Optimum Prefetching Patterns Searching: A Case Study of Matrix-Matrix Multiplication
Varintorn Khomongkonudom, Panyayot Chaikarn · 2022
Prefetching reduces data fetch latency and augments the speed of program execution. This paper presents an analysis model for selecting the optimum prefetching pattern for matrix-matrix multiplication. By calculating the number of prefetch instructions and the amount of data from prefetching, numbers of prefetching candidates are selected. Then the best prefetching pattern is obtained by selecting the candidate with lowest value of memory access. The prefetching pattern obtained from this algorithm is on average 18.86 percent faster than without prefetching. It is on average 17.54 percent faster than the automatic prefetching feature provided by the Intel C compiler.