Sourcery VSIPL++ Performance
What performance benefits does Sourcery VSIPL++ offer?
Peak performance.
Take advantage of as much of the raw processing ability of your hardware as possible for your algorithm.
Multiple processors.
Scale up smoothly to add GFLOPS by adding processors.
Leveraging low level libraries through a portable API.
Get the benefits of a portable VSIPL++ and VSIPL library without sacrificing performance relative to alternative approaches.
Scalable Performance
With the addition of a few lines of code, a Sourcery VSIPL++ application scales smoothly to run in parallel on multiple CPUs.
Leveraging Low-Level Math Libraries
Sourcery VSIPL++ makes use of the vendor-optimized math libraries available on different platforms, imposing small to negligble performance overhead (depending on the problem size) for functions that are present in those libraries. In some cases, Sourcery VSIPL++ can fuse two or more low-level functions to achieve even better performance.
Optimization for Highest Performance
Sourcery VSIPL++ includes a built-in profiler to guide developers in improving the performance of their applications.
Sourcery VSIPL++'s flexible and extensible design facilitates optimization for performance. Developers can easily
experiment with data layout choices until they find the best one for their
application and target platform. Developers can also get under the hood
and hand code their inner
loops when necessary.
Performance Combined with Productivity
If cost is a factor, then the true measure of a system's performance is not GFLOPs--it's GFLOPs per SLOC (source line of code). Sourcery VSIPL++'s productivity benefits translate into real gains on that important metric.
Portable Performance
The same VSIPL++ code runs unchanged on platforms ranging from an x86 processor to a Power Architecture® processor to a Cell Broadband Engine™ processor. Sourcery VSIPL++ makes it easy to make the data layout changes that may be needed to get the best performance when porting application code between two different systems.