AlphaSparse

AlphaSparse is a superset of traditional auto-tuners that goes beyond the scope of existing human-designed format(s) and implementation(s). It automatically creates novel machine-designed formats and SpMV kernel implementations entirely from the knowledge of input sparsity patterns and hardware architectures.

The SC'22 paper: https://dl.acm.org/doi/abs/10.5555/3571885.3571972

Hardware Configurations

• We have tested on NVIDIA RTX 2080 and A100. Generally, AlphaSparse can support all up-to-date NVIDIA GPUs.
• CUDA version: 11.x. The nvcc needs to be configured into environment variables.
• GCC version: 9.4 or any others supporting C++11.
• Operating System version: Ubuntu 20.04.3 LTS, Linux kernel 5.4.0-99-generic x86_64.
• Python version: 3.x.
• No third-part library is needed.

Download data of test matrices

In our evaluation, all input matrices are from SuiteSparse Matrix Collection (https://sparse.tamu.edu). Two ways are provided to download data as follows.

• Directly download from websites. Go to the websites of SuiteSparse Matrix Collection, click the link of specific matrix name, and click the download link named "Matrix Market". The downloaded file is zipped. By extracting the file, a ".mtx" file can be gotten.

• Use the python interface. Install the ssgetpy Python module. Run import ssgetpy and type help(ssgetpy) to get a detailed help message on using ssgetpy to search and download sparse matrices. We have provided a python script named get_data_set_from_UF.py to download all the needed matrices. The variable UF_DIR is the destination of downloaded data. The script needs wget.

Run AlphaSparse

Go to the root directory of AlphaSparse, and compile the source code.

make clean
make -j16

Create a directory named data_source.

mkdir data_source

Configure AlphaSparse according to the environment. Go to global_config.json. Set two fields named ROOT_PATH_STR and spmv_header_file need to be set according to the path of AlphaSparse.

"ROOT_PATH_STR": "{path to AlphaSparse}"
"spmv_header_file": "{path to AlphaSparse}/spmv_header_top.code"

Prepare the input matrix. Go to the directory of UF dataset. Extract the zipped file of specific matrix. And convert the .mtx file.

cd {path of UF dateset}
tar -zxvf {matrix name}.tar.gz
python3 data_prepare.py {path of UF dateset}/{matrix name}/{matrix name}.mtx data_source/{matrix name}.mtx.coo

Run AlphaSparse. The description of the best Operator Graph and the corresponding performance are shown in test.log.

./main data_source/{matrix name}.mtx.coo >> data_source/test.log

Batch Test of AlphaSparse

We have provided a script named batch_test_spmv_builder.py to test all matrices in our paper. All matrix names are stored in matrix_name_list.

python3 batch_test_spmv_builder.py >> data_source/test.log

Batch Test of its counterparts

Run ./make.sh in cuda_code/ACSR_test, cuda_code/CSR_adptive_test, cuda_code/ELL_test, cuda_code/taco-csr. Run make in cuda_code/CSR5_cuda. Run ./coo_make.sh and ./csr_make.sh.

Test cuSparse.

python3 batch_test_cusparse_from_UF.py

Test CSR5, ELL, ACSR, CSR-Adaptive.

python3 get_data_set_from_UF.py

Change CUDA version to 9.2. Run make gpu_spmv in cuda_code/merge-spmv. Run batch test of Merge and HYB.

python3 batch_test_merge_hyb_from_UF.py