性能库intel数学核心库（mkl）.ppt

* Intel? Math Kernel Library Contents Each of the BLAS has 4 data types: single and double precision real and complex data types. Most all the functions (with some exceptions) have identical functionality in each data type. The extended BLAS are a set of level 1 BLAS, which support sparse data. * Intel? Math Kernel Library Contents Intel MKL’s value-add to the LAPACK code includes: Just building the LAPACK code takes some effort Threading key portions of the functions Optimizing key functions through the use of recursion The new Fourier transforms meet the needs of a far wider audience than did the previous radix-2 FFTs. This list shows key features. Optimization of the functions will continue for some time yet, but the complex transforms are well optimized for IPF-2 now. VML and VSL offer improved performance over scalar implementations of the underlying functions provided the user can vectorize the code. * Roll Your Own/Dot Product Roll Your Own: This is a simple, straightforward dot product approach to matrix multiplication. Note that the innermost loop is a dot product, and thus can be replaced with a call to the dot product, which is shown in the second panel. * DGEMV/DGEMM The two innermost loops comprise a matrix-vector multiply, which can form the central operation of matrix multiplication. DGEMV parameters: incx = 1; incy = ldb; alpha = 1.0; beta = 0.0; transa = t; DGEMM parameters: alpha = 1.0; beta = 0.0; * Intel? Math Kernel Library Optimizations in LAPACK* Threading at higher levels (LAPACK factorization rather than at DGEMM, for instance) opens additional parallelization opportunities. The blocking strategy employed in traditional LAPACK can be extended to the factorization of the block columns to improve locality of reference and minimize vector operations. NETLIB LAPACK has numerous intrinsic function calls, which raises the need for run-time library support. All of these calls have been implemented within Intel MKL, so no run-time