New Orleans, Louisiana – November 16, 2014 – The OpenMP® Architecture Review Board (ARB) announces the release of Technical Report 3 (TR3). TR3 provides a preview for the next OpenMP specification, OpenMP 4.1. It includes significant enhancements to the OpenMP device constructs, which support execution on GPUs, DSPs and other coprocessors. It also includes two significant extensions for parallelization of loops.
The key features that are included in this Technical Report include:
- • Refinement of device constructs. The addition of unstructured data movement will greatly simplify the use of these constructs with C++ applications. In addition, improved support for asynchronous execution will simplify concurrent computation on host CPUs and accelerators as well as concurrent data transfer to and from the target device.
- • New features to handle dependences within loops. Extensions to the ordered construct and the associated clause on loop constructs provide a simple mechanism to support doacross loops (parallel loops with backward loop-carried dependences), as well as the ability to apply the SIMD construct to similar loops.
- • Creation of tasks from loops. The taskloop construct specifies that the iterations of one or more of the associated loops will be executed in parallel using OpenMP tasks. This feature will simplify the parallel execution of longer running tasks, such as I/O, with the iterations of a loop.
Feedback is welcome
The specs in this Technical Report is a work in progress. The goal of this release is to get feedback on the proposed API. Feedback can be posted on the OpenMP Forum, for which registration is required. Deadline for feedback is Feb. 6th, 2015.
“With the improvements to its asynchronous execution model and more flexible data mapping features, TR3 makes it even easier to write portable programs that leverage the performance of all CPUs and accelerators in a user’s system,” said Eric Stotzer of TI and James Beyer of Cray, co-chairs of the Accelerator Subcommittee of the OpenMP Language Committee.
“TR3 represents a significant body of work completed in the last year,” said Bronis R. de Supinski, the Chair of the OpenMP Language Committee. “The report demonstrates that we will meet our commitment to release annual updates to the OpenMP specification, either in technical reports or in new specification versions. OpenMP 4.1 will be completed and released in time for SC15.”