Gemm and storage order¶

Gemm means general matrix multiplication. It is a common routine in linear algebra.

$Gemm(A, B, C, tA, tB, \alpha, \beta) = \alpha A^{tA} B^{tB} + \beta C$

Where $A^{tA}$ means A is tA if 0 and $A'$ if tA is 1. The coefficients of a matrix are stored in memory in a one dimension array T: $A(i,j) = T[i * C + j]$ where C is the number of columns of matrix A. In that case, the storage is said as row major. In case $A(i,j) = T[j * R + i]$ where R is the number of rows, the storage is column major.

We define a matrix A with $(I, J, M, R)$ , it has I rows, J columns, the memory buffer is M and the matrix order R. In that case, we can express the transpose of this matrix by: If $A=(I,J,M,R)$ , then $A' = (J,I,M,C)$ .

Let’s use that notation for $A=(I,J,M_A,R)$ , $B=(J,K,M_B,R)$ and $C=(I,K,M_C,R)$ . We note $D = A^{tA} B^{tB} = (I, K, M_D, R)$ .

$\begin{array}{rcl} \alpha A^{tA} B^{tB} + \beta C &=& \alpha (I,J,M_A,R)^{tA} (J,K,M_B,R)^{tB} + \beta (I,K,M_C,R) \\ &=& \left( \alpha (I,J,M_A,R)^{tA} (J,K,M_B,R)^{tB} + \beta (I,K,M_C,R) \right)'' \\ &=& \left( \alpha (J,K,M_B,R)^{1-tB} (I,J,M_A,R)^{1-tA} + \beta (I,K,M_C,R)' \right)' \\ &=& \left( \alpha (K,J,M_B,C)^{tB} (J,I,M_A,C)^{tA} + \beta (K,I,M_C,C) \right)' (*)\\ &=& \left( (K,I,M_D,C) + \beta (K,I,M_C,C) \right)' \\ &=& (I,K,M_D,R) + \beta (I,K,M_C,R) \end{array}$

This trick can be used to run the computation of matrices using a column major algorithm instead of a row major algorithm by using line (*) as a replacement.

$\begin{array}{rcl} &&\alpha (I,J,M_A,R)^{tA} (J,K,M_B,R)^{tB} + \beta (I,K,M_C,R) \\ &=& \left( \alpha (K,J,M_B,C)^{tB} (J,I,M_A,C)^{tA} + \beta (K,I,M_C,C) \right)'\\ &=& \alpha (J,I,M_A,C)^{1-tA}(K,J,M_B,C)^{1-tB} + \beta (K,I,M_C,C)' \end{array}$