High Performance Dense Linear Algebra on Multicore DSPs



The project aims at developing a fully optimized software stack for dense Basic Linear Algebra Subprograms (BLAS) and advanced linear algebra implementations (LAPACK) on multi-core Texas Instruments Digital Signal Processors (DSPs) and hybrid ARM+DSP architectures. More specifically, the project is based on the BLIS [1] and libflame [2] software infrastructures, adapting their implementations to the specifics of the aforementioned architectures. As a result, the MCSDK-HPC [3] from Texas Instruments
integrates a tuned version of both libraries targeting the C66x multi-core DSP and the hybrid K2H ARM+DSP architecture.

[1] https://github.com/flame/blis
[2] https://github.com/flame/libflame
[3] http://software-dl.ti.com/sdoemb/sdoemb_public_sw/mcsdk_hpc/latest/index_FDS.html