GPU support

[kazuyukitanimura]

1. Explicit data transportation between main memory and GPU memory
   This is a mistake that I made when I started this project. I believed that the code works for CPU works exactly the same for GPU (thought it is the claim of OpenCL). But, in fact, it is a miss leading. In narray.c, there is a line
   ary->buffer = clCreateBuffer(context, CL_MEM_READ_WRITE|CL_MEM_USE_HOST_PTR, na_sizeof[type]*total, ary->ptr, NULL);
   The specification does not explain clearly, but CL_MEM_USE_HOST_PTR works fine for CPU main memory but not for GPU memory (the newest version of OpenCL might be different). For GPU memory, one has to explicitly transfer the data from main memory to GPU memory and take care of pointers. In order to this, I might have to re-write the whole program that I added to the original NArray since I kept using the main memory pointers :P
2. Overhead
   The overhead of data is quite expensive. If this is a normal OpenCL programing, one can save the reusable data on the GPU memory; however, the interpreter(or virtual machine) of ruby runs using the main memory so that all the data on the GPU memory have to be sent back to the main memory every time after each OpenCL NArray method is executed. This causes a lot of data transaction between the memories and make it difficult to get performance gain.
3. Applications
   The applications that uses this project can get nice performance gain only if the applications need to compute a large size of NArray instances (since you can parallel data more). However, considering about the large overhead mentioned above, I don't know any application that needs to compute that large NArray instances. I will get excited and more motivated once I know that there are specific applications that have to utilize OpenCL NArray.