Consider The Product Loops

What does ZPL’s performance model tell us?

Cannon:

[Res] C := 0.0; -- Initialize C

for k := 1 to n do -- Thru common dim

[Res] C := C + A*B ; -- Product & accumulate

[right of Lop] wrap A; -- Send first col right

[Lop] A := A@right; -- Shift array left

[below of Rop] wrap B; -- Send top row down

[Rop] B := B@below; -- Shift array up

end;

SUMMA:

[Res] C := 0.0; -- Initialize C

[Res] for k := 1 to n do

[ ,*] Col := >>[,k] A; -- Flood kth col of A

[*, ] Row := >>[k,] B; -- Flood kth row of B

C := C+Col*Row;-- Accumulate product

end;

Previous slide

Next slide

Back to first slide

View graphic version