[Numpy-discussion] Cython-based OpenMP-accelerated quartic polynomial solver

[Sturla Molden]

Sturla Molden <sturla.molden at gmail.com> wrote:

> OpenMP has a flush pragma for synchronizing shared variables. This means
> that OpenMP is not restricted to shared memory hardware. A "pragma omp
> flush" can just as well invoke some IPC mechanism, even network
> communication. 

By the way, while this is the case for C and Fortran, it is certainly not
the case for Cython. In a Cython prange block, a shared variable is
accessed by dereferencing its address. This requires shared memory. Pure
OpenMP in C does not, because shared variables are not accessed through
pointers, but are rather normal variables that are synchronized with a
pragma.

Cython actually requires that there is a shared address space, and it
invokes something that strictly speaking has undefined behavior under the
OpenMP standard. So thus, a prange block in Cython is expected to work
correctly on a laptop with a multicore processor, but it is not expected to
work correctly on a cluster.

IIRC, Intel's cluster OpenMP is based on MPI, which means the compiler will
internally translate code with OpenMP pragmas into equivalent code that
calls MPI functions. A program written for OpenMP can then run on any
cluster that provides an MPI implementation.