bblas_interleaved
Basic implementation of batched (GEMM, TRSM, POTRF, POSV) using an interleaved memory layout suitable for use on multicore systems and the self-hosted Intel Xeon Phi (KNL).
We can achieve exceptionally good performance for small matrix sizes when compared to MKL.
Compilation
So far the optimization flags are only supported by icc, they can be motified to support gcc, but we don't guarantee the same performance.
Execution
We acheive a relatively good performance by taking advantage of the KNL MCDRAM. To control the memory allocation we use hbw_malloc and hbw_free instead of a simple malloc and free respectively. The hbw_malloc is from the memkind library. Alternatively, all the hbw_malloc and hbw_free may be replaced by a simple malloc and free but excute the code using "numactl -m 1" to force the program to allocate all its memory in the MCDRAM.