✳flâneur — a map of the web's best reading

Block Sparse Matrix-Vector Multiplication with CUDA | by Georgii Evtushenko | GPGPU | Medium

medium.com · 1,501 words · saved by 1 readers

In the previous post, we’ve discussed sparse matrix-vector multiplication. It was shown that it’s possible to take advantage of knowledge about a position of zeroes by storing matrices in special data structures. Although we’ve improved performance and memory space requirements, we haven’t used all the information about zeroes position. In some applications, non-zeroes are gathered in blocks. The knowledge about these blocks could give us more room for optimization. In this post, I’m going to discuss the efficiency of block sparse matrix-vector multiplication on GPU. To show some real-live application results, I develop a Matrix Structural Analysis application, which is used to simulate the Golden Gate bridge structure. BCSR is one of the most popular block sparse matrix formats. In BCSR, all blocks have the same size. To understand this format imagine a sparse matrix with the block size equal to one. In this case, CSR and BCSR matrix representations are equivalent. Block size increasi

Block Sparse Matrix-Vector Multiplication with CUDA Georgii Evtushenko 7 min read · Dec 28, 2019 -- Listen Share Press enter or click to view image in full size In the previous post, we’ve discussed sparse matrix-vector multiplication. It was shown that it’s possible to take advantage of knowledge about a position of zeroes by storing matrices in special data structures. Although we’ve improved performance and memory space requirements, we haven’t used all the information about zeroes position. In some applications, non-zeroes are gathered in blocks. The knowledge about these blocks could give

Explore this link on the map →

saved by

related reading