Install dependencies

System dependencies:

• Python3
• GNU parallel

Install Python dependencies:

python3 -m pip install -r requirements.txt

Perform the following operations in order

Generate VBR Matrices and their Matrix Market equivalent

./synthesize_matrices/generate_vbr_matrices_w_all_mostly_dense.sh
./synthesize_matrices/generate_vbr_matrices_w_some_mostly_sparse.sh
python3 synthesize_matrices/gen.py -o vbr_to_mtx
python3 synthesize_matrices/gen.py -o vbr_to_mtx --dense-blocks-only

The corresponding files will be generated in Generated_VBR/, Generated_VBR_Sparse/, Generated_MMarket/ and Generated_MMarket_Sparse/.

Generate VBR matrices from manually partitioned Suitesparse matrices

python3 utils/convert_real_to_vbr.py

The corresponding files will be generated in manual_vbr/.

SABLE

Run SABLE to generate CPU code to perform SpMV and SpMM over these generated Matrices

python3 synthesize_matrices/gen.py -o vbr_to_spmv
python3 synthesize_matrices/gen.py -o vbr_to_spmv --dense-blocks-only
python3 synthesize_matrices/gen.py -o vbr_to_spmm
python3 synthesize_matrices/gen.py -o vbr_to_spmm --dense-blocks-only

The corresponding files will be generated in Generated_SpMV/, Generated_SpMV_Sparse/, Generated_SpMM/ and Generated_SpMM_Sparse/.

Run SABLE to generate CUDA code to perform SpMV and SpMM over these Matrices

python3 synthesize_matrices/gen.py -o vbr_to_spmv --cuda
python3 synthesize_matrices/gen.py -o vbr_to_spmv --cuda --dense-blocks-only
python3 synthesize_matrices/gen.py -o vbr_to_spmm --cuda
python3 synthesize_matrices/gen.py -o vbr_to_spmv --cuda --dense-blocks-only

The corresponding files will be generated in Generated_SpMV_cuda/, Generated_SpMV_cuda_Sparse, Generated_SpMM_cuda and Generated_SpMM_cuda_Sparse/.

Benchmark SABLE

python3 bench.py
python3 bench_sparse.py
python3 bench_manualvbr.py
python3 bench_cuda.py
python3 bench_sparse_cuda.py

The corresponding files will be generated in results/.

Benchmarking the CSR SpMV version

For a particular matrix:

# to compile for multi-thread execution with OpenMP
mkdir -p tmp && g++ -DOPENMP -O3 src/csr-spmv.cpp -o tmp/csr-spmv -fopenmp

# to compile for single-thread execution
mkdir -p tmp && g++ -O3 src/csr-spmv.cpp -o tmp/csr-spmv

# to execute a particular matrix
./tmp/csr-spmv <.mtx file> <threads> <num_bench_runs> <dense_vector_path>

For all of Suitesparse:

python3 bench_naive.py