Re: [hwloc-devel] [hwloc-svn] svn:hwloc r1865

[Samuel Thibault]

Bert Wesarg, le Sat 27 Mar 2010 15:06:43 +0100, a écrit :
> > +  if (needed_count <= ulongs_count)
> > +    return;
> > +
> > +  while (ulongs_count < needed_count)
> > +    ulongs_count *= 2;

Mmm, this is actually 1 << (hwloc_fls(needed_count)), isn't it?

SAmuel

Re: [hwloc-devel] [hwloc-svn] svn:hwloc r1865

[Bert Wesarg]

Hi,

thanks for sharing this early.

On Sat, Mar 27, 2010 at 08:53,   wrote:
> Modified: branches/dyncpusets/src/cpuset.c
> ==
> --- branches/dyncpusets/src/cpuset.c    (original)
> +++ branches/dyncpusets/src/cpuset.c    2010-03-27 03:53:47 EDT (Sat, 27 Mar 
> 2010)
> @@ -50,10 +63,18 @@
>  struct hwloc_cpuset_s * hwloc_cpuset_alloc(void)
>  {
>   struct hwloc_cpuset_s * set;
> -  set = calloc(sizeof(*set), 1);
> +  set = malloc(sizeof(struct hwloc_cpuset_s));
>   if (!set)
>     return NULL;
>
> +  set->ulongs_count = 1;
> +  set->ulongs = calloc(sizeof(unsigned long), set->ulongs_count);
> +  if (!set->ulongs) {
> +    free(set->ulongs);

That should be free(set).

> +    return NULL;
> +  }
> +
> +  set->infinite = 0;
>  #ifdef HWLOC_DEBUG
>   set->magic = HWLOC_CPUSET_MAGIC;
>  #endif
> @@ -70,29 +91,78 @@
>   set->magic = 0;
>  #endif
>
> +  free(set->ulongs);
>   free(set);
>  }
>
> +/* realloc until it contains at least needed_count ulongs */
> +static void
> +hwloc_cpuset_realloc_by_ulongs(struct hwloc_cpuset_s * set, unsigned 
> needed_count)
> +{
> +  unsigned ulongs_count = set->ulongs_count;
> +  unsigned i;
> +
> +  HWLOC__CPUSET_CHECK(set);
> +
> +  if (needed_count <= ulongs_count)
> +    return;
> +
> +  while (ulongs_count < needed_count)
> +    ulongs_count *= 2;

You may have notices, that I don't like exponential realloc schemes.
On the other hand, I know that you plan to use a sparse implementation
in the future, so this is probably only an intermediate step. Anyway,
in this case, because one ulong should suffice for the common case, a
linear scheme with an increment of 1 would be crazy, I would use an
exponential scheme until some limit and from that on a linear scheme
(with that limit as the increment). The cache line size would be such
limit and a good one I think.

Regards,
Bert