نقل البيانات من الجهاز وإليه

The CPU and GPU each have their own memory. The GPU cannot read your host arrays until the data lives on the device.

We call the CPU side the host and the GPU side the device. Moving data between them is a core part of GPU work.

First you reserve space on the GPU for your data. This device buffer is where inputs and outputs will live.

Next you copy your host arrays into those device buffers. This upload gets the GPU the numbers it needs.

With data on the device, you launch the kernel. It reads and writes the device buffers, never the host ones.

The output sits on the GPU until you fetch it. A download copies the result buffer back into host memory.

These calls are enqueued on the GPU, not finished at once. You must synchronize before trusting the results on the host.

The full cycle is upload, compute, download. Picture data making a round trip from host to device and back.

Copying over the bus is slow compared to compute. Each transfer can erase your speedup if you do it too often.

Chain several kernels on the same buffers before copying back. Minimizing round trips keeps the GPU win intact.

Device memory is limited, so release buffers you no longer need. Mojo's ownership rules help free them at the right time.

You want a GPU program to actually run faster end to end.

Upload inputs to the device, run the kernel, then download results and synchronize, while keeping costly transfers to a minimum. 🔁