The ASTEX Pipeline GPU Programming Environment

Short Description
GPU Programming Environment. One of the current most challenging developpers’ issues addressed by ASTEX is to efficiently partition …

Website: perso.ens-lyon.fr | Filesize: 781kb

Content
- PARA
GPGPU with ASTEX: Hot Path Based
Thread Extractor for Distributed Memory
System
E. Petit, F. Bodin, S.Matz, F. Dru, G. Papaure
epetit,fbodin,smatz,fdru,gpapaure@irisa.fr
Application
Analysis
Kernel
To export
Code and
Data Layout
Generator
Gpu Programming
Environment
Performance
Feedback
Path
Instrumentation
Hot Paths
computation
Memory
Instrumentation
Thread
Evaluation
Memory
Speculation
Computation
Thread
Selection
1
2
3
The ASTEX Pipeline
Architecture
Independent
Distributed Memory System
Heterogeneous
Co-processors
Input kernel code in C
Data layout mapping
on GPU memory system
Automatic OpenGL kernel code
generation for vertex and
fragment processors
C source
code
for CPU
C Kernel
Code
Memory
usage
information
ASTEX
Progrmmable
Vertex
Processor
Programmable
Fragment
Processor (Pixel
Processing)
Target
Rendering
(Frame Buffer)
Texture Memory
Texture
R
G
B
A
Texture
Cache
Fixed Pipeline
Transform
& Light
Primitive
CPU Assembly Rasterize
CPU Memory
Render to Texture
GPU Programming Environment
One of the current most challenging developpers’ issues addressed by ASTEX is to efficiently partition
applications into threads when programming heterogeneous systems such as CPU+GPU.
ASTEX integrates runtime measurements to identify hotpaths from the application and to then evaluate
sets of threads that minimize the amount of memory transfer and communication….