Incorrect Kmeans OpenACC code

[ouankou]

The OpenACC directives seem fine and can be mapped to OpenMP code. However, there are many errors in the source.

For example:

1. Missing headers: cluster(), allocateMemory(), deallocateMemory(), etc.
2. Mismatch data type usage: float* features vs float** features.
3. Functions with incorrect code: undeclared variables in allocateMemory() and deallocateMemory()
4. Mixed C++ and CUDA code: CUDA API tex1Dfetch().

We can either fix them all or start with the official OpenMP CPU version.

[ouankou]

The OpenMP GPU offloading version is created from scratch based on the official OpenMP and CUDA versions. Then the OpenACC version is created based on the OpenMP GPU offloading version.
The original OpenMP CPU version uses omp_get_thread_num() to do work chunking for multiple threads. However, OpenACC doesn't have similar APIs. We have to calculate the chunk id manually (check this commit passlab/NeoRodinia@3d88f2a).

The original OpenACC version doesn't work at all, so it's abandoned.