Atomic accumulation combines concurrent contributions at shared global-memory locations using atomic read-modify-write operations. In Triton, tl.atomic_add returns the location’s previous value; a mask can skip invalid addresses. Initialize the output buffer to zero before launch, or contributions will add to uninitialized values. section 4.1.3
tl.atomic_add(pointer, val, mask=None, sem=None, scope=None)
Triton’s tl.atomic_add accepts pointers, values, and an optional mask. section 4.1.3
Shared output address
Four program instances can add partial results to one address. Their execution order may vary, producing different bitwise floating-point results because addition is not associative. section 4.1.3
