Hi!
I notice that access of shared memory in SpdDownsampleMip_2 is not bank conflict free. The output will use bank[0: 15] twice and not use bank [16: 31] in each of two warps(32 threads).
We can reduce bank conflicts at best with the following access pattern in SpdStoreIntermediate: (x * 2 + y % 2, y * 2 + (y >> 1） & 1), which use every bank[0: 31] in each warp.
Correct me if i am wrong.
Hi!
I notice that access of shared memory in SpdDownsampleMip_2 is not bank conflict free. The output will use bank[0: 15] twice and not use bank [16: 31] in each of two warps(32 threads).
We can reduce bank conflicts at best with the following access pattern in SpdStoreIntermediate: (x * 2 + y % 2, y * 2 + (y >> 1） & 1), which use every bank[0: 31] in each warp.
Correct me if i am wrong.