Closed
[Bug-Report|缺陷反馈]: CMCT block_mmad_mx GM2L1搬运未按平台宏隔离预分配参数 #8041
wangzitaoAdd on 18 days agoClosed at 10 days ago
18 days ago Added label: bug-report
18 days ago Associated pull request
wangzitao
18 days ago Comments:
18 days ago Comments:
18 days ago Assign wangzitao_leo as responsible person
12 days ago Modify description of issue
10 days ago Close issue
10 days ago changed custom state from 进行中 to 已完成
9 days ago Added label: resolved


Describe the current behavior / 问题描述 (Mandatory / 必填)
matmul/common/cmct/block/block_mmad_mx.h中 GM2L1 的AscendC::DataCopy/AscendC::DataCopyPad直接传入预分配参数enablePreReadIgnoreSync=true,未像matmul/common/cmct/block/block_mmad_pertile.h那样按平台宏隔离。现状与风险:
__NPU_ARCH__ == 9201/9202上真正生效,其他架构(如 3510/5102/5162)虽声明但实现体为空;Environment Information / 环境信息 (Mandatory / 必填)
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
matmul/common/cmct/block/block_mmad_mx.h;AscendC::DataCopy(..., true)与AscendC::DataCopyPad(..., true);block_mmad_pertile.h中#if defined(__NPU_ARCH__) && (__NPU_ARCH__ == 9201 || __NPU_ARCH__ == 9202)的隔离写法。Describe the expected behavior / 预期结果 (Mandatory / 必填)
所有 GM2L1 预分配参数调用与 pertile 保持一致:9201/9202 传
true,其他架构使用无预分配参数的重载,保证跨平台编译与语义一致。Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
无(代码一致性/健壮性问题,非运行时故障)。
Special notes for this issue/备注 (Optional / 选填)
涉及 10 处调用:A/B、A/B Dn2Nz、WeightNz、Bias、ScaleA/ScaleB。
补充(同一 MR 一并修复):同文件中 HIF4/FP4 MTE1 补搬路径在 transA(或 !transB)且
ceil(curM/16) % B4_MIN_STEP != 0时,会把LoadData2DMxParams的 step 字段清零后仍走带 scale 的 LoadData 重载,9202 上报 MTE1m_step=0;现统一改为无 scale 的 LoadData 重载,scale 由首条指令按整个 curKL0 窗口搬运。