Vectorized Calculations and Use of Fast Semiconductor Memories in the Dv-Xα Method

Abstract

Two types of vectorized programs of the discrete varia tional Xα method with the basis set of Slater-type or bitals have been developed. The first type calculates a basis function every time that it is used and the second type reads a basis function from outside memories. They are examined for the most efficient way of per forming calculations with the HITAC S-810/20 supercom puter (630 MFLOPS at peak) and its available outside memory devices. The acceleration factor of computation by the first type of program over scalar processing is more than 12. The second type of program has no ad vantage over the first in small electron systems of atomic clusters with primitive basis functions. But in larger systems such as Si2O7H6 with complex basis functions, the second type of program, used with semi conductor memories, results in 11 times less CPU time than the first type of program.