Skip to content

Legate sparse part mgr merge#3

Merged
rohany merged 19 commits intolegate-sparsefrom
legate-sparse-part-mgr-merge
Nov 19, 2022
Merged

Legate sparse part mgr merge#3
rohany merged 19 commits intolegate-sparsefrom
legate-sparse-part-mgr-merge

Conversation

@rohany
Copy link
Copy Markdown
Owner

@rohany rohany commented Nov 19, 2022

No description provided.

bryevdv and others added 19 commits November 7, 2022 11:53
* checkpoint

* only compute PartSym hash once

* don't define unneeded closure on every call

* try caching get_subregion_size?
…lized (nv-legate#465)

* fix for -ll:networks none, we will init MPI if it has not been
initialized.

* add the self mpi finalize

* fix for LEGATE_NEED_NETWORK
* leave override config to client projects

* Add option to have per-file, per-stage test overrides

* remove example custom test
* use bind.sh for all rank detection

* add tests for detected_rank

* add --bind-detail option

* address review comments

* remove Launcher.rank_id

* more careful debug echo

* docs
Co-authored-by: Marcin Zalewski <mzalewski@nvidia.com>
…legate#477)

Previously we would pass this setting through an envvar in
quickstart build.sh, but Legion's cmake workflow overrides this
value, so we need to pass it as a cmake flag.

Co-authored-by: Manolis Papadakis <mpapadakis@nvidia.com>
* Support for library specific annotations

* Address review comments

* Replace itertools.chain with a simpler code
* stdout for debug output

* add --cprofile driver option
* Make the scalar extraction a constant time operation

* Comment about the format for packed return values

* Use a read-write accessor to extract the pointer of a ReturnValue

* Pick the right variant of extract_scalar based on machine configuration
Co-authored-by: Manolis Papadakis <mpapadakis@nvidia.com>
* fix for -ll:networks none, we will init MPI if it has not been
initialized.

* refactor mpi comm

* add the missing file

* refactor local comm

* remove unused files

* always build local comm

* now the both mpi and local comms are working

* move common used functions into the base class

* add a virtual destructor for BackendNetwork base class

* select network based on LEGATE_NEED_NETWORK

* minor fix for local network

* mute printf

* use debug instead of info
* Disallow cmake 3.25.0

* CUDAConfig.ctk can be "none" but not None
* Start assigning unique ids to stores and storages

* Move partition caches to the runtime

* Two mapper changes

* Start using a special Legate functor for identity projections
* Refactor the core mapper's slice_task using dispatch

* Make constructors fetch the unique ids instead of passing them as arguments

* Stop using offsets in linearization
* WIP on cycle detection check

* Don't consider cycles going through sys.modules

* Return True if cycles are found

* Remove some dead code

* Add -legate:cycle-check flag, to perform cycle check at exit

* Print out list indices on cycle listings

* Ignore getset_descriptor objects in cycle check

They seem to be triggering false positives

* Instead just skip the __globals__ pointer

* Skip based on ModuleType, vs inclusion sys.modules

* Print name for functions and class objects

* Ignore cycles going through types

Self-type references and class.__init__.__closure__ seem to produce
a lot of false positives.

* Look for ny Store, not just RegionField

* Filter for more specific Legion handle classes

* Add unit test for cycle checker

* Restrict to RegionField's, as some Futures naturally leak

* Update documentation

Co-authored-by: Manolis Papadakis <mpapadakis@nvidia.com>
* Fixes for mypy>=0.990

* Bump up the mypy version

* Fix the mypy error by inheriting DataclassMixin explicitly
@rohany rohany merged commit 140695f into legate-sparse Nov 19, 2022
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants