coll/base: fix memory allocation in mca_coll_base_alltoall_intra_basi…

[ggouaillardet]

…c_inplace

(cherry picked from commit open-mpi/ompi@0f23037)

[mellanox-github]

Test PASSed.
See http://bgate.mellanox.com/jenkins/job/gh-ompi-release-v1.8-pr/870/ for details.

[ggouaillardet]

:bot:assign: @bosilca
:bot🏷️bug
:bot:milestone:v2.0.0

[jsquyres]

@bosilca gave this a 👍 in #687.

[bosilca]

Actually, I was holding on this one, because I think the whole allocation of memory for collective operations is wrong. But, I need a little bit more time to think about

[jsquyres]

@bosilca Doh. Should I revert?

[bosilca]

Leave it as is by now. Once I have a better understanding of the issue, I
will try to propose a cleaner approach.

On Mon, Oct 19, 2015 at 1:52 PM, Jeff Squyres [email protected]
wrote:

@bosilca https://github.com/bosilca Doh. Should I revert?

—
Reply to this email directly or view it on GitHub
#686 (comment)
.

[bosilca]

The more I look at this patch, the more confused I get. The only thing this change does is slightly reducing the allocated memory. However, now the allocated memory and the max_size are out of sync, as one includes the gap while the other doesn't. In addition the allocated buffer is not consistent with the amount of data allocated, as it has not been moved back by the true_lb (as in the other functions).

[bosilca]

I just pushed a PR to clean the handling of temporary buffers (mainly in collectives). Please take a look @ open-mpi/ompi#1047. However, it is just a bandaid, the root of all these problems still exists.

[ggouaillardet]

my patch does change the size of the allocated memory.
if extent < true_extent (in Fortran, think of a matrix row), the allocated memory is slightly increased
i agree this patch does not handle non-zero true_lb.
i will review open-mpi/ompi#1047 from now

[bosilca]

You're right, in the case where the extent was altered by a resize you used to formula that gives the right span of the data.