[Puzzle 12, Test2] Question regarding shared memory

[VeritasYin]

Hi,

For this puzzle, it seems that shared cuda arrays are lazily initialized. Since SIZE is 15 and TPB is 8, the last element of Block 1 is not overwritten, but the value got carried over from Block 0, even after calling cuda.syncthreads(). So is this a normal behavior or actually a leak?

My original implentation assumed that the shared array would all be intialized as 0, so it was safe to perform the sum over the entire block. Given the above findings, there are two options for those unaligned block: 1) fill the rest as 0 or 2) check the index bounds for sum ops ((i+step) < size). While both pass the test case, the latter gives a weird dependency graph for Block 1 as attached, which raises my concern.