van de Geijn and Watts SUMMA
Scalable
Universal MM Algorithm
Claimed
to be the best practical algorithm
Uses
overlap, pipelining, decomposition
Initialize
C, blocking all arrays the same
broadcast
(segment of) 1st A column to processors in row
broadcast
(segment of) 1st B row to processors in column
for
i = 2 through n
broadcast
(segment of) next A column to all processors in row
broadcast
(segment of) next B row to all processors in column
compute
i-1 term in dot product for all elements of block
compute
last term for all elements of block
