Question

When a warp executes a conditional branch, what specific mechanism causes threads that do not meet the condition to become inactive during that instruction cycle?

Accepted Answer

When a warp executes a conditional branch, the mechanism used to handle threads that do not meet the condition is called a branch execution mask, also known as a predicate mask. A warp is a group of 32 threads in an NVIDIA GPU that execute the same instruction simultaneously in a lock-step fashion known as SIMT, or Single Instruction, Multiple Threads. Because every thread in a warp must execute the exact same instruction at the same time, the hardware cannot physically stop some threads while others proceed if they need to follow different code paths. When a branch instruction is encountered where threads have different outcomes, the hardware creates a bitmask representing which threads follow which path. A bitmask is a sequence of 32 bits where each bit corresponds to one thread in the warp, typically set to 1 if the thread is active and 0 if it is inactive. During the execution of instructions within a specific branch, the hardware hardware logic checks this mask. If a thread&#x27;s corresponding bit is set to 0, that thread is effectively disabled, meaning its state is not updated by the instruction, and it performs no meaningful work for that specific cycle. The warp continues executing in this manner, switching the mask to toggle between different paths until all threads have completed their respective instructions and re-converge at a common point in the code.

Home → All Courses → Engineering and Technology Courses → High-Performance GPU Programming for AI Workloads → Flashcard

When a warp executes a conditional branch, what specific mechanism causes threads that do not meet the condition to become inactive during that instruction cycle?