Vector addition, execution time

Table II lists the execution times for a DO-Loop which performs a simple addition of two vectors A and B. Times are given for both scalar and vector additions using one or two pipes as well as times for the same operation on the Cyber 174. Note that, for vector operations, increasing the array size is much less expensive than the corresponding increase on a scalar machine. For very small arrays (e.g. N 5) scalar execution on the ASC is faster because Vector overhead consumes a much higher percentage of the execution time.

Example 1 Vector Addition CUDA C extends C by allowing the programmer to define C functions, called kernels, that, when called, are executed N times in parallel by N different CUDA threads, as opposed to only once like regular C functions. 

The pipelines achieve their highest sustained flow rate in vector mode. In this situation, a single instruction is interpreted and a single operation performed on many pairs of operands. For example, if A, B, and C are arrays, only one vector instruction is required for computing the sum C(I) = A(I) + B(I), 1=1,100. The A and B values stream continuously into the pipes, additions are performed in discrete steps and results flow back to CM at the rate of one per CPU clock cycle per pipe. This is in contrast to scalar execution which requires five instructions to be executed 100 times 2 fetches, 1 add, 1 store, and 1 counter incremention. 

---