Technical data

Vector Processing Concepts

1.9 COMBINING VECTOR OPERATIONS TO IMPROVE EFFICIENCY

Some of the techniques available to increase vector instruction efﬁciency

include overlapping and chaining.

1.9.1 Instruction Overlap

Overlapping instructions involves combining two or more instructions

to overlap their execution to save execution time. If a vector processor

has independent function units, it can perform different operations on

different operands simultaneously. Overlapping provides a signiﬁcant

gain in performance. If a register must be reused or if data is not yet

available, overlapping may not be possible.

1.9.2 Chaining

Chaining, a special form of instruction overlap, is possible with multiple

function units. Chaining is passing the result of one operation in one

function unit to another function unit. For example, an add instruction

followed by a store command can "combine" so each element of the vector

is stored as soon as the result is obtained. The processor does not have to

wait for the add instruction to ﬁnish before storing the data.

VADD V1,V2,V3

VSTORE V3

As results are generated by the add instruction, they are immediately

available for input to the waiting store instruction. The store instruction

can then begin processing the data.

Instruction chaining only works if all the data to be processed is available

at the beginning of the pipeline. If the result of one operation must be

used as input to another operation in the same data stream, instruction

chaining cannot be used.

1–18