Towards dense linear algebra for hybrid GPU accelerated manycore systems

By S. Tomov, Jack J. Dongarra, Marc Baboulin. Parallel Computing, 2009

496 citations recorded by the source corpus.

Subfields: Computational Theory and Mathematics, Hardware and Architecture

View Paper: https://www.semanticscholar.org/paper/96c9d47ea4fe159e75a42d17dce463adcfdea646

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computational Theory and Mathematics

1. Bo Li
2. J. Demmel
3. Jack J. Dongarra
4. Bo Lang
5. Piotr Luszczek
6. David J. Gardner
7. Daniel R. Reynolds
8. Carol S. Woodward

In Hardware and Architecture

1. Jack J. Dongarra
2. J. Demmel
3. Piotr Luszczek
4. J. Langou
5. Jakub Kurzak
6. S. Tomov
7. Bo Li

Citations in the corpus

• Enabling New Flexibility in the SUNDIALS Suite of Nonlinear and Differential/Algebraic Equation Solvers (2020, background): “The current set of SUNDIALS-provided SUNLinearSolvers includes direct solvers for dense (native, LAPACK [2], and MAGMA [72]), banded (native and LAPACK), and sparse (KLU [23, 24], SuperLU_MT [26], SuperLU_Dist [53], and cuSOLVER [57]) matrix-based linear systems, as well as an ensemble of…”
• Numerical linear algebra on emerging architectures: The PLASMA and MAGMA projects (2009, methodology): “An example of the new class of communication-optimal algorithms is the RBT LP LU(NB) algorithm that we developed for Multi-core + GPU systems [11].”
• StarPU: a unified platform for task scheduling on heterogeneous multicore architectures (2011, methodology): “Figure 5 shows the speed of an LU decomposition with partial pivoting (SGETRF) running on a single nehalem CPU aside with a QUADRO FX5800 GPU, either with StarPU or with the MAGMA library [8].”
• GWCNN: A Metric Alignment Layer for Deep Shape Analysis (2017, unknown): “We implement this layer using the Torch library [CKF11] and execute the metric alignment algorithm and its differentiation on GPU using NVIDIA’s cuDNN [CWV ∗ 14] and MAGMA [TDB10].”
• Orthogonal Weight Normalization: Solution to Optimization over Multiple Dependent Stiefel Manifolds in Deep Neural Networks (2017, methodology): “We also evaluate computational cost per iteration in our current Torch-based implementation, where the convolution relies on the fastest cudnn package and the eigen-decomposition on MAGMA [33].”