Parallella Community

Just to avoid any confusion, STDCL is a C API, similar to the level of CUDA, but users are not required to use STDCL in order to use the OpenCL implementation for Parallella that is also provided by the COPRTHR SDK.

I was not really thinking about the fact that STDCL is a C API. While I did not intend on giving offense, I seem to have gotten carried away in my arguments. For that I apologize. STDCL is not my kind of library, but I can still appreciate the amount of work that went into wrapping the OpenCL API's into an easier to use C API.

What JAR is referring to is that the standard is very much designed around GPU architectures ca. 2009. Yes, you can program other compute devices with OpenCL, but they were not the driving architectures in the design. This can be seen most notably in the memory model and the lack of strong signaling support. If you were to design a programming model around multi-core devices that are not massively multi-threaded, with fully divergent cores, strong inter-core communication, and more general memory architectures, it would not look like OpenCL, but it would have some things in common with it. Everything evolves for a reason and GPGPU was the reason for OpenCL. This is clear in the design. That is not a criticism. I have invested considerable time in OpenCL, it can be leveraged for other "compute devices", and we are determined to leverage OpenCL to provide a productive parallel programming API for Parallella. OpenCL is a great foundation, but it must be adapted to support non-GPU architectures.

I am not sure I can find a current example that uses these extensions. They were used to share data between cores a very long time ago and I am not sure they are still compatible with current code base. Its been 2 years since I tested these extensions. The need for these extensions contributed to the evolution of our currently recommended programming model for Epiphany, which is threaded MPI. This model has exhibited the highest performance reported to date using a standard programming API. There are technical reasons that lead me to believe APIs designed for SMP architectures will not prove to be as effective.

-DAR

DAR
Thank you for your response, i will have a look for MPI

Gideon