## Pull Request Overview

[Stevenanthony21b]

Pull Request Overview

This PR ensures that the compute shader local sizes (x, y, z) are rounded up to multiples of the subgroup size. It adds helper functions to compute trailing-zero counts, round values to powers-of-two multiples, and selects the minimal adjustment cost for the trio of dimensions.

• Added count_trailing_zeros, round_up_pow2_mul, and adjust_xyz helper functions
• Modified set_local_size_xyz to call adjust_xyz before assigning the local sizes
• Brought in <limits.h> for LONG_MAX

Comments suppressed due to low confidence (1)

src/pipeline.cpp:197

• No unit tests were added for the new rounding logic in adjust_xyz. Consider adding tests for various subgroup_size and (w,h,c) combinations, including edge cases like zero or non-power-of-two sizes.

adjust_xyz(&w, &h, &c, d->subgroup_size);

Originally posted by @copilot-pull-request-reviewer in Tencent/ncnn#2483 (review)