Eventos para medir y sincronizar

A CUDA event is a marker you drop into a stream. The GPU records when it reaches that point, letting you time and order work. ⏱️

Events use a cudaEvent_t handle, just like streams. You create it, record it, query it, and destroy it when finished.

Call cudaEventCreate for each event you need. Timing usually wants two: one before the work and one after.

cudaEventRecord places the event into a stream. The GPU stamps it the moment it finishes everything queued before it.

Record start, launch your kernel, then record stop, all in the same stream. The gap between them is the real GPU execution time.

Events are async, so the host must wait. cudaEventSynchronize blocks until the given event has actually been recorded on the GPU.

cudaEventElapsedTime returns the milliseconds between two recorded events. This is far more accurate than a host-side CPU clock.

Host timers include launch overhead and miss true GPU timing. Events sit inside the stream, so they capture exactly what the device did.

Events also synchronize streams. cudaStreamWaitEvent makes one stream pause until an event recorded in another has happened.

With stream-wait-event you express a dependency: streamB starts its work only after streamA reaches its event, without blocking the host.

Free events you no longer need with cudaEventDestroy. Like streams, they are cheap but should not be leaked in long-running apps.

Which function gives you the time between two events?

Events let you time GPU work precisely and order one stream after another. They are your tool for measuring and coordinating overlap. 🎯