What question did this study set out to answer?

To automate the selection of the highest-performing sparse matrix storage format for sparse matrix-vector multiplication (SpMV) on GPUs using machine learning.

August 1, 2016

Sparse Matrix Format Selection with Multiclass SVM for SpMV on GPU

View Full Paper

Ask AI

Bookmark

Authors

MBMohamed Akrem BenatiaPolytechnic School of Algiers WJWeixing JiBeijing Institute of Technology YWYizhuo WangBeijing Institute of Technology

Discussion

Member takes

Implication

Computational evaluation demonstrates accurate sparse matrix format selection across two GPU architectures, achieving over 98% of optimal multiplication performance.

Key Points

To automate the selection of the highest-performing sparse matrix storage format for sparse matrix-vector multiplication (SpMV) on GPUs using machine learning.
Extracted low-overhead matrix sparsity features to characterize input structures for GPU execution.
Trained a multiclass Support Vector Machine (SVM) classifier to select among four standard sparse matrix formats: COO, CSR, ELL, and HYB.
Evaluated performance on two distinct GPU architectures: NVIDIA Fermi GTX 580 and Maxwell GTX 980 Ti.
Achieved greater than 98% of the execution performance possible with an optimal format selection.
Demonstrated effective format classification across disparate GPU microarchitectures without manual tuning.

Cite This Study

Benatia et al. (2016) studied this question.

synapsesocial.com/papers/69da12720d540cafc5838409 https://doi.org/10.1109/icpp.2016.64

View Full Paper

Ask AI

Bookmark

Implication

Computational evaluation demonstrates accurate sparse matrix format selection across two GPU architectures, achieving over 98% of optimal multiplication performance.

Sparse Matrix Format Selection with Multiclass SVM for SpMV on GPU

Discussion

Member takes

Implication

Key Points

Cite This Study

Sparse Matrix Format Selection with Multiclass SVM for SpMV on GPU

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider