entry_proxy.hpp: `explicit entry_proxy(unsigned int mem_offset...` about viennacl-dev

[karlrupp]

the direct use of std::size_t instead of vcl_size_t is only a lack of unification after the vcl_size_t typedef was introduced. I consider this to be the necessary task associated with this issue.

The use of unsigned int, however, is intentional when dealing with sparse matrices for now. The reason is that size_t may differ on the host and on the GPU, hence one cannot use size_t for both and has to fall back to integer types.

[albertz]

And unsigned int is the same on both? How can that be? Isn't sizeof(int) quite dependent on the compiler and the target arch (32bit or 64bit)? And wouldn't it maybe sth different on the GPU, like always 32bit or so? (Just curious because I never did GPU programming.)

[karlrupp]

OpenCL defines the type 'cl_uint', which is guaranteed to be the same on both the host and the device. For all operations on OpenCL buffers, there is actually a lot of extra logic in place to make sure things get converted to the correct types.

However, 'cl_uint' is not necessarily available on the OpenMP backend, because OpenCL might be switched off. For this reason, 'unsigned int' is consistently used for indices for buffers in CPU RAM. Similarly, 'unsigned int' is controlled by the CUDA compiler to be the same on the host and on the device, hence we considered this to be the most consistent way of dealing with stuff.

On a related note: We make the index type a global typedef just like vcl_size_t so that user codes dealing with signed integer indices can better interface without the need for (potentially costly) conversions...