;
   P_2 = list_sfxs1, L*N = 100000*1000, T_ms = 10255
;  P_2 = list_sfxs2, L*N = 100000*1000, T_ms = 10296
;  P_2 = list_tails, L*N = 100000*1000, T_ms = 14592
;
   P_2 = list_sfxs1, L*N = 1000000*100, T_ms =  6955
;  P_2 = list_sfxs2, L*N = 1000000*100, T_ms =  6534
;  P_2 = list_tails, L*N = 1000000*100, T_ms =  9738.

然后,我们使用版本4.3.2(64位)重复上一个查询 3

?- run_benchs([list_sfxs1,list_sfxs2,list_tails], P_2, L, N, T_ms).
   P_2 = list_sfxs1, L*N = 10*10000000, T_ms =  1580
;  P_2 = list_sfxs2, L*N = 10*10000000, T_ms =  1610
;  P_2 = list_tails, L*N = 10*10000000, T_ms =  1580
;
   P_2 = list_sfxs1, L*N = 100*1000000, T_ms =   710
;  P_2 = list_sfxs2, L*N = 100*1000000, T_ms =   750
;  P_2 = list_tails, L*N = 100*1000000, T_ms =   840
;
   P_2 = list_sfxs1, L*N = 1000*100000, T_ms =   650 
;  P_2 = list_sfxs2, L*N = 1000*100000, T_ms =   660
;  P_2 = list_tails, L*N = 1000*100000, T_ms =   740
;  
   P_2 = list_sfxs1, L*N = 10000*10000, T_ms =   620
;  P_2 = list_sfxs2, L*N = 10000*10000, T_ms =   650
;  P_2 = list_tails, L*N = 10000*10000, T_ms =   740
;
   P_2 = list_sfxs1, L*N = 100000*1000, T_ms =   670
;  P_2 = list_sfxs2, L*N = 100000*1000, T_ms =   650
;  P_2 = list_tails, L*N = 100000*1000, T_ms =   750
;
   P_2 = list_sfxs1, L*N = 1000000*100, T_ms = 12610
;  P_2 = list_sfxs2, L*N = 1000000*100, T_ms = 12560
;  P_2 = list_tails, L*N = 1000000*100, T_ms = 33460.

摘要:

  • alias-thingy 可以和显着提高性能
  • 在上述测试中,与SWI-Prolog相比,SICStus Prolog 4 提供 10X加速

脚注1:为什么将(@)/2置于规则头部的特技? 最终得到非idiomatic Prolog代码?
脚注2:我们对总运行时间感兴趣。为什么?因为垃圾收集成本显示更大的数据量! 脚注3:为了便于阅读,答案序列已经过后处理 脚注4:release 4.3.0起可用。目前的目标架构包括IA-32AMD64