Glimpse: Mathematical Embeddings of Hardware Specification for Neural Compilation*
TimeThursday, July 14th5:10pm - 5:30pm PDT
Location3002, Level 3
Event Type
Research Manuscript
AI/ML Design: System and Platform
DescriptionWe develop an optimizing compiler Glimpse that can expedite the compilation time on multiple generations of hardware by learning to transfer and reuse the experiences among multiple compilation runs. We first devise Blueprint that summarizes the key architectural features of hardware, and use them to synthesize prior distributions of search space. Then, we integrate meta-learning into the optimizing compiler that utilizes the hardware signatures. We also implement a statistical sampling mechanism that can reduce the number of measurements. Experimentation against state-of-the-art compilers with GPUs shows that Glimpse provides significantly faster compilation time while also achieving the best inference time.