About lower-level ir optimizatin detection
Kyrie-Zhao opened this issue · comments
I appreciate your efforts, nice work! Well, seems that nn-meter does not consider the tensor/code level optimization, while optimizations in this stage can make quite huge effects on run-time performance, especially for AI accelerators. Anyone interested in this topic? Will nn-meter extend the work to fully considerations of compiler level acceleration?
Hi Kyrie, thanks for the interest and suggestion! Yep, nn-meter currently does not consider the compiler level optimizations. As we discussed in the paper (Discussion section), we will consider this kind of optimizations to further develop nn-meter.
Very welcome to communicate about the compiler level accelerations!