I have the following doubts:
- do openMP make pthread calls?
OpenMP is a cross-platform standard. The standard can be implemented in any way the implementor wants. Obviously on a platform without the POSIX threads library like Windows, OpenMP will not be implemented via pthreads. Since pthreads itself is a cross-platform standard, the OpenMP library may use it or go straight for the platform-specific low-level interface.
However, the OpenMP implementations provided by GCC and Clang do indeed call pthreads, as far as I know. At the very least they are compatible so that you can mix-and-match the libraries, e.g. use pthread's thread-local variables in conjunction with OpenMP's.
- how threads are created in openMP?
Again, specific to the implementation. Normally you don't need to worry about it
- is openMP a replacement to pthreads? or openMP and pthreads are entirely different?
The OpenMP interface caters to very specific styles of parallelization, like classic fork-join parallelization of loops. Pthreads is more general-purpose but requires you to do a lot of the things manually that OpenMP provides, such as distributing work across threads.
- if openMP and pthreads are different, then which gives better parallelism at C level i.e in openBLAS math library, openBLAS + openMP is better or openBLAS + pthreads is better?
When the programming model of OpenMP fits your use-case, it will save you work and brings with it low-level performance tunings that fit this style of parallelization. For example OpenMP has a thread-pool, handles CPU binding, and its synchronization primitives are tuned / tuneable to its style of parallelization (using longer spin-counts instead of sleeping directly).
As far as OpenBLAS or FFTW are concerned, I see the main benefit in that the OpenMP version can reuse the thread pool instead of using one thread pool per library. This reduces the number of context switches.
