Search code examples
matrix-multiplicationblasintel-mkl

Does Intel MKL matrix product gemm work in-place?

The BLAS gemm routine in the Intel MKL normally works with three matrices. It's like giving you f(A, B, C) = alpha A * B + beta C where alpha and beta are scaling factors.

But can one write f(A, B, A) with alpha=1,beta=0 in order to simply get A = A * B? I mean the two A's in f(A, B, A) are the same variable. (All square matrices here.)

Certainly, if we set a third variable C=A, f(A, B, C) works. But it will be much better without even making this copy C.

Solution

  • No, this is not allowed. You either have to introduce a temporary buffer or find another way to do the job.