OSCAR API v2.1: Extensions for an Advanced Accelerator Control Scheme to a Low-Power Multicore API

The number of cores in smartphones and tablet-PCs are rapidly increasing along with their required high computational power. However, almost all applications on those devices have not used multiple cores for their high speed and low power execution since the application development environments, which allow the application developers easy and prompt development of parallelized application, are not available. In addition to the development of parallelized applications, low-power consumption techniques and efficiently use of accelerators such as GPUs are required to application developers. In order to provide more productive application development environment for multicores, an automatic parallelizing compiler, OSCAR compiler, which parallelizes C and Fortran programs automatically by applying multi-grain parallelization, local-memory and cache optimization and low-power optimization, has been developed. Furthermore, the OSCAR API has been also developed as an interface between the OSCAR Compiler and various kinds of shared memory multicores including homogeneous and heterogeneous manycores with SMP, cc-NUMA and PGAS architectures from various vendors, such as ARM, Intel, IBM, AMD, Tilera, Fujitsu, Renesas Electronics, and so on. The OSCAR API v1.0 and v2.0 have been opened and their specifications are available from http://www.kasahara.cs.waseda.ac.jp/. In the OSCAR API v1.0, fundamental thread control, memory control, DMA-control, power control and flexible group barrier are supported. In the OSCAR API v2.0, various kinds of heterogeneous multicores are supported. In this paper, the API extensions in OSCAR API v2.1, which includes control schemes for asynchronously executable accelerators, and hint directives for low-power optimizations, are described in addition to brief review of the OSCAR API v1.0 and v2.0. A flexible and low-overhead accelerator control scheme that allows us overlapped execution of CPUs, accelerators and DMA controllers can be realized by newly added accelerator control APIs.