B5 - parallel for

#pragma omp parallel for
for (i = i_start; i <= i_end; inc_exp) {
	// statements
}

• this creates a work-sharing construct, and also shares work

• the loop index is private by default

• the loop must not have any early exit statements (break, but continue is allowed)

• it must have a single entry point at the top and a single exit point at the bottom

• the index start and end values must not change during the loop

• increment expressions only change i by a fixed amount each iteration

• ordered clauses can be used, which make each thread wait until the previous iteration has finished its ordered section before doing its own

#pragma omp parallel for ordered
 for (i=0; i<n; i++) {
	 a[i] = costly_func(i);
	 #pragma omp ordered
	 printf("%d %f\n", i, a[i])
 }

• only loops in which there is no data-dependence on another iteration can be parallelised using this

// cannot parallelise
for (i=1; i<n; i++) {
	a[i] = a[i] = a[i-1];  // depends on previous iteration
}

// can parallelise
for (i=1; i<n; i+=2) {  // only odd indices are modified
	a[i] = a[i] = a[i-1];  // data-dependence on even (uniterated) indices
} 

last/first private

int x = 10;

#pragma omp parallel private(x)
{
    printf("Thread %d: x = %d\n",
           omp_get_thread_num(), x);
}

• in the above code, each thread gets an independent, uninitialized copy of x

int x = 10;

#pragma omp parallel firstprivate(x)
{
    printf("Thread %d: x = %d\n",
           omp_get_thread_num(), x);
    x += omp_get_thread_num();
}

printf("Original x = %d\n", x);

• here, each thread gets an independent, initialised copy
• the original value remains unchanged, so after the parallel construct, x = 10

int x = 10;

#pragma omp parallel for lastprivate(x)
for (int i = 0; i < 4; i++) {
    x = i * 10;
}

printf("x = %d\n", x);

• now, the last logical iteration's value is copied
• the output will be 10