FlashSparse

FlashSparse: Minimizing Computation Redundancy for Fast Sparse Matrix Multiplications on Tensor Cores

FlashSparse is accepted by PPoPP 2025. See the published version of the paper. FlashSparse significantly reduces the computation redundancy for unstructured sparsity (for SpMM and SDDMM) on Tensor Cores through a Swap-and-Transpose mapping strategy.

1. Clone this project.

git clone --recursive https://github.com/ParCIS/FlashSparse.git

• Requirements:

• Ubuntu 16.04+
• cmake >= 3.29
• CUDA >= 11.8
• one NVIDIA RTX4090 GPU and one H100 PCIe GPU.

2. Environment Setup.

Conda environments need to be set up on machines with H100 PCIe and RTX4090 GPUs following the steps below.

2.1 Install via Conda.

• 2.1.1 Install conda on system. (Toturial).
• 2.1.2 Create a conda environment:

conda create -n env_name python=3.9

• 2.1.3 Install PyTorch (Toturial):

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia 

3. Install FlashSparse.

cd FlashSparse/
bash compile.sh

4. Download datasets. (optional)

cat dataset.txt
The sparse matrices can be downloaded from the SuiteSparse repository, while GNN-related datasets are available on the official websites of DGL and PyG. After downloading, the matrices can be partitioned and preprocessed using FS_Block_gpu.preprocess_gpu_fs(). The preprocessed results are then used for computation with FlashSparse.(total 515 sparse matrices).

5. Install Sparse-Kernel Baselines.

Install RoDe, Sputnik, GNNAdvisor, GE-SpMM, cuSPARSE, DTC-SpMM and TC-GNN:

cd Baseline/
bash compile.sh

6. Install GNNs-Framework Baselines.

6.1 Install Deep Graph Library (DGL) (Toturial):

pip install  dgl -f https://data.dgl.ai/wheels/torch-2.1/cu118/repo.html

6.2 Install Pytorch-Geometric (PyG) (Toturial):

pip install torch_geometric

7. Running FlashSparse on H100 PCIe and RTX4090 GPUs.

7.1 SpMM test

• Go to project eva/kernel/spmm/ directory.
• bash ./test_spmm_shell.sh to run all SpMM experiments. (about 200 minutes)
• Check the results in result/FlashSparse/spmm/*.csv.

7.2 SDDMM test

• Go to project eva/kernel/sddmm/ directory.
• bash ./test_sddmm_shell.sh to run all SDDMM experiments. (about 100 minutes)
• Check the results in result/FlashSparse/sddmm/*.csv.

7.3 GCN and AGNN tests

• Go to project eva/end2end/gcn/ directory.
• python eva_gcn_fs.py to run GCN experiments.
• python eva_gcn_baseline.py to run GCN experiments.
• Check the results in result/FlashSparse/gcn/fs_gcn_128.csv. (about 5 minutes)
• Check the results in result/Baseline/agnn/baseline_gcn_128.csv. (about 15 minutes)

• Go to project eva/end2end/agnn/ directory.
• python eva_agnn_fs.py to run AGNN experiments.
• python eva_agnn_baseline.py to run AGNN experiments.
• Check the results in result/FlashSparse/agnn/fs_agnn_32.csv. (about 5 minutes)
• Check the results in result/Baseline/agnn/baseline_agnn_32.csv. (about 15 minutes)

8. Running Baselines on H100 PCIe and RTX4090 GPUs (Optional).

8.1 Evaluating RoDe, Sputnik and cuSPARSE.

• Go to project Baseline/RoDe/script/ directory.
• bash download.sh to download the same 515 matices in a specific format for RoDe. (optional)
• bash test_spmm_shell.sh to run all SpMM experiments. (about 300 minutes)
• bash test_sddmm_shell.sh to run all SDDMM experiments. (about 300 minutes)
• Check the results in result/Baseline/spmm/rode*.csv and result/Baseline/sddmm/rode*.csv.

8.2 Evaluating DTC-SpMM.

• Go to project Baseline/DTC-SpMM/ directory.
• bash test_spmm_shell.sh to run all SpMM experiments. (about 20 minutes)
• Check the results in result/Baseline/spmm/dtc*.csv.

8.3 Evaluating GNNAdvisor, TCGNN, GE-SpMM.

• Go to project eva/kernel/spmm/ directory.
• bash test_spmm_shell_base.sh to run all SpMM experiments. (about 100 minutes)
• Check the results in result/Baseline/spmm/base*.csv.

• Go to project eva/kernel/sddmm/ directory.
• bash test_sddmm_shell_base.sh to run all SDDMM experiments. (about 20 minutes)
• Check the results in result/Baseline/sddmm/base*.csv.

8.4 Summarize the results in all baselines.

• Go to project result/Baseline/spmm/ directory.
• python summarize.py to summarize all results.

• Go to project result/Baseline/sddmm/ directory.
• python summarize.py to summarize all results.

9 Reproducing the experinmental figures and tables in FlashSparse.

9.1 Reproduce the Figure 11 and Table 5. (both on H100 and RTX4090)

• Go to project eva/plot/kernel_spmm/ directory.
• python plot_figure11_ac.py and check the figure in figure11.png (The plotted figure11.png on H100 corresponds to Figure 11(a) in the paper, and on RTX4090 corresponds to Figure 11(c) in the paper.)
• python plot_figure11_bd.py and check the figure in figure11_sub.png. (The plotted figure11_sub.png on H100 corresponds to Figure 11(b) in the paper, and on RTX4090 corresponds to Figure 11(d) in the paper.)
• python profile_table5.py and check the result in table5.txt. (The profiled table5.txt on H100 corresponds to Table5(left) in the paper, and on RTX4090 corresponds to Table5(right) in the paper.)

9.2 Reproduce the Figure 13 and Table 6. (both on H100 and RTX4090)

• Go to project eva/plot/kernel_sddmm/ directory.
• python plot_figure13_a.py and check the figure in figure13(a).png.
• python plot_figure13_b.py and check the figure in figure13(b).png. (The plotted figure13(a).png and figure13(b).png on H100 correspond to Figure 13(a)(b) in the paper, and on RTX4090 corresponds to Figure 13(c)(d) in the paper.)
• python profile_table6.py and check the result in table6.txt. (The profiled table6.txt on H100 corresponds to Table6(left) in the paper, and on RTX4090 corresponds to Table6(right) in the paper.)

9.3 Reproduce the Figure 12. (only on H100 or RTX4090)

• Go to project eva/plot/ablation/memory/ directory.
• python spmm.py and check the result in memory_spmm.csv. (about 20 minutes)
• python sddmm.py and check the result in memory_sddmm.csv. (about 20 minutes)

• python plot_spmm.py and check the figure in spmm_mem.png.
• python plot_sddmm.py and check the figure in sddmm_mem.png.

9.4 Reproduce the Figure 14. (both on H100 and RTX4090)

• Go to project eva/plot/ablation/throughput/ directory.
• python plot_spmm.py and check the figure in figure14(a).png.
• python plot_sddmm.py and check the figure result in figure14(b).png. (The plotted figure14(a).png and figure14(b).png on H100 correspond to Figure 14(a)(b) in the paper, and on RTX4090 corresponds to Figure 14(c)(d) in the paper.)

9.5 Reproduce the Figure 15. (both on H100 and RTX4090)

• Go to project eva/plot/ablation/access/ directory.
• python plot.py and check the figure in figure15.png. (The plotted figure15.png on H100 correspond to Figure 15(left) in the paper, and on RTX4090 corresponds to Figure 15(right) in the paper.)

9.6 Reproduce the Table 7. (only on H100 or RTX4090)

• Go to project eva/plot/ablation/format/ directory.
• python format.py and check the result in result.csv. (about 25 minutes)
• python profile.py and check the output`.

9.7 Reproduce the Figure 16. (only on H100 or RTX4090)

• Go to project eva/plot/gcn/ directory.
• python plot.py and check the figure in figure16_gcn.png.

• Go to project eva/plot/agnn/ directory.
• python plot.py and check the figure in figure16_agnn.png.

9.8 Reproduce the Table 8. (only on H100 or RTX4090)

• Go to project eva/accuracy/gcn/ directory.
• python eva_gcn.py. (about 1 minutes)
• Check the result in result/Baseline/gcn/accuracy.csv.