Energy Efficient/Optimal Exascale Accelerators
Abstract
The context of the project is to develop an accelerator ecosystem with generic, retargettable compliers and run-time systems, for time/energy optimal mapping of programs to graphics processing units (GPUs). To attain energy efficiency/optimality we developed transformations that first perform iteration space tiling, followed by a specific scheduling strategy called overlapped multi-pass wavefront parallelization. It showed significant gains for CPU based stencil programs. Specific STIR project deliverables were to to extend this approach to GPUs.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 12, 2023
- Accession Number
- AD1223950
Entities
People
- Sanjay Rajopadhye
Organizations
- Colorado State University