Kernel Coding Bounds - 2024.1 English

AI Engine Kernel and Graph Programming Guide (UG1079)

Document ID
Release Date
2024.1 English

In this example, a total of 1024 int16 x int16 multiplications are required for computing 128 output value. Given that 32 16-bit multiplications can be performed per cycle in an AI Engine, the compute bound for the kernel is as follows.

Compute bound = 32 cycles / invocation

Matrix B can be stored in the vector register because it is only 16*16-bit =256 bits. It does not need to be fetched from the AI Engine data memory or tile interface for each MAC operation. Considering the data “a” needed for computation, there are total 64*8*2=1024 bytes to be fetched from memory. Given that AI Engine allows two 256 bits (32 bytes) loads per cycle, the memory bound for the kernel is as follows.

Memory bound = 1024 / (2*32) = 16 cycles / invocation

It is seen that the compute bound is larger than the memory bound. Hence the purpose of vectorization can be to achieve the theoretical limit of MAC operations in the vector processor.