A kernel launch can fail without updating its output, leaving old or stale data. Call `cudaGetLastError()` immediately after a launch to check for an error. section 3.16 section 11.12
Catch an oversized shared-memory request
If a kernel requests more shared memory per block than the hardware allows, its launch fails and the output may remain stale. Check with `cudaGetLastError()` after the launch to catch the failure. section 11.12
