El lanzamiento con triple ángulo

You do not call a kernel like a normal function. You launch it, telling the GPU how many threads to spin up at the same time. 🚀

The launch uses CUDA's special triple angle brackets. Inside them you give a launch configuration before the normal argument list.

The first value sets how many blocks launch. A block is a group of threads that run together and share fast on-chip memory.

The second value sets threads per block. Total threads launched equals blocks times threads per block.

After the brackets, you pass the kernel's normal arguments in parentheses, just like any C++ call.

A launch is asynchronous: the CPU queues the work and keeps going immediately. The GPU runs the kernel in the background.

The launch numbers do not have to be literals. You often compute the block count from your data size at runtime.

For grids and blocks in 2D or 3D, you use a dim3 value. It bundles x, y, and z sizes into the launch.

A safe default for threads per block is 128 or 256. Values must be multiples of 32 and stay at or below 1024.

If you ask for too many threads per block, the launch fails silently. You must check for errors right after launching.

Read a launch as: run this kernel across this many blocks of this many threads, with these arguments. Simple once it clicks.

Time to check the launch configuration.

You launch with kernel<<>>(args). Blocks times threads is your total, and launches run async. You are launching like a pro! 🎉