OpenMP and GCC Optimization

So I encountered and interesting little affect while compiling some C code in OpenMP using GCC. Seems that my code ran fine without any automatic optimization flags (-O2, -O3), but when I turned them on, went haywire. The sequential version wasn’t affected.

I tracked down the problem and found that it was actually do to my use of the private variable in the OpenMP directive. Here is a sample of the problem.

 

#pragma omp parallel

{
	#pragma omp for private(i, j, new_j, skip_next)                      	for(i = 0; i < n; i++)
        	{                                                                    	// Set each thread on one row of the loop
                	skip_next = 0;
                	for(j = 0; j < n; j++)                                       	{
				....

The problem actually occured because I did not have ‘skip_next=0′ on line 5 of this code. This didn’t seem to matter when optimization was disabled but did cause incorrect semantics when I turned it on. Seems like OpenMP says you must assume that all privates are uninitialized when you start a new directive block. If you want them to be private but start with some default value, do what I did above or add the variable to the firstprivate() attribute instead of the private attribute:

 

#pragma omp for firstprivate(skip_next) private(i, j, new_j)