变更摘要

本 PR 针对 gmmaq（grouped matmul activation quant）算子做性能优化，核心改动位于 GroupedMatmulActivationQuantTiling950::DoLibApiTiling()（文件 grouped_matmul_activation_quant_tiling.cpp）。改动在 M 分块计算完成后新增了一个实验性逻辑：当输入满足特定目标形状（inputParams_.mSize == 51200、nSize == 1536、kSize == 4096、groupNum == 80）且 baseM/baseN 均为 256（GmmConstant::BASIC_BLOCK_SIZE_256）时，将 basicTiling_.baseM 下调一个 16 行的 cube 单元至 224（EPILOGUE_DB_EXPERIMENT_BASE_M），从而为 Epilogue 腾出可分配两个 UB 输出槽（双缓冲）的空间。该实验按形状范围收敛，待性能结果确认后再决定是否推广，改动不涉及接口、数据结构和配置变更。

主要改动

• 新增形状限定实验分支: 在 basicTiling_.baseM = std::min(alignedBaseM, maxBaseMByUb) 计算之后，新增一段注释为 Experiment 的条件判断，通过 inputParams_ 的 mSize/nSize/kSize/groupNum 精确匹配目标形状，保证实验只作用于该特定场景。
• 下调 M 分块以适配 Epilogue UB 双缓冲: 当 baseM 与 baseN 均为 GmmConstant::BASIC_BLOCK_SIZE_256 时，将 basicTiling_.baseM 由 256 改为新增常量 EPILOGUE_DB_EXPERIMENT_BASE_M（224），即减少一个 16 行 cube 单元，使 Epilogue 能分配两个 UB 输出槽，属于 GroupedMatmulActivationQuantTiling950::DoLibApiTiling() 中纯本地、不改变外部接口的 M 分块调整。

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.
You can self-configure the PR merge rules for this repository. For more details, please refer to Here.
For more, you also can visit HICANN.

PR Approval Progress

✅ Congratulations! All modules have met the lgtm and approve requirements.

Module Approval Details

💡 Tip:

• Committer can comment /approve or /lgtm
• Commenting /approve implies both code review (lgtm) and intent to merge (approve)

CLA Signature Pass

zhoushaolong, thanks for your pull request. All authors of the commits have signed the CLA. 👍

compile

compile

/approve

/lgtm

The following label is not ready.

ci-pipeline-passed: The ci-pipeline-passed label is expired as added over 72 hours. Please use /compile to recompile.

compile