iSpLib: A Library for Accelerating Graph Neural Networks using Auto-tuned Sparse Operations
Authors:
Md Saidul Hoque Anik,
Pranav Badhe,
Rohit Gampa,
Ariful Azad
Abstract:
Core computations in Graph Neural Network (GNN) training and inference are often mapped to sparse matrix operations such as sparse-dense matrix multiplication (SpMM). These sparse operations are harder to optimize by manual tuning because their performance depends significantly on the sparsity of input graphs, GNN models, and computing platforms. To address this challenge, we present iSpLib, a PyT…
▽ More
Core computations in Graph Neural Network (GNN) training and inference are often mapped to sparse matrix operations such as sparse-dense matrix multiplication (SpMM). These sparse operations are harder to optimize by manual tuning because their performance depends significantly on the sparsity of input graphs, GNN models, and computing platforms. To address this challenge, we present iSpLib, a PyTorch-based C++ library equipped with auto-tuned sparse operations. iSpLib expedites GNN training with a cache-enabled backpropagation that stores intermediate matrices in local caches. The library offers a user-friendly Python plug-in that allows users to take advantage of our optimized PyTorch operations out-of-the-box for any existing linear algebra-based PyTorch implementation of popular GNNs (Graph Convolution Network, GraphSAGE, Graph Inference Network, etc.) with only two lines of additional code. We demonstrate that iSpLib obtains up to 27x overall training speedup compared to the equivalent PyTorch 2.1.0 and PyTorch Geometric 2.4.0 implementations on the CPU. Our library is publicly available at https://github.com/HipGraph/iSpLib (https://doi.org/10.5281/zenodo.10806511).
△ Less
Submitted 21 March, 2024;
originally announced March 2024.
Photo-nuclear cross sections on $^{197}$Au
Authors:
J. Song,
D. Rotsch,
J. A. Nolen,
R. Gampa,
R. M. de Kruijff,
T. Brossard,
C. R. Howell,
F. Krishichayan,
Y. K. Wu,
S. Mikhailov,
M. W. Ahmed,
R. V. F. Janssens
Abstract:
A method was developed for measuring photonuclear reactions concurrently at several discrete photon beam energies on a stack of different target materials via a single irradiation. Concentric ring targets of the materials (in order from front to back targets: Au, TiO$_2$, Zn, Os, and Au) were irradiated at the High Intensity Gamma-ray Source (HI$γ$S). As a proof of principle, we report the result…
▽ More
A method was developed for measuring photonuclear reactions concurrently at several discrete photon beam energies on a stack of different target materials via a single irradiation. Concentric ring targets of the materials (in order from front to back targets: Au, TiO$_2$, Zn, Os, and Au) were irradiated at the High Intensity Gamma-ray Source (HI$γ$S). As a proof of principle, we report the result of the cross section measurements from the front Au target. The excitation functions of the $^{197}$Au($γ$,n)$^{196}$Au and $^{197}$Au($γ$,3n)$^{194}$Au reactions were determined in the incident photon energy range of 13-31 MeV using quasi-monoenergetic photon beams provided at HI$γ$S. The cross sections of the combined ground state (2$^{-}$) and short-lived first isomeric state (m1, 5$^{+}$), and of the second isomeric state (m2, 12$^{-}$) in the $^{196}$Au production are obtained separately by subtracting the $γ$ rays from the internal conversion of the second isomeric state. The excitation function of the second isomeric state via the photon-induced reaction $^{197}$Au($γ$,n)$^{196m2}$Au was measured for the first time. By using the activation method rather than direct neutron counting, the exclusive cross sections for the ($γ$,n) and ($γ$,3n) reactions were determined. Comparing the yields from the front and back gold targets validates our ability to simulate the effect of photon scattering in the target stack and provides a method for assessing the systematic uncertainty of our technique.
△ Less
Submitted 23 February, 2024;
originally announced February 2024.
