Closed
[Bug-Report|缺陷反馈]: ascend960dt/ascend960pr平台算子二进制配置缺失,ConcatD融合pass未覆盖DAV_9201 #3681
yangjinwenAdd on 21 days agoClosed at 21 days ago
21 days ago Added label: bug-report
21 days ago Associated pull request
21 days ago Modify description of issue
Yyangjinwen
21 days ago Change title to "[Bug-Report|缺陷反馈]: ascend960dt/ascend960pr平台算子二进制配置缺失,ConcatD融合pass未覆盖DAV_9201", originally titled "[Bug-Report|缺陷反馈]: ascend960dt平台Concat系列等算子二进制配置缺失,ConcatD融合pass未覆盖DAV_9201".
21 days ago Change title to "[Bug-Report|缺陷反馈]: ascend960dt/ascend960pr平台算子二进制配置缺失,ConcatD融合pass未覆盖DAV_9201", originally titled "[Bug-Report|缺陷反馈]: ascend960dt平台Concat系列等算子二进制配置缺失,ConcatD融合pass未覆盖DAV_9201".
21 days ago Modify description of issue
21 days ago Assign yangjinwen as responsible person
21 days ago Close issue
21 days ago Added label: resolved
一、问题描述 (必填)
ascend960dt/ascend960pr 平台上存在以下问题:
conversion/concat_d/op_graph/fusion_pass/concat_to_concatd_fusion_pass.cpp中IsTargetPlatform()的 RegBase 架构白名单仅包含DAV_3510/DAV_5102,未包含 ascend960dt 对应的DAV_9201,导致该融合 pass 在 ascend960dt 平台整体跳过(日志输出 "Platform NpuArch xx is not regbase(DAV_3510/DAV_5102), skip whole graph.")。scripts/kernel/binary_config/ascendc_config.json中 Concat、ConcatD、ConcatV2、Pack、BatchToSpaceND、MatrixSetDiag、SpaceToDepth、DepthToSpace、Rint 共 9 个算子缺失 ascend960dt 计算单元(compute_units)及对应编译选项配置,相关算子无法在 ascend960dt 平台编译运行。conversion/matrix_set_diag工程的 CMakeLists(COMPUTE_UNIT/TILING_DIR)、matrix_set_diag_def.cpp(AICore 配置注册)及op_host/config/下两平台的matrix_set_diag_binary.json(int8/int16/int32/int64)均未覆盖。二、环境信息 (可选)
• 昇腾硬件型号:ascend960dt / ascend960pr
• 仓库:cann/ops-math
三、重现步骤 (可选)
在 ascend960dt/ascend960pr 平台加载包含 Concat/ConcatD 节点的计算图,观察融合 pass 日志跳过整个图;同时上述算子在 ascend960dt 上无法编译生成二进制,MatrixSetDiag 无法调用。
四、预期结果 (可选)
关联 PR:https://gitcode.com/cann/ops-math/merge_requests/6523