Agency 0.3.0

Summary

Various changes driven by ISO C++ proposal P0443.

Breaking Changes

• bulk_sync_execute has been eliminated
• is_bulk_synchronous_executor has been eliminated
• bulk_async_execute has been eliminated
• is_bulk_asynchronous_executor has been eliminated
• async_execute has been eliminated
• is_asynchronous_executor has been eliminated
• sync_execute has been eliminated
• is_synchronous_executor has been eliminated
• bulk_then_execute has been eliminated
• is_bulk_continuation_executor has been eliminated
• then_execute has been eliminated
• is_continuation_executor has been eliminated
• is_bulk_executor has been eliminated
• is_simple_executor has been eliminated
• future_value and future_value_t have been renamed future_result and future_result_t, respectively.
• executor_execution_category and executor_execution_category_t have been replaced with the bulk_guarantee executor property
• execution_categories.hpp and the functional therein has been eliminated
• execution_agent_traits<A>::execution_category has been replaced with execution_agent_traits<A>::execution_requirement
• cuda::deferred_future has been eliminated

New Features

• require and properties:
  • bulk
  • single
  • then
  • always_blocking
  • bulk_guarantee
• basic_span

Containers

• cuda::vector

Executors

• Various executors now have equality operations.

Utilities

• pointer_adaptor
• cuda::device_ptr
• cuda::scoped_device

New Experimental Features

Utilities

• experimental::domain
• overload

Known Issues

• #437 Nvcc emits warnings regarding standard library functions calling __host__-only functions

Resolved Issues

• #428 Warnings regarding ignored CUDA annotations have been eliminated

Agency 0.2.0

Summary

Agency 0.2.0 introduces new components for creating parallel C++ programs. A suite of new containers allow easy management of collections of objects in parallel programs. agency::array and agency::vector provide familiar C++ components in CUDA codes while components like shared and shared_vector allow groups of concurrent execution agents to cooperatively own an object. New executors target CUDA cooperative kernels, OpenMP, loop unrolling, and polymorphism. Finally, new speculative features allow programmers to experiment with interfacing with native CUDA APIs, multidimensional arrays, and range-based programming.

Breaking Changes

• cuda::split_allocator has been renamed cuda::heterogeneous_allocator

New Features

Control Structures

• async : Composes with an executor to create a single asynchronous function invocation.
• invoke : Composes with an executor to create a single synchronous function invocation.

Containers

• array : Statically-sized object container based on std::array.
• vector: Dynamically-sized object container based on std::vector.
• shared: Container for a single object shared by a group of concurrent execution agents.
• shared_array : Container for a statically-sized collection of objects shared by a group of concurrent execution agents.
• shared_vector: Container for a dynamically-sized collection of objects shared by a group of concurrent execution agents.
• tuple : A product type based on std::tuple.

Execution Policies

• concurrent_execution_policy_2d : Induces a two-dimensional group of concurrent execution agents.
• sequenced_execution_policy_2d : Induces a two-dimensional group of sequenced execution agents.
• parallel_execution_policy_2d : Induces a two-dimensional group of parallel execution agents.
• unsequenced_execution_policy_2d : Induces a two-dimensional group of unsequenced execution agents.
• OpenMP-specific execution policies
  • omp::parallel_execution_policy : Induces a group of parallel execution agents using an OpenMP executor.
  • omp::unsequenced_execution_policy : Induces a group of unsequenced execution agents using an OpenMP executor.

Executors

• cuda::concurrent_grid_executor : Creates concurrent-concurrent execution agents using CUDA 9's cooperative grid launch.
• omp::parallel_for_executor AKA omp::parallel_executor : Creates parallel execution agents using OpenMP's parallel for loop.
• omp::simd_executor AKA omp::unsequenced_executor : Creates unsequenced execution agents using OpenMP's SIMD for loop.
• experimental::unrolling_executor : Creates sequenced execution agents using an unrolled for loop.
• variant_executor : Creates execution agents with polymorphic execution guarantees using an dynamic, underlying executor.

Utilities

• cuda::device : Creates a cuda::device_id from a device enumerant.
• cuda::devices: Creates a collection of cuda::device_id from a sequence of device enumerants.
• cuda::all_devices: Creates a collection of cuda::device_id corresponding to all devices in the system.

New Experimental Features

Containers

• experimental::ndarray : Dynamically-sized multidimensional object container.
• experimental::ndarray_ref: View of a multidimensional object container.

Execution Policies

• cuda::experimental::static_grid : Induces a statically-sized group of parallel-concurrent execution agents using CUDA grid launch.
• cuda::experimental::static_con : Indicues a statically-sized group of concurrent execution agents using CUDA grid launch.

Utilities

• cuda::experimental::make_async_future : Creates a cuda::async_future from underlying CUDA resources.
• cuda::experimental::make_dependent_stream : Creates a cudaStream_t from a cuda::async_future.
• Fancy ranges
  • experimental::interval() : Creates a range of integers specified by two end points.
  • experimental::iota_view : A range of increasing integers.
  • experimental::transformed_view : A transformed view of an underlying range.
  • experimental::zip_with_view : A zipped-and-then-transformed view of multiple underlying ranges.

New Examples

• fork_executor.cpp

Known Issues

• #428 nvcc 9.0 emits spurious warnings regarding ignored annotations

Resolved Issues

• #347 Various warnings at aggressive reporting levels have been eliminated
• #289 async_future::bulk_then() needs to schedule the outer_arg's destruction
• #352 .rank() generates results in the wrong order

Acknowledgments

• Thanks to Steven Dalton and other Github users for submitting bug reports.
• Thanks to Steven Dalton, Michael Garland, Mike Bauer, Isaac Gelado, Saurav Muralidharan, and Cris Cecka for continued input into Agency's overall design and implementation.

Agency v0.1.0

Summary

Agency 0.1.0 introduces new control structures such as bulk_invoke() for creating parallel tasks. A suite of new execution policies compose with these control structures to require different kinds of semantic guarantees from the created tasks. A new library of executors controls the mapping of tasks onto underlying execution resources such as CPUs, GPUs, and collections of multiple GPUs. In addition to these basic components, this release also introduces experimental support for a collection of utility types useful for creating Agency programs.

New Features

New Control Structures

• bulk_invoke
• bulk_async
• bulk_then

New Execution Policies

• concurrent_execution_policy
• sequenced_execution_policy
• parallel_execution_policy
• unsequenced_execution_policy
• CUDA-specific execution policies
  • cuda::concurrent_execution_policy
  • cuda::parallel_execution_policy
  • cuda::grid

New Executors

• concurrent_executor
• executor_array
• flattened_executor
• parallel_executor
• scoped_executor
• sequenced_executor
• unsequenced_executor
• vector_executor
• CUDA-specific executors
  • cuda::block_executor
  • cuda::concurrent_executor
  • cuda::grid_executor
  • cuda::grid_executor_2d
  • cuda::multidevice_executor
  • cuda::parallel_executor

New Experimental Utilities

• experimental::array
• experimental::bounded_integer
• experimental::optional
• experimental::short_vector
• experimental::span
• Fancy ranges based on the range-v3 library
  • experimental::chunk_view
  • experimental::counted_view
  • experimental::stride_view
  • experimental::zip_view

New Examples

• concurrent_ping_pong.cpp
• concurrent_sum.cpp
• fill.cpp
• hello_async.cpp
• hello_lambda.cpp
• hello_then.cpp
• hello_world.cpp
• ping_pong_tournament.cpp
• saxpy.cpp
• version.cpp
• CUDA-specific example programs
  • async_reduce.cu
  • black_scholes.cu
  • hello_device_lambda.cu
  • multigpu_saxpy.cu
  • saxpy.cu
  • simple_on.cu
  • transpose.cu

Known Issues

• #255 Agency is not known to work with any version of the Microsoft Compiler
• #256 Agency is not known to work with NVIDIA Compiler versions prior to 8.0
• #257 Agency is not known to work with NVIDIA GPU architectures prior to sm_3x

Acknowledgments

• Thanks to Michael Garland for significant input into Agency's overall design.
• Thanks to the NVIDIA compiler team, especially Jaydeep Marathe, for enhancements to nvcc's C++ support.
• Thanks to Steven Dalton, Mark Harris, and Evghenni Gaburov for testing this release during development.
• Thanks to Duane Merrill and Sean Baxter for design feedback.
• Thanks to Olivier Giroux for contributing an implementation of synchronic.