Adjusting primitives for graph : SHORT REPORT / NOTES
0 likes14 views
The document discusses performance evaluations of various graph algorithms, specifically focusing on vector operations such as multiplication and summation using different execution modes including sequential, OpenMP, and CUDA. It compares the performance of using different storage types like float and bfloat16, as well as various CUDA launch configurations. Additionally, it explores strategies for in-place operations and their impact on performance metrics.
Adjusting primitives for graph : SHORT REPORT / NOTES
- 1. Adjusting primitives for graph Graph algorithms, like PageRank Compressed Sparse Row (CSR) is an adjacency-list based graph representation that is Multiply with different modes (map) Sequential OpenMP CUDA 1. Performance of sequential execution based vs OpenMP based vector multiply. 2. Comparing various launch configs for CUDA based vector multiply. Sum with different storage types (reduce) float bfloat16 1. Performance of vector element sum using float vs bfloat16 as the storage type. Sum with different modes (reduce) Sequential OpenMP CUDA (memcpy, in-place) 1. Performance of sequential execution based vs OpenMP based vector element sum. 2. Performance of memcpy vs in-place based CUDA based vector element sum. 3. Comparing various launch configs for CUDA based vector element sum (memcpy). 4. Comparing various launch configs for CUDA based vector element sum (in-place). Sum with in-place strategies of CUDA mode (reduce) sum-loop sum-reduce one-loop atomic-add block-loop template, next-pow2 launch one-reduce, next-pow2 launch block-loop template, prev. pow2 launch one-reduce, prev-pow2 launch grid-loop 1. Comparing various launch configs for CUDA based vector element sum (in-place).