GPU SpMV Deliverable

Implementations

• CPU CSR reference implementation
• GPU simple CSR kernel: one CUDA thread per row
• GPU adaptive CSR kernel:
  • normal rows: one thread per row
  • long rows: one CUDA block per row with shared-memory reduction
• cuSPARSE CSR SpMV comparison

Build

make

Data

Download the SuiteSparse Matrix Collection inputs used for the benchmarks:

./scripts/download_williams_matrices.sh

This creates data/*.mtx for the following matrices:

• pdb1HYS
• consph
• cant
• pwtk
• mac_econ_fwd500
• mc2depi
• cop20k_A
• scircuit
• webbase-1M
• rail4284

Validation

make run

Benchmark

./build/benchmark data/*.mtx

Output

Benchmark CSV files are written to results/.

Reproducibility Environment

The included benchmark results were produced with the following environment:

• Node: edu01
• Loaded module: CUDA/12.3.2
• GPU: NVIDIA A30, 24576 MiB
• NVIDIA driver: 550.144.03
• NVIDIA-SMI reported CUDA version: 12.4
• CUDA compiler: nvcc release 12.3, V12.3.107
• GCC: 11.4.1 20230605 (Red Hat 11.4.1-2)
• G++: 11.4.1 20230605 (Red Hat 11.4.1-2)
• CPU: Intel Xeon Silver 4309Y CPU @ 2.80GHz