matmul

Why naive matmul on a GPU is bandwidth-starved, and how tiling with __shared__ memory reduces global memory traffic by a factor of the tile size. The classic optimisation, with the kernel that demonstrates it.