hiidea / ia-spgemm

An Input-aware Auto-tuning Framework for Parallel Sparse Matrix-Matrix Multiplication

IA-SPGEMM

IA-SPGEMM is a An Input Auto-tuning Sparse General Matrix-Matrix Multiplication on Multicore and Manycore Architure. Currently supported components include:

• SpGEMM algorithms for COO, DIA and ELL sparse storage format
• Feature extraction and density representation
• MatNet (mix CNN and BP)

All tests default calculate the square of A for matrix inputs.

The tool extracts all of the features and density representation as MatNet inputs.

It is easy to use and provide unified interface.

Getting Started

In IA-SPGEMM system, the goal is to search an optimal format and algorithm that minimizes computing overhead.

Setting up an IA-SPGEMM is easy.

(1) run SpGEMM code on CPU with auto-tuning in double precision

cd ./IA-SPGEMM-CPU_release;
make;
./spgemm-cpu Inputs/dia.mtx;

(2) run SpGEMM code on GPU with auto-tuning in double precision

cd ./IA-SPGEMM-GPU_release;
make;
./spgemm-gpu Inputs/dia.mtx;

Intel & AMD CPU example

NVIDIA GPU example

Requirement

• Intel MKL 2018
• CUSP v0.5.1
• cuSPARSE v8.0
• Python 3.6.2
• tensorflow 1.4.0
• keras 2.1.0

MatNet

Details of the neural network

Weights are in IA-SPGEMM-CPU_release/NetWeights and IA-SPGEMM-GPU_release/NetWeights

MatNet structure is below: