Evaluation of Speedup & Energy with Multigrain Parallelizing Compiler
summary
Automatic source-to-source OSCAR compiler applies multigrain parallelism, task-graph scheduling and L3-aware data localization to C/C++/Fortran and compiles with GCC/Clang. On Intel IceLake-SP and AMD Zen‑4 16-core systems OSCAR yields larger runtime and energy wins over OpenMP — up to 10.6x and 80% versus sequential for NAS CG, and about 7–9% runtime and 4–9% energy improvements versus OpenMP across benchmarks; cache tuning cuts L3 misses and gives ~4% runtime and 4–7% energy gains.
AI-written summary · about
chapters (13)
- 0:00Oscar compiler research
- 1:10Energy-aware computing motivation
- 2:37Intel vs AMD architectures
- 3:23Oscar automatic parallelization
- 5:10Cache optimization in Oscar
- 6:57Benchmark setup
- 8:23Intel performance results
- 10:09AMD performance results
- 11:51Energy reduction vs OpenMP
- 12:14L3 cache impact on energy
- 13:43CCX communication on AMD
- 14:04Oscar compiler benefits
- 14:54Q&A on optimizations
related talks
14mOn-Stack Replacement for Program Generators and Source-to-Source Compilers
01h 01mIntroduction to Vector Parallelism
24mMapping Sparse Applications to Accelerated Computing Systems
29mEmpirical Observations about Profile-Guided Optimizations for Mainstream C/C++ Compilers
54mGCC/Clang Optimizations for Embedded Linux
01h 00mBringing Clang and C++ to GPUs: An Open-Source, CUDA-Compatible GPU C++ Compiler
17mUnlocking 15% More Performance: A Case Study in LLVM Optimization for RISC-V
24mTowards Understanding the Costs of Avoiding Out-of-Thin-Air Results