tl.reduce combines tensor values using a programmer-provided binary function. Like other Triton reduction operations, it works on a tensor tile local to one program instance; its axis selects a tile dimension, not the full global tensor. section 4.1.2
A reduction across the launch grid needs orchestration beyond tl.reduce: write partial results to global memory and reduce them in another kernel, or use atomic operations. section 4.1.2
Figure 4.3 Two-pass reduction dataflow. The first kernel reduces tiles independently, writing one partial sum per program instance to global memory. The second kernel reduces those partials to a single output value
Shows the two-pass pattern: program instances reduce tiles independently, then a second kernel combines their partial sums. section 4.1.2
