# 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
```bash
cd ./IA-SPGEMM-CPU_release;
make;
./spgemm-cpu Inputs/dia.mtx;
```
(2) run SpGEMM code on GPU with auto-tuning in double precision
```bash
cd ./IA-SPGEMM-GPU_release;
make;
./spgemm-gpu Inputs/dia.mtx;
```
**Intel & AMD CPU example**
<img src="https://github.com/AnonymousPPOPP2019/IA-SPGEMM/blob/master/IA-SPGEMM-CPU_release/1.jpg"/>
**NVIDIA GPU example**
<img src="https://github.com/AnonymousPPOPP2019/IA-SPGEMM/blob/master/IA-SPGEMM-GPU_release/2.jpg"/>
## 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:
<img src="https://github.com/AnonymousPPOPP2019/IA-SPGEMM/blob/master/model.png"/>
一种用于并行稀疏矩阵-矩阵乘法的输入感知自动.zip （70个子文件） 
IA-SpGEMM-master 
model.png 27KB IA-SPGEMM-CPU_release detail 
format.h 1KB ell common_ell.h 5KB dia common_dia.h 6KB dense common_dense.h 1KB csr common_csr.h 8KB coo common_coo.h 5KB common.h 850B utime.h 458B utils.h 381B MatNet.py 3KB Makefile 360B
1.jpg 336KB main.cpp 20KB Inputs sample.mtx 4KB b1_ss.mtx 663B relat3.mtx 1KB Trec5.mtx 2KB small.mtx 611B ch3-3-b2.mtx 1KB dia.mtx 607B Ragusa18.mtx 1KB LFAT5.mtx 1KB NetWeights Amd_weights.h5 214KB Intel_weights.h5 214KB __pycache__ MatNet.cpython-36.pyc 3KB MatNet.cpython-35.pyc 3KB imgs 
img1.txt 32KB img2.txt 32KB mmio.h 16KB spgemm-cpu 594KB mmio.c 17KB IA-SPGEMM-GPU_release spgemm-gpu 5.56MB main.cu 15KB P100_weights.h5 214KB 2.jpg 580KB detail format.h 2KB ell common_ell.h 6KB dia common_dia.h 6KB ell_dev common_ell_dev.h 11KB dia_dev common_dia_dev.h 8KB dense common_dense.h 1KB csr common_csr.h 6KB csr_dev common_csr_dev.h 8KB coo common_coo.h 5KB cusp common_cusp.h 830B common.h 2KB coo_dev common_coo_dev.h 23KB utime.h 936B cusparse common_cusparse.h 5KB utils.h 381B MatNet.py 3KB Makefile 977B Inputs sample.mtx 4KB b1_ss.mtx 663B relat3.mtx 1KB Trec5.mtx 2KB small.mtx 611B ch3-3-b2.mtx 1KB dia.mtx 607B NetWeights P100_weights.h5 214KB __pycache__ MatNet.cpython-35.pyc 3KB imgs img1.txt 32KB img2.txt 32KB mmio.h 16KB mmio.c 17KB NetWeights P100_weights.h5 214KB Amd_weights.h5 214KB Intel_weights.h5 214KB README.md 2KB资源评论
