Hint for Q11: 1D Convolution about gpu-puzzles HOT 4 CLOSED

The key thing to remember is that shared memory is per block so if it is two blocks you get that same shared memory twice. One chunk for the first block and one for the second. Threads only see the shared memory for their block.

from gpu-puzzles.

Are we allowed to extend the shared memory space to TPB + MAX_CONV + (MAX_CONV - 1) since, for the convolution on the last few indices in the first block, we would need access? Currently, my max global reads are increasing to 4 and write to 1

from gpu-puzzles.

Ah, yes sorry. You are allowed to have multiple shared memory arrays. You shouldn't need to create 1 bigger one though.

from gpu-puzzles.

Thanks a lot!

from gpu-puzzles.