triton-ascend-case-elemwise-broadcast-2d

To optimize 2D elementwise division with broadcast on Ascend, you can use Triton to avoid splitting small dimensions and preload the divisor outside the inner loop, which maximizes kernel occupancy and UB utilization.

Handling large-broadcast axes paired with small non-broadcast axes in 2D tensors is best achieved by processing small dimensions completely without splitting, maximizing Unified Buffer (UB) utilization and overall kernel performance.

Preloading the divisor outside the inner loop improves Triton broadcast operations by reusing the preloaded divisor across iterations, which reduces redundant memory access and helps balance register pressure during computation.

Yes, you need to configure fixed NUM_BLOCKS to control kernel parallelism and SUB_M to balance UB utilization and register pressure, ensuring efficient grid-based scheduling for 2D broadcast division on Ascend.

This Triton Ascend broadcast optimization supports 2D tensor configurations with 131072x16 and 2048x131072 shapes, utilizing along-M kernel-level parallelism and inner row tiling for performance gains.