Changes: 24.11#

General

  • Rename the Python package from legate.core to legate. Note that the layout of the package itself, and the names of modules have not changed. This only affects the install command. Where previously a user would write

    $ conda install legate-core

    They now should write

    $ conda install legate

  • Add LEGATE_INLINE_TASK_LAUNCH environment variable to request inline task launch. When enabled, this instructs Legate to forgo the usual Legion task calling convention in favor of launching the task immediately on the submitting thread. This feature is currently considered experimental.

  • Rename bind.sh to legate-bind.sh.

  • Add LEGATE_CUDA_DRIVER environment variable to optionally override the location of the CUDA driver shared library. If not set, defaults to the system-default name of the library (usually “libcuda.so.1”).

C++#

General

  • Remove legate_c.h and all macros defined therein.

  • Remove legate_preamble.h.

  • Deprecate LEGATE_NO_VARIANT, LEGATE_CPU_VARIANT, LEGATE_GPU_VARIANT, and LEGATE_OMP_VARIANT. Users should use the corresponding legate::VariantCode instead.

  • Deprecate LegateVariantCode. Users should use legate::VariantCode instead.

  • Deprecate legate::destroy(). Users should call legate::finish() instead.

  • Add legate::has_finished().

  • Add LocalTaskID and GlobalTaskID, to describe local and global Task ID’s. Use these across the public and private API whenever local or global Task ID’s are requested. Unfortunately, this change could not be made in a backwards-compatible way, and users are therefore required to make the following changes. Users that previously did

    struct MyTask : LegateTask<MyTask> {
  static constexpr auto TASK_ID = 0;
};

runtime->create_task(..., MyTask::TASK_ID);

    Must now either

    struct MyTask : LegateTask<MyTask> {
  static constexpr auto TASK_ID = legate::LocalTaskID{0};

    or

    runtime->create_task(..., legate::LocalTaskID{MyTask::TASK_ID});

  • Add LocalRedopID and GlobalRedopID, to describe local and global reduction operator ID’s.

  • Remove partition.h and restriction.h from the public API

  • Add LEGATE_MPI_WRAPPER environment variable.

Data

  • Improve ScopedAllocator::ScopeAllocator() alignment validation. The alignment is now required to be a strictly positive power of 2. This was technically always the case, but previous versions did not check this, leading to silent failures down the line. The constructor will now validate the alignment and throw an exception if it is improper.

  • Change passing nullptr to ScopedAllocator::deallocate(). This is now a no-op (previously this threw a std::invalid_argument exception).

  • Clarify that ScopedAllocator::allocate(0) will return a nullptr. It always did this, but this behavior is now explicitly documented. Coupled with the change above, it means that all pointers returned from allocate() are safe to pass to deallocate().

Mapping

  • Add NodeRange::hash().

  • Remove mapping::Mapper::set_machine().

  • Add mapping::InstanceMappingPolicy::redundant that forces eager collection of freshly created redundant copies. Add mapping::InstanceMappingPolicy::with_redundant() and mapping::InstanceMappingPolicy::set_redundant() to allow mappers to set new values to the flag.

  • Change enum values for mapping::TaskTarget, mapping::StoreTarget, mapping::AllocPolicy, mapping::InstLayout, and mapping::DimOrdering::Kind. Users should only consider the names of enums (and their members) to be stable, and should not depend on the values themselves.

Partitioning

Tasks

  • Add VariantOptions::elide_device_ctx_sync and VariantOptions::with_elide_device_ctx_sync() to allow specifying that a particular task variant need not perform device context synchronization after task completion.

  • Add TaskContext::get_task_stream() to retrieve the current tasks’ active CUDA stream.

  • Deprecate TaskRegistrar::record_task(std::int64_t std::unique_ptr<TaskInfo>). Users should use LegateTask::Registrar combined with TaskRegistrar::record_all_task() instead.

  • Change TaskRegistrar::register_all_tasks() to take a mutable reference to the library argument instead of by value.

  • Deprecate TaskInfo::has_variant(). Users should use TaskInfo::find_variant() directly instead.

  • Change TaskInfo::find_variant() to return a std::optional. If the optional has a value, the find succeeded and the contained value is the VariantInfo. Otherwise the optional does not contain a value.

  • Add TaskContext::num_scalars() to query the number of Scalar arguments for a task.

  • Move the implementation detail of the CPU communicator (i.e., legate::comm::coll::BackenedNetwork, legate::comm::coll::LocalNetwork, and legate::comm::coll::MPINetwork) to the detail namespace. As a consequence, the following headers are removed from the public interface:backend_network.h, thread_comm.h, local_network.h, and mpi_network.h.

  • Add TaskContext::concurrent_task_barrier() to perform a blocking barrier across all the leaf tasks in a concurrent task launch.

Types

  • Remove ReductionOpKind::DIV and ReductionOpKind::SUB. Partial reduction results are combined in an arbitrary order; since division and subtraction are neither commutative nor associative, it is impossible to use these reliably as reduction operators.

  • Change the return type of legate::array_type from legate::Type to legate::FixedArrayType.

  • Change the return type of legate::struct_type from legate::Type to legate::StructType.

  • Change the return type of legate::list_type from legate::Type to legate::ListType.

  • Change the return type of legate::point_type from legate::Type to legate::FixedArrayType. legate.core.types.point_type now returns legate.core.types.FixedArrayType instead of legate.core.types.Type.

  • Change the return type of legate::rect_type from legate::Type to legate::StructType. legate.core.types.rect_type now returns legate.core.types.StructType instead of legate.core.types.Type.

Runtime

  • Add optional default_options argument to Runtime::create_library() to specify the library-default task variant options.

  • Add optional default_options argument to Runtime::find_or_create_library() to specify the library-default task variant options.

  • Add Library::get_default_variant_options() to retrieve the library-default task variant options.

  • Add Runtime::issue_mapping_fence() to issue a mapping fence that prevents all the downstream tasks from being mapped ahead of the fence.

  • Remove Library::get_mapper_id(). All libraries share the same (internal) underlying Legion mapper, so this method is pointless.

  • Remove Library::register_mapper(). The mapper should be passed when the library is created, and can no longer be changed after the fact.

  • Remove restriction on stores with zero-sized types. It is now possible to create a LogicalStore of type Type::Code::NIL and, for example, empty binary types.

  • Add Runtime::prefetch_bloated_instances() to prefetch instances bloated by a certain amount for a given logical store.

Utilities

  • Deprecate legate::cuda::StreamPool. Users should use TaskContext::get_task_stream() instead within tasks.

  • Deprecate legate::cuda::StreamView. Users should implement their own version of this class.

Python#

General

Data

Mapping

Partitioning

Tasks

  • Add support for default values on scalar arguments in Python tasks.

  • Add support for = None as a default argument for store-type arguments in Python tasks. None is currently the only allowed default value for stores. Users may use any of the supported “optional” type hint variants (x | None, Union[x, None], Optional[x]) to inform the runtime that a store argument may be None.

  • Add support for passing simple Legate Data Interface objects with a single, non-nullable store, as arguments to Python tasks.

Types

  • Add Type.from_python_object() to allow automatic type deduction at runtime. Make type parameters more optional in several cases as a result.

Runtime

  • Add Runtime.prefetch_bloated_instances() to prefetch instances bloated by a certain amount for a given logical store.

Utilities

On this page