로드-동기화-계산 패턴

Tiling follows one rhythm in every kernel: load a tile into shared memory, sync, then compute from the fast copy.

In the load phase, each thread reads one element from global memory and stores it into a shared-memory tile its whole block can see.

Loading is a team job. Every thread fetches its slice, so together the block stages one full tile with a single coalesced pass.

Before anyone reads a neighbor s value, all threads must finish loading. __syncthreads is the barrier that guarantees the tile is ready.

Skip the barrier and a thread may read a tile slot that its neighbor has not written yet. That is a race and gives wrong results. 💥

Now every value lives on chip. In the compute phase threads read tile entries freely, since shared memory is orders faster than global.

Because the tile sits in shared memory, a value loaded once is reused by many threads with almost no extra cost. That is the whole win.

If the compute phase writes back into the same tile for a next step, add a second __syncthreads before reusing those slots.

Every thread in the block must reach __syncthreads. If some skip it inside a branch, the kernel deadlocks or misbehaves.

Load, sync, compute is a template you will reuse for convolutions, matrix multiply, and reductions throughout this category.

The shared tile must fit the block. Tile width usually equals blockDim, so each thread owns exactly one slot to load and reuse.

Why is __syncthreads needed between load and compute?

Tiling is load, sync, compute: stage a tile cooperatively, barrier so it is complete, then reuse it fast on chip. ✅