This repo accompanies the paper Probing the Efficacy of Hardware-Aware Weight Pruning to Optimize the SpMM routine on Ampere GPUs, published at PACT'22. It includes the kernels associated with CLASP, which has been presented in that conference. CLASP is a column-vector pruning-aware implementation of the SpMM routine that supports the characteristics of the Ampere platform. It aims to take advantage of the knowledge pushed into the pruning technique to generate the sparse input matrices (e.g. column-vector), and boost the performance achieved on half precision.
git clone --recurse-submodules [email protected]:UDC-GAC/CLASP.git
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCUDA_ARCHS="86" && make -j 12
Note: If you find a problem like this
Policy "CMP0104" is not known to this version of CMake
Please, comment this line cmake_policy(SET CMP0104 OLD) in include/sputnik/CMakeLists.txt
./src/benchmark_spmm --sparsity-type cvs --spmm CLASP --gemm cuBlas --precision half --block-size 16 --m 1024 --k 256 --n 256 --d 0.2 --check
Results can be compared with Sputnik's library using the following notation:
./src/benchmark_spmm --sparsity-type csr --spmm sputnik --gemm cuBlas --precision half --m 1024 --k 256 --n 256 --d 0.2 --check
(Recommended before time measurement) Lock the clocks:
sudo nvidia-smi -i 0 -pm 1
sudo nvidia-smi -lgc 1750 -i 0
@inproceedings{castro2022probing,
author = {Castro, Roberto L. and Andrade, Diego and Fraguela, Basilio B.},
title = {Probing the Efficacy of Hardware-Aware Weight Pruning to Optimize the SpMM routine on Ampere GPUs},
booktitle = {Proceedings of the International Conference on Parallel Architectures and Compilation Techniques, {PACT} 22},
year = {2022},
}
Apache-2.0 License
-- Roberto López Castro