Matrix–Vector Operation

CIM_MVM

CIM_MVM rs,rt,re,rf,[flags]

Performs the multiplication operation between the input feature vector and the weight matrix

• The address of the input vector in the local memory is GRF[rs], with a length of GRF[rt];
• The address of the weight matrix in the computing-in-memory array is GRF[re];
• The batch operation count is GRF[rf]. When this value is greater than 1, this instruction executes batch multiplication operations;
• flags represents the flag bits, including SP_V, SP_B, etc., used to support the structured pruning algorithm and bit-level sparsity algorithm respectively. This field can be extended to support new operations and custom functions.

Example:

CIM_MVM $1,$2,$3,$4,SP_V,SP_B