Par4all: Auto-Parallelizing C and Fortran for the CUDA Architecture
Par4all: Auto-Parallelizing C and Fortran for the CUDA Architecture
Par4all: Auto-Parallelizing C and Fortran for the CUDA Architecture
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
•Table des matières ◮<br />
HPC Project hardware: WildNode from Wild Systems 2<br />
HPC Project software <strong>and</strong> services 3<br />
We need software tools 4<br />
1 Par4All<br />
Outline 5<br />
Use <strong>the</strong> Source, Luke... 6<br />
PIPS 7<br />
Current PIPS usage 9<br />
Current PIPS usage 10<br />
2 <strong>CUDA</strong> generation<br />
Outline 11<br />
Basic <strong>CUDA</strong> execution model 12<br />
Challenges in automatic <strong>CUDA</strong> generation 13<br />
<strong>Auto</strong>matic parallelization 14<br />
Outlining 15<br />
From array regions to GPU memory allocation 17<br />
Communication generation 19<br />
Loop normalization 21<br />
�Par4All in <strong>CUDA</strong> — GPU conference 10/1/2009<br />
From preconditions to iteration clamping 22<br />
Complexity analysis 24<br />
Optimized reduction generation 25<br />
<strong>Fortran</strong> to <strong>CUDA</strong> 26<br />
Par4All accel runtime — <strong>the</strong> big picture 29<br />
3 Results<br />
Outline 33<br />
Results on a customer application 34<br />
Comparative per<strong>for</strong>mance 35<br />
Keep it simple (precision) 37<br />
4 Conclusion<br />
Outline 39<br />
Take advantage of C99 40<br />
From an open source project to genetic algorithms 41<br />
Future work 42<br />
Conclusion 44<br />
Par4All is currently supported by... 47<br />
You are here! 48<br />
HPC Project, Mines ParisTech, TÉLÉCOM Bretagne, RPI Ronan KERYELL et al. 46 / 46