变更摘要

本 PR 对 norm 系列算子的 arch35 路径做去硬编码清理：将代码中写死的 32 字节 UB block（及 256 字节向量寄存器）常量，统一替换为平台查询接口（host 侧 Ops::Base::GetUbBlockSize(context)/commonParams.blockSize，kernel 侧 platform::GetUbBlockSize()/Ops::Base::GetVRegSize()），涉及 add_layer_norm_quant、layer_norm_grad、layer_norm_grad_v3、layer_norm_v3、layer_norm_v4 等算子的 tiling 与 kernel 代码，使 UB/VL block 大小随平台自适应。

主要改动

• kernel 侧 block 常量改为平台接口：在 layer_norm_v3_common.h、layer_norm_v4_regbase_common.h、layer_norm_v4_welford.h 中将 constexpr static int64_t BLOCK_SIZE = 32 改为 platform::GetUbBlockSize()；layer_norm_grad_api.h 中 BLOCK_SIZE 改为 Ops::Base::GetUbBlockSize()，VREG_SIZE = 256 改为 VECTOR_LENGTH = Ops::Base::GetVRegSize()。
• host 侧 tiling 使用 commonParams.blockSize：layer_norm_v3_regbase_no_reduce_tiling.cpp、layer_norm_v3_regbase_norm_not_equal_params_tiling.cpp、layer_norm_v3_regbase_two_pass_perf_tiling.cpp、layer_norm_v3_welford_multi_reduce_tiling.cpp、layer_norm_v4_regbase_two_pass_perf_tiling.cpp 删除 *_BLOCK_SIZE = 32/*_B32_ALIGN_NUM 等常量，改用 commonParams.blockSize / sizeof(float) 计算 B32 对齐数。
• add_layer_norm_quant 去硬编码：host 侧 add_layer_norm_quant_tiling_arch35.cpp 删除 BLOCK_SIZE，colsAligned_ 及 blkFp32Nums、constTmpBufSize 改用 this->blockSize_；kernel 侧 add_layer_norm_quant_regbase_helper.h 中 FLOAT_BLOCK_ELEM = 8 改为 blockSize_ / sizeof(float)。
• layer_norm_grad/layer_norm_grad_v3 统一 block 获取方式：layer_norm_grad_base.h 删除自定义返回 32U 的 GetUbBlockSize()，相关调用改用 API 中定义的 BLOCK_SIZE；layer_norm_grad_v3_base.h 的 GetUbBlockSize() 改为返回 Ops::Base::GetUbBlockSize()。
• 对齐函数参数化：layer_norm_v3_regbase_norm_not_equal_params_tiling.cpp 的 AlignB32(val) 改为 AlignB32(val, alignNum)，由调用方基于 commonParams.blockSize 传入对齐数，去除对固定 32B 对齐的假设。

代码审查

✅ 未发现问题

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

wangpengbo26, thanks for your pull request. All authors of the commits have signed the CLA. 👍