threadIdx, blockIdx, blockDim

Every thread runs the same code, so each one needs to ask: which piece of data is mine? CUDA answers with built-in variables. 📍

Inside a kernel, threadIdx tells a thread its position within its own block. The first thread in a block has threadIdx.x equal to 0.

The variable blockIdx tells a thread which block it belongs to within the grid. Block numbering also starts at 0.

You also get blockDim, the number of threads per block. If you launched 256 threads, then blockDim.x is 256 for every thread.

You never set these variables yourself. CUDA fills them in for each thread the moment the kernel launches.

Each of these has x, y, and z fields for 1D, 2D, or 3D launches. For simple arrays you usually only touch the x field.

On their own, threadIdx repeats in every block. To get a unique spot you must combine blockIdx and blockDim with threadIdx.

With blockDim 4, thread 2 in block 0 is element 2, but thread 2 in block 1 is element 6. The block shifts the start point.

These built-ins are read-only inside the kernel. Treat them as trusted coordinates describing exactly where your thread sits.

There is one more friend: gridDim holds the total number of blocks in the grid. It is handy when you loop over more data than threads.

These three variables are the raw materials for computing a global index, which is how every thread claims its own element.

Pick the variable that does the stated job.

You met the built-ins: threadIdx for position in a block, blockIdx for the block, and blockDim for block size. Together they pinpoint a thread. ✅