This particular thread seems to be more about dynamic allocations of
gsl_vector inside the multimin routines - I wasn't aware that was
happening but in theory it should be possible to rewrite things so there
are no dynamic allocations.
The first thread you linked seems to have a legitimate reason to reduce
the malloc calls for gsl_vector, since he's allocating a large number of
small vectors it seems.
Anyway, Gerard, you have a good grasp of the various issues and what
needs to be done about it - would you be interested in taking the lead
on this and designing a nice API for user-defined alloc functions?
Patrick
On 05/01/2014 07:55 PM, Gerard Jungman wrote:
As before, I apologize for the flood of messages.
But I figure better a few shorter messages than
one long message with multiple sub-topics.
Here is another discussion that I found interesting:
http://comments.gmane.org/gmane.comp.lib.gsl.general/4385
As usual, the details could be debated. But I found the
following quote to be very telling:
"Rhys, I did try to use views. They do not help because the
gsl routines allocate vectors internally and there is not
much that I can do about it... except for maybe hacking
gsl and changing gsl_vector_alloc myself."
People seem to be led to this conclusion from multiple
directions.
Another choice quote:
"The other issue is: the implementation of gsl_vector just
seems inefficient to me. Looking at the code, it seems like
a single vector requires 3 calls to malloc and free (one for
the data, one for the gsl_block, and one for the gsl_vector
itself). The manual states that the block is there for
"consistency", and I can see how memory management becomes
easier with it. But it seems to be a case of generality at
the expense of performance. Also, the stride and the owner
flag are part of the gsl_vector object to make it work with
gsl_views, but then people who never need views pay the
performance price anyway."
