장치 열거 및 선택

A single machine can hold several GPUs. To use them all, your program first needs to discover how many devices are present before it sends any work.

One call tells you how many GPUs the runtime can see. cudaGetDeviceCount writes that number into an int you hand it.

Each GPU gets an integer id from 0 up to count minus one. That device id is how you point the runtime at one specific card.

You choose which GPU your next calls target with cudaSetDevice. After this, allocations and launches go to that card.

At any moment exactly one GPU is the current device for the calling thread. Every cudaMalloc or kernel launch lands on whatever device is current.

Before trusting a GPU you can read its specs. cudaGetDeviceProperties fills a struct with name, memory size, and compute capability.

The properties struct is rich. Fields like name and totalGlobalMem help you pick the best card or skip one that is too small.

A pointer from cudaMalloc belongs to whichever GPU was current then. Using it while another device is current is an error waiting to happen.

Behind each GPU sits a context that holds its allocations and streams. Switching the current device switches you into that device's context.

A common pattern is a loop that calls cudaSetDevice for each id, then does setup on that card. This is how you spread work across every device.

If a helper changes the current device, switch it back when done. Leaving the current device changed can surprise the rest of your code.

Recall which call chooses the GPU your next allocations and launches will use.

You count GPUs, pick one with cudaSetDevice, and inspect it with properties. Each device owns its own memory and context. Next: splitting work across them. ✨