No matter what you do, you're going to have to target the Phi much as you would target OpenCL, and run dedicated code on it that way.
There is a two pronged attack Intel is using with Phi.
a. Apps that run on High Performance Clusters (i.e., computational clusters). Typically these are clusters of Linux boxes. That software ports over to Phi cards relatively easily because the Phi card is running Linux. This is the low-hanging-fruit so approach is more evolved.
If Phi caps the encroachment of more mainstream GPGPU cards from supercomputer design wins, then Intel will be happy.
b. Intel maps the OpenCL model onto the Phi card. You can run OpenCL code on the Phi card. You probably won't get a the highest match to peak theoretical performance, but you can. This isn't as mature and Intel dog and pony demos don't really push this option.... yet.
http://software.intel.com/en-us/art...ming-guide-for-the-intel-xeon-phi-coprocessor
http://software.intel.com/en-us/art...l-applications-for-intel-xeon-phi-coprocessor
These cards are not currently supported under Windows (as in, there's no drivers for them at all).
This too is why the OpenCL lags behind. In the Windows ( and OS X ) space there is a wider base of apps that are "shipping" computation off to "helpers" with OpenCL. In Linux, that more likely would be done with MPI (e.g. OpenMPI) and OpenMP which have been around alot longer.
Even if there were, I know for a fact that neither MAXON, ChaosGroup, LAUBlab, or Adobe is interested in supporting this card...... It would be analogous to them porting to a third option- OpenCL, CUDA, and now the Phi-
It would more so be a variant of a OpenCL where the n-way of the architecture leaks back through the OpenCL interface. The much larger blocker for folks like Adobe , MAXON , etc
far more seems to be they don't have Linux apps and don't "plug-into" Linux based HPC clusters well.
This is primarily why I said Apple could make a difference if they wanted to try. It is the "hook two Unix instances to do work together" drivers ( which pragmatically includes the OpenCL glue to do the dispatching of the computations/data to the Phi card) that is the essentially the missing part.
I suspect though that it is a much higher Apple priority to stoke the fires under the chairs under Intel folks in the integrated graphics group. With Haswell GPUs, Intel is only just catching up to the full OpenCL API. The catch-22 is that the performance of the HD 4000 and HD 5000 series isn't a huge barn-burner (compared to what AMD and Nvidia are currently delivering) while they do have a barn-burner in Phi but the OpenCL interface is missing or immature on anything but Linux.
and we don't even have full support for both OpenCL and CUDA in the CG industry.
In industries where they are charging a relatively high amount per computation grid node. .... They are just going to drag their feet on this. If one node could do the work of three that could result in a 2/3% reduction in income for them. Very few companies are going to accelerate investment into that.
