Nice idea but..
Some results:
100 elements:
old method: 00:00:00.0126610 (26x)
stackalloc bit method: 00:00:00.0004750 (1x)
array bit method: 00:00:00.0010700 (2x)
heap bit method: 00:00:00.0038830 (8x)
1,000 elements:
old method: 00:00:00.0139250 (24x)
stackalloc bit method: 00:00:00.0005670 (1x)
array bit method: 00:00:00.0010890 (2x)
heap bit method: 00:00:00.0034920 (6x)
10,000 elements:
old method: 00:00:00.0136110 (12x)
stackalloc bit method: 00:00:00.0011240 (1x)
array bit method: 00:00:00.0016450 (1.4x)
heap bit method: 00:00:00.0043110 (3x)
50,000 elements:
old method: 00:00:00.0175970 (3x)
stackalloc bit method: 00:00:00.0085630 (1.5x)
array bit method: 00:00:00.0055010 (1x)
heap bit method: 00:00:00.0099590 (1.8x)
100,000 elements:
old method: 00:00:00.0210330 (2x)
array bit method: 00:00:00.0100430 (1x)
heap bit method: 00:00:00.0154150 (1.5x)
1,000,000 elements:
old method: 00:00:00.1243730 (1.2x)
array bit method: 00:00:00.0973110 (1x)
heap bit method: 00:00:00.1285650 (1.3x)
10,000,000 elements:
old method: 00:00:00.9252570 (1x)
array bit method: 00:00:00.9632300 (1.05x)
heap bit method: 00:00:01.1098490 (1.20x)
I think the "heap bit" method is not that good :)
JCO
On 3/19/07, Miguel de Icaza <[EMAIL PROTECTED]> wrote:
Hey!
> is that really better than user a uint[] array? I have thinked in
> other solutions... I will test them later.
Well, basically you could use the same code path FindAllStackBits, but
you would have to change it like this:
List <T> FindAllStackBits (uint *bits, Predicate <T> match)
And you would have to pre-allocate bits in the caller:
if (size < 8*8*1024)
bits = stackalloc uint [n];
else
bits = (uint *) Marshal.AllocHGlobal (...);
That way it always uses a fast path, the only difference is that for
large memory blocks we use memory on the heap instead of using stack
memory.
Miguel
> JCO
>
> On 3/19/07, Miguel de Icaza < [EMAIL PROTECTED]> wrote:
> Hello Juan,
>
> > Because there's no way to know the size of the stack in any
> moment in
> > time, Miguel suggested me to limit this new method to use
> only 8KB of
> > the stack... so it will be useful for list with 8 * 1024 * 8
> elements
> > or less. Bigger lists will use the old algorithm.
>
> I just had an idea: if the memory to be allocated is too
> large, instead
> of using stackalloc, you could use:
>
> using System.Runtime.InteropServices ;
>
> IntPtr Marshal.AllocHGlobal (size);
>
> and release it with:
>
> Marshal.FreeHGGlobal (IntPtr p);
>
> That way we could always use the fast mechanisms.
>
> > It also includes few testcases.
> >
> > Juan C. Olivares
> > www.juancri.com
> >
> > On 3/18/07, Juan C. Olivares < [EMAIL PROTECTED]> wrote:
> > I'll check...
> >
> > Juan Cristóbal Olivares
> >
> > www.juancri.com
> >
> > On 3/18/07, Miguel de Icaza <[EMAIL PROTECTED]>
> wrote:
> > Hello,
> >
> > I feel a bit strange that the bools are
> allocated
> > using
> > numbers.Count but the loop uses this._items.
> I
> > rather keep them in
> > parallel.
> >
> > Also, another idea might be to use three
> code
> > paths depending on the
> > number of items:
> >
> > * [0,N] use bools.
> > * [N,M] use bitbools.
> > * [M,infinity] use the old method.
> >
> > Thread stacks could not be very large,
> and you
> > would like to avoid a
> > crash in that condition. Actually, I
> wonder if
> > stackalloc returns a
> > null if the requested size is larger than
> the
> > available stack.
> >
> > Miguel.
> >
> > > Attached.
> > >
> > > Juan C. Olivares
> > > www.juancri.com
> > >
> > > On 3/18/07, Miguel de Icaza <
> [EMAIL PROTECTED]>
> > wrote:
> > > Hello,
> > >
> > > >
> > > > Do you have any comment?. I
> could send a
> > patch...
> > >
> > > These numbers look fantastic. We
> would
> > love to see a patch.
> > >
> > > > best regards
> > > >
> > > > Juan C. Olivares
> > > > www.juancri.com
> > > >
> > > >
> > > >
> > > >
> >
> _______________________________________________
> > > > Mono-list
> > maillist - [email protected]
> > > >
> >
> http://lists.ximian.com/mailman/listinfo/mono-list
> > >
> > >
> > >
> _______________________________________________
> > > Mono-list maillist -
> [email protected]
> > >
> http://lists.ximian.com/mailman/listinfo/mono-list
> >
> >
> >
> >
> >
> >
> >
> >
> >
> >
>
>
>
>
> --
> Juan Cristóbal Olivares
--
Juan Cristóbal Olivares
_______________________________________________
Mono-list maillist - [email protected]
http://lists.ximian.com/mailman/listinfo/mono-list
