GitLab

Missing commit from merge request !145.

[feature] duplicate operator for layouts

See merge request !145

Add a function to create a layout from an existing one. We call it
duplicate to avoid the confusion with dma copies.

- ownership consistent with destroy operator: underlying layouts are
  also copied
- data is copied by the implementation, and the generic code takes care of
  the ops pointer.

Now that we have unit tests in multiple directories, it's easier to just
change the ignore rules to match .log and .trs files across the entire
repo.

Add valgrind checks to the CI

See merge request !143

Verbose logs and one single OPENMP thread to avoid weird valgrind
scheduling issues.

Fix #68

Limit the size of the memory tested to avoid tricky overflows.

Detected by valgrind.

return-early test always succeeding for the wrong raison.

target_dims is not the same size as the dimensions of this layout.

Detected by valgrind, but without consequences.

nanosleep is not guaranteed to sleep under valgrind, which make our
async test code run forever. Change for a random amount of work in a
loop, that's somehow better.

Generic layout_destroy

See merge request !144

Refactor destroy API for layouts so that we can solve a long standing
ownership issue with complex layouts. Putting the destroy operator
inside the implementation-specific ops allows us to call a generic
destroyer on embedded layouts.

Note that this introduce a ownership policy for pad & reshape layouts
that take ownership of the targeted layout at creation time.

Related to #55.

ECP-CI Fixes for Theta

See merge request !142

Extra recursive target across the repository to build the test programs
without running them. Allows us to cross-compile both the lib and the
test programs on frontend systems, without launching on the
compute nodes.

Make us of that to fix the theta ECP-CI job using icc.

Some implementations require the file descriptor to be -1 when asking
for an anonymous mmap, making this test fail. The fix is to test both
anonymous mappings and file-backed mappings separately.

Some filesystems do not support the full posix range of file-backed
mmaps too (parallel file systems), so the file used for tests must be on
a simpler filesystem.

DMA Cuda

See merge request !140

Nicolas Denoyelle authored Jul 10, 2020 and Swann Perarnau committed Jul 10, 2020

Simple copy operator and dma engine for CUDA.
Requires cuda layouts to work.

Chimera ci

See merge request !139

Nicolas Denoyelle authored Jul 09, 2020 and Swann Perarnau committed Jul 09, 2020

Fix benchmark checks for cuda version too.

Layout cuda fix

See merge request !137

Layout cuda

See merge request !81

hwloc topology query in aml

See merge request !136

Nicolas Denoyelle authored May 29, 2020 and Swann Perarnau committed May 29, 2020

Compute a distance matrix no matter what.
It never happens that the system has a distance matrix between
something else that NUMANODES. Sometimes this matrix does not even
exist.
This commit will compute a hop distance matrix if the system does
not have any distance matrix. It will also reshape existing distance
matrix to match the distance query.

Add hwloc area by performance criterion

See merge request !135

Nicolas Denoyelle authored May 19, 2020 and Swann Perarnau committed May 19, 2020

Area that will allocate similarly to mbind_preferred,
but using numa sorted by a performance criterion.

Benchmarks Update

See merge request !86

Florence Monna authored May 19, 2020 and Swann Perarnau committed May 19, 2020

BLAS Level 1 benchmarks with flat and tile examples are implemented, except for the rotations generator functions.

basic hwloc area

See merge request !133

Provide a nix shell for the repository

See merge request !132

Use the local nix shell for the CI steps that can be migrated. Some of
the build steps could be refactored further by using the argopkgs "real
package".

Full nix definition of a shell environment capable of compiling,
debugging, style-check, and building docs for AML.

Package definition pinned from the argopkgs master.

[ci/fix] quiet git-clang-format to avoid errors

See merge request !131

If git-clang-format is not quiet, it's output changes between "no
modified files to format" and "clang-format did not modify files"
depending on whether the branch contains files to inspect but properly
formatted or no files to inspect at all. This breaks the CI step as is.

[ci/fix] bad file/regex for clang-format CI

See merge request !130

Forgot to double check output of the CI based on clang-format version
and actual diff output.