Intel is moving Ct, a high-level research project that supports data parallelism in current multicore and manycore architectures, out of the lab and into the real world of high-performance software development.
Ct, short for "C/C++ for throughput computing," lets programmers abstract data-parallel programming away from the hardware while supporting deterministic behavior to avoid races and deadlocks. It does this in part by adding parallel collection objects and methods to C++ using C++ templates.
Moreover, Ct supports forward-scaling across multicore and manycore processors. This means that programs written today can be ready for tomorrow's hardware without having to rewrite code due to the release of new architectures.
According to Intel, Ct provides several benefits for developers:
- Reduce errors in parallel programming by providing determinism, which helps provide certain guarantees about safety. Safety helps avoid data races and deadlocks, the two most often encountered parallel programming bugs.
- Parallel programming that is readable with an expressive syntax that stays close the domain expert's mode of expression. Ct technology excels by providing a framework that allows programs to keep a programming notation close to the notation used by non-programmer experts.
- Forward-scaling across multicore and manycore processors. SSE, AVX, Larrabee, and beyond and can be supported in a common binary.
- Fit into existing programs. Ct allows for effective data-parallelism to be added into existing programs using existing tools and programming languages. Ct extends C++ for data-parallelism which allows for compatible and incremental addition of parallel programming into existing programs without the need for completely new and incompatible programming languages.
Intel will release a new product beta using Ct technology by the end of the year and these new products will deliver data-parallel capabilities through standard C++ templates. Using high-level abstractions, these products let C++ developers build applications for optimized performance on several to hundreds of cores. Intel says this will complement existing Intel software development products for both data and task parallelism, including full support for Intel Threading Building Blocks and OpenMP.