| 摘要 |
Speed with which sparse matrices are processed can be increased by using improved compressed representations of the matrices. Structured compressed representations reduce the number of cache misses experienced during matrix processing by decreasing the number of times the cache has to be accessed randomly. Further, representations of the matrix that divide and regroup rows and columns of the matrix based on their number of non-zero entries allows to assign the most appropriate kernel function for processing of these portions of a matrix, overcoming the limitations of the GPU-based hardware. As a result, the speed of processing can be increased without disturbing the original structure of the matrix. |
