تعداد اتجاه النسخ

Every cudaMemcpy ends with a direction flag of type cudaMemcpyKind. It tells the runtime which way the bytes travel. 🧭

Use cudaMemcpyHostToDevice to upload inputs from CPU RAM into GPU memory before a kernel runs.

Use cudaMemcpyDeviceToHost to download results from GPU memory back into CPU RAM after computing.

To copy between two GPU buffers without touching the CPU, reach for cudaMemcpyDeviceToDevice.

There is even cudaMemcpyHostToHost, a plain CPU copy. You rarely need it since standard memcpy already does the job.

With cudaMemcpyDefault, the runtime infers direction from the pointers. It needs unified addressing to work safely.

The flag must agree with your real pointers. A device pointer with a HostToHost kind triggers an error or a crash.

Swapping HostToDevice and DeviceToHost is a frequent bug. The compiler stays silent, so read each call carefully. 🐛

The enum name spells the data flow: HostToDevice means from host to device. Match it to your source and destination.

Look at which buffer is the source and which is the destination, then pick the kind that names that exact path.

Prefer the named direction over cudaMemcpyDefault. Being explicit makes your intent obvious and catches mistakes sooner.

You are copying one device buffer into another device buffer, never touching the CPU. Which kind fits?

You met every cudaMemcpyKind: HostToDevice, DeviceToHost, DeviceToDevice, HostToHost, and Default. Match the flag to your pointers every time. 🎉