إعادة تشغيل الرسوم البيانية لتقليل العبء

Graphs exist to be replayed. One launch call submits the entire recorded sequence to the GPU at once.

You replay with cudaGraphLaunch, passing the instantiated exec and a stream. That is the whole submission.

Instead of ten launches per step, you pay the CPU launch cost just once for the entire graph.

In an iterative solver you call launch every step. The savings amortize across thousands of identical iterations.

The expensive instantiate step happens before the loop, so each replay inside the loop stays cheap.

If only parameters change, use cudaGraphExecUpdate to patch the exec instead of rebuilding from scratch.

The gain is largest with short kernels: when GPU work is brief, launch overhead dominated, and graphs erase it.

Submitting work as a batch also reduces CPU jitter, letting the GPU stay busy with fewer gaps on the timeline.

Graphs assume a stable structure. If the work pattern changes every step, the rebuild cost can outweigh the savings.

When finished, release both objects with cudaGraphExecDestroy and cudaGraphDestroy to avoid leaks.

The full lifecycle is three beats: capture the work, instantiate it once, then replay it many times.

Why do graphs cut launch overhead?

Replay with cudaGraphLaunch to fire a whole sequence in one call, amortizing setup across iterations. Update in place and free when done. Well done! 🏁