Software integration of identical DLP threads via compilation for VLIW processors
Maolin Guan, Nan Wu, Mei Wen, Chunyuan Zhang · 2010
Based on the characteristics of data level parallelism (DLP) multi-threading programs appearing in the practical application, this paper proposes a new method that implements software integration of identical DLP threads via compilation for VLIW processors. This method translates DLP into ILP by merging the operations in corresponding basic blocks divided from different threads into one basic block to extend the instruction window that the compiler can schedule, and optimizes the control flow of the program after thread integration to ensure the correctness of the program. The experimental results show that this technique can accelerate the program execution very well without exerting more burdens on the programmer, while the hardware overhead can be ignored. Generally speaking, integration of 2~4 threads can get a speedup of 1.34~2.07.