GPU Pro 5

342 IV Mobile Devices

which means that on each iteration 0 ≤ k<N all work-items first read from A

and then from B.The

, operator for 0 ≤ j<λ

0

is the innermost one, due to

the order in which threads are created by the device.

12

7.5.5 Blocking

As a first step towards better implementations, we will introduce blocking,a

program transformation that will allow us to use vector operations for memory

accesses and arithmetic operations. We will later exploit blocking to improve

cache usage.

Let us assume that matrix order N is divisible by blocking factors ΔI,ΔJ

and ΔK

reg.

. Imagine that

• matrix A consists of

N

ΔI

×

N

ΔK

reg.

submatrices of ΔI ×ΔK

reg.

elements each,

• matrix ...