mirror of
https://github.com/boykopovar/AnyPS5.git
synced 2026-10-22 08:33:12 -05:00
On a host subgroup narrower than 32 lanes, threadBit reads an EXEC written with s_mov_b32 or s_mov_b64 at bit LaneId & 31 of EXEC_LO, and LaneId is the subgroup invocation id. Every lane at or above the host subgroup size therefore reads the bit of a lower lane, in wave32 and wave64 alike. An EXEC produced by a compare stays consistent because its ballot is subgroup-local as well. Found while investigating the llvmpipe failure of #751.