NAME

dispatch_apply — schedule blocks for iterative execution

SYNOPSIS

#include <dispatch/dispatch.h>

void
dispatch_apply(size_t iterations, dispatch_queue_t queue, void (^block)(size_t));

void
dispatch_apply_f(size_t iterations, dispatch_queue_t queue, void *context, void (*function)(void *, size_t));

DESCRIPTION

The dispatch_apply() function provides data-level concurrency through a "for (;;)" loop like primitive:

size_t iterations = 10;

// 'idx' is zero indexed, just like:
// for (idx = 0; idx < iterations; idx++)

dispatch_apply(iterations, DISPATCH_APPLY_AUTO, ^(size_t idx) {
	printf("%zu\n", idx);
});

Although any queue can be used, it is strongly recommended to use DISPATCH_APPLY_AUTO as the queue argument to both dispatch_apply() and dispatch_apply_f(), as shown in the example above, since this allows the system to automatically use worker threads that match the configuration of the current thread as closely as possible. No assumptions should be made about which global concurrent queue will be used.

Like a "for (;;)" loop, the dispatch_apply() function is synchronous. If asynchronous behavior is desired, wrap the call to dispatch_apply() with a call to dispatch_async() against another queue.

Sometimes, when the block passed to dispatch_apply() is simple, the use of striding can tune performance. Calculating the optimal stride is best left to experimentation. Start with a stride of one and work upwards until the desired performance is achieved (perhaps using a power of two search):

#define STRIDE 3

dispatch_apply(count / STRIDE, DISPATCH_APPLY_AUTO, ^(size_t idx) {
	size_t j = idx * STRIDE;
	size_t j_stop = j + STRIDE;
	do {
		printf("%zu\n", j++);
	} while (j < j_stop);
});

size_t i;
for (i = count - (count % STRIDE); i < count; i++) {
	printf("%zu\n", i);
}

IMPLIED REFERENCES

Synchronous functions within the dispatch framework hold an implied reference on the target queue. In other words, the synchronous function borrows the reference of the calling function (this is valid because the calling function is blocked waiting for the result of the synchronous function, and therefore cannot modify the reference count of the target queue until after the synchronous function has returned).

This is in contrast to asynchronous functions which must retain both the block and target queue for the duration of the asynchronous operation (as the calling function may immediately release its interest in these objects).

FUNDAMENTALS

dispatch_apply() and dispatch_apply_f() attempt to quickly create enough worker threads to efficiently iterate work in parallel. By contrast, a loop that passes work items individually to dispatch_async() or dispatch_async_f() will incur more overhead and does not express the desired parallel execution semantics to the system, so may not create an optimal number of worker threads for a parallel workload. For this reason, prefer to use dispatch_apply() or dispatch_apply_f() when parallel execution is important.

The dispatch_apply() function is a wrapper around dispatch_apply_f().

CAVEATS

Unlike dispatch_async(), a block submitted to dispatch_apply() is expected to be either independent or dependent only on work already performed in lower-indexed invocations of the block. If the block's index dependency is non-linear, it is recommended to use a for-loop around invocations of dispatch_async().