Tail-Safety Constraints【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skillsRead this file when a kernel has tile tails, odd row splits, or partial GM boundaries.GoalKeep tail handling correct without corrupting the stable local-tensor shape assumptions used by lowering and simulation.1. Core ruleApplyvalid_m,valid_n, andvalid_kat GM read and write boundaries. Do not shrink local tensor shapes for every tail tile.Repository expectation:local buffers remain full tile-sizedonly the GM boundary slices use the tail sizes2. Why this rule existsStable local tensor shapes make lowering and simulator behavior predictable. If you start shrinking local buffers for tails, it becomes much easier to create shape drift between the intended logical tile and the actual staged buffer.3. Standard cube - vec half-row writeback patternFor many cube - vec kernels, use the standard half-row split:half_rows CeilDiv(valid_m, 2)row_begin GetSubBlockIdx() * half_rowsrow_end Min(row_begin half_rows, valid_m)row_count row_end - row_beginThis keeps odd-row tails stable across the two vec subblocks.4. Common symptoms of tail bugsTail issues often look like this:aligned shapes pass but odd shapes failonly the last tile is wrongone vec subblock is correct and the other is garbageoutput shape looks right but the boundary rows or columns are corruptedWhen this happens, inspect the GM boundary slices first. Do not start by changing the local buffer shape.Special case:if the kernel is a normalized online softmax with runningrow_max/row_sum, GM-boundary slicing alone is not enough; invalid score columns must behave like-infbeforerowmaxreadagent/references/constraints/online-softmax-tail.md5. Quick checklistBefore accepting tail logic, verify:valid_m,valid_n,valid_kcome from the current tile boundarylocal buffers still use fullTILE_*shapesGM load boundaries use the valid sizesGM store boundaries use the valid sizesvec half-row split usesCeilDivand clamps withMinat least one odd-size case has been testedfor normalized online softmax, verify score-domain invalid columns are masked beforecmaxFiles to studyagent/example/kernels/a5/basic_cube_vec_mix.pyagent/example/kernels/a5/matmul_half_splitn_bias10p2_vf.pyagent/example/kernels/a5/matmul_rowwise_norm.pyagent/example/kernels/a5/vec_cube_abs_sqrt_matmul.pyagent/example/kernels/a5/vec_unaligned_gm_to_ub_pad.py【免费下载链接】cannbot-skillsCANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体本仓库为其提供可复用的 Skills 模块。项目地址: https://gitcode.com/cann/cannbot-skills创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考