__shfl_down_sync для свёрток

A shuffle lets one lane read another lane's register value directly. Data hops between threads with no shared memory and no barrier in between.

The workhorse for reductions is shfl_down_sync. Each lane grabs a value from a lane a fixed number of positions higher in the warp.

The call takes an active mask, the value to share, and an offset. Lane i receives the value held by lane i plus offset.

A warp sum just adds the shuffled value back in. Lane i takes its neighbor's number and folds it into its own running total.

Like a tree reduction, the offset starts at 16 and halves each step: 16, 8, 4, 2, 1. After five steps the whole warp is summed.

A warp has 32 lanes, and 2 to the fifth power is 32. So five halving steps are exactly enough to combine all 32 values into one.

After the loop, the complete warp total sits in lane 0. The other lanes hold partial junk, so only lane 0 should write the result out.

Because exchange happens through registers in lockstep, you skip syncthreads entirely. The sync suffix already coordinates the masked lanes.

A warp shuffle reduction beats the shared-memory version: fewer instructions, no bank conflicts, and no barrier stalls. It is the fast path for the last 32.

If some lanes already exited, a full mask is wrong. Capture the live set with activemask so each shuffle only touches present lanes.

Big reductions often reduce each warp with shuffles, then combine warp results in shared memory. Shuffles handle the cheap inner level beautifully.

Think about which lane holds the answer when the loop finishes.

You summed a warp with shfl_down_sync: halve the offset five times, and lane 0 holds the total, no barriers needed. Next: ballot and vote. ✨