[Gomp-discuss] Dissecting OpenMP

[Scott Robert Ladd]

Here is a very simple OpenMP program; as I undertsnad it, one of our 
first goals is to compile a program such as this with GOMP.

#include <stdio.h>

int main(void)
{
    #pragma omp parallel
    puts("Hello, world!");

    return 0;
}

You can't get much simpler than that.

The OpenMP pragma is likely to do the following:

1) Take the associated code block (in this case, a single statement), 
and convert it into a function. Let's call this par_reg_func.

2) Find out how many processors are available (or configured, since 
environment variables can force a number of processors different from 
the actual number recognized by the OS).

3) Launch a thread, per processor, that invokes par_reg_func.

The situation becomes more complex when parallalizing a loop, since the 
extracted "parallel region function" would need to be parameterized, 
shared variables passed, etc.

In a manner of speaking, OpenMP writes new code the replace what the 
programmer has specified. Private variables must become local variables 
in the generated parallel function, for example.

I am not certain what our policy should be in terms of discussing how 
other compilers -- Intel, in particular -- generate code for code such 
as the above. I've been tempted to look at how Intel implements OpenMP, 
but an concerned with "clean room" issues.

--
Scott Robert Ladd
Coyote Gulch Productions (http://www.coyotegulch.com)
Software Invention for High-Performance Computing