-
函數庫接口標准:BLAS (Basic Linear Algebra Subprograms)和LAPACK (Linear Algebra PACKage)
1979年,Netlib首先用Fortran實現基本的
向量乘法、矩陣乘法的函數庫(該庫沒有對運算做過多優化)。后來該代碼庫對應的接口規范被稱為BLAS。
LAPACK也是Netlib用Fortan編寫的代碼庫,實現了高級的線性運算功能,例如矩陣分解,求逆等,底層是調用的BLAS代碼庫。后來LAPACK也變成一套代碼接口標准。
后來,Netlib還在BLAS/LAPACK的基礎上,增加了C語言的調用方式,稱為CBLAS/CLAPACK
因此,BLAS/LAPACK都有兩個含義,一個是Netlib通過Fortran或C實現的代碼庫,一個是這個兩個代碼庫對應的接口標准。
http://www.icl.utk.edu/~mgates3/docs/
現在大多數函數庫都是基於BLAS/LAPACK接口標准實現
https://en.wikipedia.org/wiki/List_of_numerical_libraries
-
開源函數庫
開源社區對對BLAS/LAPACK的實現,比較著名是 ATLAS(Automatically Tuned Linear Algebra Software)和OpenBLAS。它們都實現了BLAS的全部功能,以及LAPACK的部分功能,並且他們都對計算過程進行了優化。
-
商業函數庫
商業公司對BLAS/LAPACK的實現,有Intel的MKL,AMD的ACML。他們對自己的cpu架構,進行了相關計算過程的優化,實現算法效率也很高。
NVIDIA針對其GPU,也推出了cuBLAS,用以在GPU上做矩陣運行。
Matlab用的是MKL庫,可以用version –lapack來查看函數庫的版本
Octave 默認用的是OpenBLAS庫, version -blas
附錄:Lapack中的函數命名規則
============================================================================
lapack naming: x-yy-zzz, or x-yy-zz
x (data type)
------------------------------
s float
d double
c float-complex
z double-complex
ds input data is double, internal use float
zc input data is double-complex, internal use float-complex
Matrix type (yy) | full | packed | RFP | banded | tridiag | generalized problem
================================================================================
general | ge gb gt gg
symmetric | sy sp sf sb st
Hermitian | he hp hf hb
positive definite| po pp pf pb pt
--------------------------------------------------------------------------------
triangular | tr tp tf tb tg
upper Hessenberg | hs hg
trapezoidal | tz
--------------------------------------------------------------------------------
orthogonal | or op
unitary | un up
--------------------------------------------------------------------------------
diagonal | di
bidiagonal | bd
(zzz) algorithm
------------------------------
* Triangular factorization
-trf — factorize: General LU, Cholesky decomposition
-tri — calculate the inverse matrix
* Orthogonal factorization
-qp3 — QR factorization, with pivoting
-qrf — QR factorization
* Eigenvalue
-ev — all eigenvalues, [eigenvectors]
-evx — expert; also subset
-evd — divide-and-conquer; faster but more memory
-evr — relative robust; fastest and least memory
* SVD singular value decomposition
-svd — singular values
* Linear system, solve Ax = b
-sv — solve
-sdd — divide-and-conquer; faster but more memory
* Linear least squares, minimize ||b?Ax||2
-ls — full rank, rank(A) = min(m,n), uses QR.
-lsy — rank deficient, uses complete orthogonal factorization.
-lsd — rank deficient, uses SVD.