AW: AW: [Caml-list] generic Hashtbl.to_array

AW: AW: [Caml-list] generic Hashtbl.to_array

[Christoph Bauer]

[-- Attachment #1: Type: text/plain, Size: 887 bytes --]


 

The dirtiest solution:

let to_array t =
 let a =  Array.make (Hashtbl.length t) (Obj.magic 0)  in
   ignore
     (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0) ; 
     a  
 
 


Does it work correctly for floats? 

 

Looks good for floats. 
 
# let to_array t =
   let a =  Array.make (Hashtbl.length t) (Obj.magic 0)  in
     ignore
       (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0) ;
       a
 
  ;;
val to_array : ('a, 'b) Hashtbl.t -> ('a * 'b) array = <fun>
# let h = Hashtbl.create 0;;
val h : ('_a, '_b) Hashtbl.t = <abstr>
# Hashtbl.add h 1.0 2.0;;
- : unit = ()
# to_array h;;
- : (float * float) array = [|(1., 2.)|]
# Gc.compact ();;
- : unit = ()
#
 
 BTW, the array should store a pointer to a tuple of two floats, so
I thinkt float or ints doesn't matter. I won't use this solution, because
 it isn't better than others. 

 Christoph Bauer 



[-- Attachment #2: Type: text/html, Size: 3244 bytes --]

Re: AW: [Caml-list] generic Hashtbl.to_array

[Tom]

[-- Attachment #1: Type: text/plain, Size: 1361 bytes --]

I'm sorry to say that, but I believe that you results are flawed...

If we look at the code of to_array_1 and to_array_5, there is no possibility
that the former was faster... if nothing else, it has an additional if jump
each and every loop. I simply couldn't believe your results.

Upon inspecting your code with Toploop, I found out some flaws...

let h () =
  let h = Hashtbl.create 100000 in
    for i = 0 to 99999 do            (* <<< not Hashtbl.length h, as it
returns 0 for ampty hashtable *)
      Hashtbl.add h (Random.int max_int) (Random.int max_int);
    done;
    h

let to_array_1 t =
  let dummy =  Array.init 0 (fun _ -> raise Not_found) in
    fst
      (Hashtbl.fold

         (fun k v (a, i) ->
            if i = 0 then
              let a = Array.make (Hashtbl.length t) (k, v) in
                (a, 1)           (* <<<<< Not 0, as it causes no progress *)
            else (a.(i) <- (k, v); (a, i + 1)))
         t (dummy, 0))

I also corrected my implementation:

let mgc = Obj.magic 0      <<< So that the function is executed only once.
let to_array_5 t =
 let a =  Array.make (Hashtbl.length t) mgc in
   ignore
     (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0) ;
     a

I tried to do some benchmarking, but I do not have much time... anyhow, my
implementation is faster as far as I tested it.

Believe in your dreams!

[-- Attachment #2: Type: text/html, Size: 2241 bytes --]

Re: [Caml-list] generic Hashtbl.to_array

[Stéphane Glondu]

We shouldn't talk about Obj.magic, but...

Tom a écrit :
> I also corrected my implementation:
> 
> let mgc = Obj.magic 0      <<< So that the function is executed only once.

Does this provide any benefit? It seems to me that Obj.magic is the
(inlined) identity (so basically Obj.magic 0 is compiled directly into
the integer 0).

-- 
Stéphane

AW: [Caml-list] generic Hashtbl.to_array

[Christoph Bauer]

[-- Attachment #1: Type: text/plain, Size: 4182 bytes --]

 


  _____  

Von: Tom [mailto:tom.primozic@gmail.com] 
Gesendet: Mittwoch, 26. Juli 2006 15:53
An: Christoph Bauer
Betreff: Re: [Caml-list] generic Hashtbl.to_array


Once again...

I'm sorry to say that, but I believe that you results are flawed...


Upon inspecting your code with Toploop, I found out some flaws... 


let h n =
 let m = n * 100000 in
 let h = Hashtbl.create m in

   for i = 0 to m - 1 do       (* <<< not Hashtbl.length h, as it returns 0
for ampty hashtable *)
     Hashtbl.replace h (Random.int <http://random.int/>  max_int) (
Random.int <http://random.int/>  max_int);
   done;
   h


 
 

 
thanks, sorry for this stupid bug.  to_array_5 is now the clear winner.
 
Christoph Bauer
 
 n=8
             Rate      to_array_2 to_array_3 to_array_1c to_array_1
to_array_4 to_array_1b to_array_5
 to_array_2 34.6+-0.4/s         --       -50%        -71%       -72%
-74%        -74%       -79%
 to_array_3 68.8+-0.6/s        99%         --        -42%       -45%
-49%        -49%       -58%
to_array_1c  118+-  2/s       241%        71%          --        -6%
-12%        -13%       -27%
 to_array_1  126+-  0/s       264%        83%          7%         --
-6%         -7%       -22%
 to_array_4  134+-  2/s       287%        95%         14%         6%
--       [-1%]       -17%
to_array_1b  135+-  1/s       290%        96%         15%         7%
[1%]          --       -17%
 to_array_5  162+- 10/s       369%       136%         37%        29%
21%         20%         --
 
 
 (* compile with
 
ocamlopt -o to_array -I benchmark-0.7 unix.cmxa benchmark-0.7/benchmark.cmx
to_array.ml
 
*)
 
open Benchmark
 
let to_array_1 t =
  let dummy =  Array.init 0 (fun _ -> raise Not_found) in
    fst
      (Hashtbl.fold
         (fun k v (a, i) ->
            if i = 0 then  
              let a = Array.make (Hashtbl.length t) (k, v) in
                (a, 1)
            else (a.(i) <- (k, v); (a, i + 1)))
         t (dummy, 0))
 
let to_array_2 t =
  let init _ = fun () -> raise Not_found  in
  let a = Array.init (Hashtbl.length t) init in
    ignore
      (Hashtbl.fold (fun k v i -> a.(i) <- (fun () -> (k, v)); i+1) t 0);
    Array.map (fun f -> f ())  a
 
let to_array_3 t =
  Array.of_list (Hashtbl.fold (fun a b c -> (a, b) :: c) t [])
 

let to_array_1b t =
  let a = ref (Array.init 0 (fun _ -> raise Not_found)) in
    ignore
      (Hashtbl.fold
         (fun k v i ->
            if i = 0 then
              (a := Array.make (Hashtbl.length t) (k, v);
               1)
            else
              ((!a).(i) <- (k, v); i + 1))
         t 0);
    !a
 
 
 
let to_array_4 t =
  let init = ref None in
    begin try Hashtbl.iter (fun k v -> init := Some (k,v); raise Exit) t
      with Exit -> ()
    end;
    match !init with
      | None -> [| |]
      | Some i ->
          let a = Array.make (Hashtbl.length t) i in
            ignore (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0);
            a
 
 
 
let to_array_5 =
  let init = Obj.magic 0 in
    fun t ->
      let a =  Array.make (Hashtbl.length t) init  in
        ignore
          (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0) ; 
        a 
 
let to_array_1c t =
  let r =
    Hashtbl.fold (fun k v seed ->
                    match seed with
                 Some (a,i) -> a.(i) <- (k,v); Some (a,i+1)
                      | None -> let a =  Array.make (Hashtbl.length t) (k,v)
in
                                  Some (a,1))
      t None 
  in
    match r with
        None -> Array.init 0 (fun _ -> raise Not_found)
      | Some (a, _) -> a
 

      
let h n = 
  let m = n * 1000 in
  let h = Hashtbl.create m in
    for i = 0 to m - 1 do
      Hashtbl.replace h (Random.int max_int) (Random.int max_int);
    done;
    h
      
let main () =
  let n = try int_of_string Sys.argv.(1) with _ -> 1 in
  let h = h n in
  let res = throughputN ~repeat:5 1
    [("to_array_1", to_array_1, h);
     ("to_array_1b", to_array_1b, h);
     ("to_array_1c", to_array_1c, h);
     ("to_array_2", to_array_2, h);
     ("to_array_3", to_array_3, h);
     ("to_array_4", to_array_4, h);
     ("to_array_5", to_array_5, h);
 
      ] in
      tabulate res
 

let () =  main () 


[-- Attachment #2: Type: text/html, Size: 12793 bytes --]

Re: [Caml-list] generic Hashtbl.to_array

[Tom]

[-- Attachment #1: Type: text/plain, Size: 53 bytes --]

At last the result I desire (and have predicted ;) )

[-- Attachment #2: Type: text/html, Size: 57 bytes --]

generic Hashtbl.to_array

[oleg]

I wonder about the following solution. At least it traverses the
hashtable exactly once (and it does not ignore the result of the
fold).


let to_array9 t =
  let Some (a,_) =
    Hashtbl.fold (fun k v seed ->
      match seed with
	Some (a,i) -> a.(i) <- (k,v); Some (a,i+1)
      | None -> let a =  Array.make (Hashtbl.length t) (k,v) in
                Some (a,1))
      t None
  in a
;;

Re: [Caml-list] generic Hashtbl.to_array

[Damien Doligez]

On 2006-07-26, at 04:16, oleg@pobox.com wrote:

> let to_array9 t =
>   let Some (a,_) =
>     Hashtbl.fold (fun k v seed ->
>       match seed with
> 	Some (a,i) -> a.(i) <- (k,v); Some (a,i+1)
>       | None -> let a =  Array.make (Hashtbl.length t) (k,v) in
>                 Some (a,1))
>       t None
>   in a
> ;;

It fails whenever the hash table is empty, and the compiler
warns you about it.  Replace your "let Some (a,_) = ..."
with a pattern-matching and you have a nice solution.

-- Damien

AW: [Caml-list] generic Hashtbl.to_array

[Christoph Bauer]

> 
> let to_array_4 t =
>    let init = ref None in
>    begin try Hashtbl.iter (fun k v -> init := Some (k,v); 
> raise Exit) t

My guess: hashtbl has to loop over the first empty buckets.
And this eats the cpu cycles.

>    with Exit -> ()
>    end;
>    match !init with
>    | None -> [| |]
>    | Some i ->
>      let a = Array.make (Hashtbl.length t) i in
>        ignore (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i 
> + 1) t 0);
>        a

Regards,
Christoph Bauer

Re: [Caml-list] generic Hashtbl.to_array

[Tom]

[-- Attachment #1: Type: text/plain, Size: 180 bytes --]

The dirtiest solution:

let to_array t =
 let a =  Array.make (Hashtbl.length t) (Obj.magic 0)  in
   ignore
     (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0) ;
     a

[-- Attachment #2: Type: text/html, Size: 319 bytes --]

Re: [Caml-list] generic Hashtbl.to_array

[Stéphane Glondu]

I know I am late.

Tom a écrit :
> The dirtiest solution:
> 
> let to_array t =
>  let a =  Array.make (Hashtbl.length t) (Obj.magic 0)  in
>    ignore
>      (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0) ;
>      a

What about:

let to_array h =
  let res = ref [||] in
  let rec assign = ref (fun i x ->
			  res := Array.make (Hashtbl.length h) x;
			  !res.(i) <- x;
			  assign := Array.set !res) in
  ignore (Hashtbl.fold (fun k v i -> !assign i (k, v); i+1) h 0);
  !res;;

Sorry if it has already been proposed (but I have not seen it).

-- 
Stéphane Glondu

Re: [Caml-list] generic Hashtbl.to_array

[Christoph Bauer]

Hi,

> let to_array t =
>    let init = ref None in
>    begin try Hashtbl.iter (fun k v -> init := Some (k,v); 
> raise Exit) t
>    with Exit -> ()
>    end;
>    match !init with
>    | None -> [| |]
>    | Some i ->
>      let a = Array.make (Hashtbl.length t) i in
>      ignore (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0);
>      a
> ;;

it's curious, but this solution is slower than the others!

[skaller's solution seems to be the same, so I
include only this one in the "benchmark"]

             Rate     to_array_4  to_array_3 to_array_1b  to_array_2
to_array_1
 to_array_4 407+-0/s          --        -16%        -16%        -17%
-17%
 to_array_3 486+-2/s         19%          --       [-0%]       [-1%]
-1%
to_array_1b 487+-0/s         20%        [0%]          --       [-0%]
-1%
 to_array_2 489+-2/s         20%        [1%]        [0%]          --
-1%
 to_array_1 491+-0/s         21%          1%          1%          1%
--

from http://ocaml-benchmark.sourceforge.net/doc/Benchmark.html

Benchmark.tablulate results prints a comparison table for a list of results
obtained by Benchmark.latencyN or Benchmark.throughputN with each function
compared to all the others. The table is of the type 


              Rate name1 name2 ...   OR          s/iter name1 name2 ...
        name1  #/s    --   r12             name1   #       --   r12
        name2  #/s   r21    --             name2   #      r21    --
        ...                                ...                            

where name1, name2,... are the labels of the tests sorted from slowest to
fastest and rij says how much namei is faster (or slower if < 0) than namej
(technically it is equal to (ri - rj) expressed in percents of rj where ri
and rj are the rates of namei and namej respectively). 

If several results are associated to a given name, they are used to compute
a Student's statistic to check whether the rates are significantly
different. If ri and rj are not believed to be different, rij will be
printed between brackets.

(* compile with

ocamlopt -o to_array -I benchmark-0.7 unix.cmxa benchmark-0.7/benchmark.cmx
to_array.ml

*)

open Benchmark

let to_array_1 t =
  let dummy =  Array.init 0 (fun _ -> raise Not_found) in
    fst
      (Hashtbl.fold
         (fun k v (a, i) ->
            if i = 0 then  
              let a = Array.make (Hashtbl.length t) (k, v) in
                (a, 0)
            else (a.(i) <- (k, v); (a, i + 1)))
         t (dummy, 0))

let to_array_2 t =
  let init _ = fun () -> raise Not_found  in
  let a = Array.init (Hashtbl.length t) init in
    ignore
      (Hashtbl.fold (fun k v i -> a.(i) <- (fun () -> (k, v)); i+1) t 0);
    Array.map (fun f -> f ())  a

let to_array_3 t =
  Array.of_list (Hashtbl.fold (fun a b c -> (a, b) :: c) t [])


let to_array_1b t =
  let a = ref (Array.init 0 (fun _ -> raise Not_found)) in
    ignore
      (Hashtbl.fold
         (fun k v i ->
            if i = 0 then
              (a := Array.make (Hashtbl.length t) (k, v);
               i)
            else
              ((!a).(i) <- (k, v); i + 1))
         t 0);
    !a



let to_array_4 t =
   let init = ref None in
   begin try Hashtbl.iter (fun k v -> init := Some (k,v); raise Exit) t
   with Exit -> ()
   end;
   match !init with
   | None -> [| |]
   | Some i ->
     let a = Array.make (Hashtbl.length t) i in
       ignore (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0);
       a



let h () = 
  let h = Hashtbl.create 100000 in
    for i = 0 to (Hashtbl.length h) do
      Hashtbl.add h (Random.int max_int) (Random.int max_int);
    done;
    h
      
let main () =
  let h = h () in
  let res = throughputN ~repeat:5 1
    [("to_array_1", to_array_1, h);
     ("to_array_1b", to_array_1b, h);
     ("to_array_2", to_array_2, h);
     ("to_array_3", to_array_3, h);
     ("to_array_4", to_array_4, h); ] in
      tabulate res


let () =  main () 

Re: [Caml-list] generic Hashtbl.to_array

[skaller]

On Tue, 2006-07-25 at 14:00 +0200, Christoph Bauer wrote:
> Hi,
> 
> > let to_array t =
> >    let init = ref None in
> >    begin try Hashtbl.iter (fun k v -> init := Some (k,v); 
> > raise Exit) t
> >    with Exit -> ()
> >    end;
> >    match !init with
> >    | None -> [| |]
> >    | Some i ->
> >      let a = Array.make (Hashtbl.length t) i in
> >      ignore (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0);
> >      a
> > ;;
> 
> it's curious, but this solution is slower than the others!
> 
> [skaller's solution seems to be the same, so I
> include only this one in the "benchmark"]

This is not entirely surprising, since Ocaml wastes time 
first initialising the array.. and then assigning the same 
cells new values, recycling the same store through the 
cache twice. However there is no need for a secondary
data structure, which may delay hitting limits of the
next level of caching (for example VM paging disk).

I see in your tests the number of elements is 100,000.
It would be interesting to increase this in a series
of tests to see if the performance tradeoffs change:
cache effects would predict a 'kink' when you hit
cache boundaries?

-- 
John Skaller <skaller at users dot sf dot net>
Felix, successor to C++: http://felix.sf.net

generic Hashtbl.to_array

[Christoph Bauer]

Hi,

what is the best way to write Hashtbl.to_array?

Hashtbl.to_array : ('a, 'b) Hashtbl.t -> ('a * 'b) array

The simples idea has the problem, that you don't have
a initial value to make the result array:

let to_array t =
  let a =  Array.init make (Hashtbl.length t) ?init?  in
    ignore
      (Hashtbl.fold
         (fun k v i ->
            a.(i) <- (k, v); i + 1)
         t 0);
    a

The best solution I found is

let to_array t =
  let dummy =  Array.init 0 (fun _ -> raise Not_found) in
    fst
      (Hashtbl.fold
         (fun k v (a, i) ->
            if i = 0 then  
              let a = Array.make (Hashtbl.length t) (k, v) in
                (a, 0)
            else (a.(i) <- (k, v); (a, i + 1)))
         t (dummy, 0))

Is there a better one?

Thanks,
Christoph Bauer

Re: [Caml-list] generic Hashtbl.to_array

[Erick Tryzelaar]

Christoph Bauer wrote:
> Hi,
>
> what is the best way to write Hashtbl.to_array?
>
> Hashtbl.to_array : ('a, 'b) Hashtbl.t -> ('a * 'b) array
>
> The simples idea has the problem, that you don't have
> a initial value to make the result array:

The easiest is to use a temporary list:

# let x = Hashtbl.create 2;;
val x : ('_a, '_b) Hashtbl.t = <abstr>
# Hashtbl.add x 5 3;;
- : unit = ()
# Hashtbl.add x 7 2;;
- : unit = ()
# Array.of_list (Hashtbl.fold (fun a b c -> (a, b) :: c) x []);;
- : (int * int) array = [|(7, 2); (5, 3)|]


You could also try inverting the Hashtbl fold into an iterator+closure 
and pass the closure into the Array.init function, but I'm not sure how 
complicated/efficient that would be.

I suppose it just depends on how efficient you need it to be. If it's 
just some simple stuff, I'd just use the intermediary list.

-e

Re: [Caml-list] generic Hashtbl.to_array

[Damien Doligez]

Hello,

On 2006-07-25, at 10:29, Christoph Bauer wrote:

> The simples idea has the problem, that you don't have
> a initial value to make the result array:

You can get it from the hash table itself:

let to_array t =
   let init = ref None in
   begin try Hashtbl.iter (fun k v -> init := Some (k,v); raise Exit) t
   with Exit -> ()
   end;
   match !init with
   | None -> [| |]
   | Some i ->
     let a = Array.make (Hashtbl.length t) i in
     ignore (Hashtbl.fold (fun k v i -> a.(i) <- (k, v); i + 1) t 0);
     a
;;

-- Damien