Pull Request已成功合入, 合并人@CANN-robot
(感谢 myx_32 的贡献)Thanks for your pull-request.
The full list of commands accepted by me can be found at here。
You can get sig-info at here
PR Approval Progress
✅ Congratulations! All modules have met the lgtm and approve requirements.
Module Approval Details
| module | lgtm status | approve status |
|---|---|---|
| branch-9.0.0 | ✅ dengjing_aoe, YeZZzzz1 (2/2) | ✅ YeZZzzz1 (1/1) |
💡 Tip:
- Committer can comment
/approveor/lgtm- Commenting
/approveimplies both code review (lgtm) and intent to merge (approve)
CLA Signature Pass
myx_32, thanks for your pull request. All authors of the commits have signed the CLA. 👍


🔵 source code change are detected, tasks labels is removed in this pull request!


问题/功能描述
针对特定硬件平台(David)上使用N_BUFFER_33调度策略时,矩阵乘法分片算法性能不佳的问题进行优化。原算法直接使用硬件L0C缓存大小计算分片参数,在该特定场景下可能导致非最优的分片策略。本次修改旨在通过调整关键参数的计算逻辑,提升该场景下的计算性能。
修改方案描述
在GetParasCombo函数中,为核心参数mnMax的计算逻辑引入了调度类型相关的条件分支。对于非N_BUFFER_33的调度类型,沿用原有的L0C缓存大小进行计算。对于N_BUFFER_33调度类型,则使用一个经过优化的固定值(128KB)替代硬件L0C大小进行计算,从而为David平台上的该特定调度模式生成更高效的计算核函数参数组合。


🔵 source code change are detected, tasks labels is removed in this pull request!


流水线任务触发成功
任务链接 [4df96d865d4d4523b918952449cad8c0][流水线指导]
| 任务名称 | 状态 | 日志 | 下载链接 |
|---|---|---|---|
| Compile_Ascend_X86 | ✅ SUCCESS | >>>>> | >>>>> |
| Compile_Ascend_ARM | ✅ SUCCESS | >>>>> | >>>>> |
| SCA | ✅ SUCCESS | >>>>> | |
| anti_virus | ✅ SUCCESS | >>>>> | |
| codecheck | ✅ SUCCESS | >>>>> | |
| Check_Pr | ✅ SUCCESS | >>>>> | |
| pre_comment | ✅ SUCCESS | >>>>> | |
| StaticCheck_codespell_check | ✅ SUCCESS | ||
| StaticCheck_link_validity_check | ✅ SUCCESS | ||
| StaticCheck_resource_existence_check | ✅ SUCCESS | ||
| StaticCheck_tag_closed_check | ✅ SUCCESS | ||
| StaticCheck_markdownlint | ✅ SUCCESS | ||
| UT_Test_3 | ✅ SUCCESS | ||
| UT_Test_2 | ✅ SUCCESS | ||
| UT_Test_1 | ✅ SUCCESS | ||
| UT_Test | ✅ SUCCESS | ||
| UT_Test_report | ✅ SUCCESS | >>>>> |
[2026-03-30 16:31:21] CI执行结束


/lgtm


Review Guide
This pull-request passes review.
Committers who wrote a comment of /approve are: YeZZzzz1.
Reviewers who wrote a comment of /lgtm are: YeZZzzz1, dengjing_aoe.


描述
现象:a5场景下,部分case(在a3场景下执行正常)开启N_BUFFER_33特性,可能会出现tiling计算失败/aicore报错的问题
修改:针对a5环境下N_BUFFER_33特性使用场景,适配tiling计算逻辑
关联的Issue
#350
测试
对obp迁移用例(修改前部分用例会失败)进行批跑,结果正常无报错

文档更新
类型标签