Par4all: Auto-Parallelizing C and Fortran for the CUDA Architecture

•Table des matières ◮
HPC Project hardware: WildNode from Wild Systems 2
HPC Project software and services 3
We need software tools 4
1 Par4All
Outline 5
Use the Source, Luke... 6
PIPS 7
Current PIPS usage 9
Current PIPS usage 10
2 CUDA generation
Outline 11
Basic CUDA execution model 12
Challenges in automatic CUDA generation 13
Automatic parallelization 14
Outlining 15
From array regions to GPU memory allocation 17
Communication generation 19
Loop normalization 21
�Par4All in CUDA — GPU conference 10/1/2009
From preconditions to iteration clamping 22
Complexity analysis 24
Optimized reduction generation 25
Fortran to CUDA 26
Par4All accel runtime — the big picture 29
3 Results
Outline 33
Results on a customer application 34
Comparative performance 35
Keep it simple (precision) 37
4 Conclusion
Outline 39
Take advantage of C99 40
From an open source project to genetic algorithms 41
Future work 42
Conclusion 44
Par4All is currently supported by... 47
You are here! 48
HPC Project, Mines ParisTech, TÉLÉCOM Bretagne, RPI Ronan KERYELL et al. 46 / 46