8

I have to calculate some products in the form A'A or more general A'DA, where A is a general mxn matrix and D is a diagonal mxm matrix. Both of them are full rank; i.e.rank(A)=min(m,n).

I know that you can save a substantial time is such symmetric products: given that A'A is symmetric, you only have to calculate the lower --or upper-- diagonal part of the product matrix. That adds to n(n+1)/2 entries to be calculated, which is roughly the half of the typical n^2 for large matrices.

This is a great saving that I want to exploit, and I know I can implement the matrix-matrix multiply within a for loop . However, so far I have been using BLAS, which is much faster than any for loop implementation that I could write by myself, since it optimizes cache and memory management.

Is there a way to efficiently compute A'A or even A'DA using BLAS? Thanks!

ztik
  • 3,482
  • 16
  • 34
enanone
  • 923
  • 11
  • 25

3 Answers3

10

You are looking for dsyrk subroutine of BLAS.

As described in the documentation:

SUBROUTINE dsyrk(UPLO,TRANS,N,K,ALPHA,A,LDA,BETA,C,LDC)

DSYRK performs one of the symmetric rank k operations

C := alpha*A*A**T + beta*C,

or

C := alpha*A**T*A + beta*C,

where alpha and beta are scalars, C is an n by n symmetric matrix and A is an n by k matrix in the first case and a k by n matrix in the second case.

In the case of A'A storing upper triangular is:

CALL dsyrk( 'U' , 'T' ,  N , M ,  1.0  , A , M , 0.0 , C , N )

For the A'DA there is no direct equivalent in BLAS. However you can use dsyr in a for loop.

SUBROUTINE dsyr(UPLO,N,ALPHA,X,INCX,A,LDA)

DSYR performs the symmetric rank 1 operation

A := alpha*x*x**T + A,

where alpha is a real scalar, x is an n element vector and A is an n by n symmetric matrix.

do i = 1, M
    call dsyr('U',N,D(i,i),A(1,i),M,C,N)
end do
soegaard
  • 30,661
  • 4
  • 57
  • 106
ztik
  • 3,482
  • 16
  • 34
2

The dsyrk routine in BLAS suggested by @ztik is the one for A'A. For A'DA, one possibility is to use the dsyr2k routine which can perform the symmetric rank 2k operations:

C := alpha*A**T*B + alpha*B**T*A + beta*C.

Set alpha = 0.5, beta = 0.0, and let B = DA. Note that this way assumes your diagonal matrix D is real.

Qimen Xu
  • 21
  • 2
  • Since D is diagonal, you can also take the square root of it and pre-multiply it against A and then use `dsyrk` on the new matrix. – John Mar 27 '23 at 13:49
0

SYRK is fine for A'A. For A'DA you can use SYMM for one side of it eg V = A'D and then use intel MKL's GEMMT for W = V A. GEMMT is like GEMM except it takes advantage of the fact that the result matrix is symmetric, and thus only has to do roughly half the work.