MATanalyze

The current file works for different algorithms given bandwidth,dimensions,multipliers,nnz of the sparse matrix, matrix A off chip(sparse), matrix B on chip(dense). The type of Matrix multiplication assumed is Sparse-Dense.

Downloading the files :

From the terminal,

git clone https://github.com/ashuthosh-mr/MATanalyze

Inputs :

The parameters could be changed in the param.py file :

• Dimensions : M,K,N = 256,256,256
• Number of NNZ in Matrix A : nnz = 256
• Algorithm used, rw for row-wise, cw for column-wise, in for inner product, op for outer product : algorithm = "rw"
• Number of multipliers : mul = 512
• Number of values every cycle (assuming compute freq = data transfer freq) : bw = 2
• Location of Matrix A (off if off chip, on if on chip) : mat1 = "off"
• Location of Matrix B (off if off chip, on if on chip) : mat2 = "on"

Running the code :

From the terminal go to the MATanalyze directory that contains the following files : mat.py and param.py
After making the necessary changes to the param.py file,

python param.py 

Output :

The output will contain the following information :

• Memory bound or compute bound?
• Latency
• Minimum On chip memory necessary