Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
对于wo大于1024场景，tiling决策优先走hw分轴计算的hw模式，由于ho比较小没法进行多核切分，导致性能不行
Ascend950PR_9589
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.401 [conv2d_v2_base_tiling_fast_tiling.cpp:138][OPS_NN][GetTilingSplitMode][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling by check HWMode first with orgWo 1790, basicBlockM 1024, basicBlockN 64. [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.412 [conv2d_v2_base_tiling_fast_tiling.cpp:310][OPS_NN][Conv2dOpTilingSetShape][3323863] OpName:[conv_1] ExtendConv2D AscendC: api got: singleCoreCi: 64, singleCorekh: 3, singleCorekw: 3 [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.416 [conv2d_v2_base_tiling_fast_tiling.cpp:50][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling from fast_tiling. [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.447 [conv2d_v2_base_tiling_fast_tiling.cpp:72][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling from HWmode formulaic algorithm, belong to fast_tiling. [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.450 [conv2d_v2_base_tiling_fast_tiling.cpp:75][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: batchDim / hoDim / woDim / nDim / groupDim / minCost: 1, 2, 1, 4, 1, 10375. [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.457 [conv2d_v2_base_tiling_print.cpp:46][OPS_NN][PrintOpTilingData][3323863] OpName:[conv_1] ExtendConv2D AscendC: ops tilingdata: hin: 7, win: 1792, hout: 5, wout: 1790, batch: 1, cin: 64, cout: 64, kh: 3, kw: 3, batchDim: 1, hoDim: 2, woDim: 1, nDim: 4, strideH: 1, strideW: 1, dilationH: 1, dilationW: 1, padTop: 0, padLeft: 0, groups: 1, cinOpt: 0, coutOpt: 0, groupOpt: 0, enlarge: 0, groupDim: 1, hasBias: 1 [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.466 [conv2d_v2_base_tiling_fast_tiling.cpp:235][OPS_NN][Conv2dApiTilingSetShape][3323863] OpName:[conv_1] ExtendConv2D AscendC: api got: orgCo: 64, orgkH: 3, orgkW: 3, orgHi: 7, orgWi: 1792, singleCo: 16, singleHo: 3, singleWo: 1790, singleM: 0, singleGroups: 0, singleGroupOpt: 0, enlarge: 0 [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.547 [conv2d_v2_base_tiling_tilingkey.cpp:425][OPS_NN][SetTilingKey][3323863] OpName:[conv_1] ExtendConv2D AscendC: c04 mode status is: 0 [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.551 [conv2d_v2_base_tiling_tilingkey.cpp:428][OPS_NN][SetTilingKey][3323863] OpName:[conv_1] ExtendConv2D AscendC: tiling key: 2883776. fmpTiling[0], weightTiling[0], l1PingPong[0], l0PingPong[3], outputOrder[0], iterOrder[0], groupType[0], enableSmallChannel[0], weightUbTrans[0], fmapCppyMode[0], innerBatch[0], batchOne[1], noPad[1], smallWeight[2], smallKernel[0]. [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.564 [conv2d_v2_base_tiling_print.cpp:164][OPS_NN][PrintLibApiTilingData][3323863] OpName:[conv_1] ExtendConv2D AscendC: api tilingdata: singleCoreHo: 3, singleCoreWo: 1790, singleCoreBatch: 1, orgHi: 7, orgWi: 1792, orgHo: 5, orgWo: 1790, groups: 1, orgCi: 64, orgCo: 64, kernelH: 3, kernelW: 3, singleCoreCi: 64, singleCoreCo: 16, hoL1: 3, woL1: 1792, kAL1: 576, kBL1: 576, nBL1: 16, hoL0: 1, woL0: 1024, kL0: 32, nL0: 16, pBufferFlag: 7, multiNBL1: 1, strideH: 1, strideW: 1, dilationH: 1, dilationW: 1, padTop: 0, padBottom: 0, padLeft: 0, padRight: 0, aL1SpaceSize: 573440, singleCoreGroups: 0, singleCoreGroupOpt: 0, enlarge: 0, bUbNStep: 0, iterateMNOrder: 0, biasFullLoadFlag: 1, fixpParamsFullLoadFlag: 1, hf32Enable: 0, hf32TransMode: 0, hasBias: 1, hasScale: 1, offsetx: 15, roundMode: 1, innerBatch: 1 [DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.567 [conv2d_v2_base_tiling_print.cpp:165][OPS_NN][PrintLibApiTilingData][3323863] OpName:[conv_1] bUbKStep: 0, orgHixWi: 12544, kernelHxkernelW: 9, kernelHxkernelWxkernelD: 9, cinAInCore: 64, cinATailInCore: 64, cinBInCore: 64, cinBTailInCore: 64, mStep: 1024, kStep: 1, nStep: 1, fmapKStride: 64, weightKStride: 1, cinOffsetBlockInGM: 802816, coutOffsetBlock: 576, nL1DivBlockSize: 1, dualOutput: 0, quantMode0: 2, reluMode0: 1, clipMode0: 0, quantMode1: 0, reluMode1: 0, clipMode1: 0, khL1: 0, kwL1: 0, khUb: 0, kwUb: 0, unionDataXt: 271593537
走hw合轴的m模式，所有核都绑在m轴上
assign
Thanks for sending an issue! Please fill in the following template to help quickly solve your problem.
Describe the current behavior / 问题描述 (Mandatory / 必填)
对于wo大于1024场景，tiling决策优先走hw分轴计算的hw模式，由于ho比较小没法进行多核切分，导致性能不行
Environment / 环境信息 (Mandatory / 必填)
Ascend950PR_9589
Steps to reproduce the issue / 重现步骤 (Mandatory / 必填)
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.401 [conv2d_v2_base_tiling_fast_tiling.cpp:138][OPS_NN][GetTilingSplitMode][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling by check HWMode first with orgWo 1790, basicBlockM 1024, basicBlockN 64.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.412 [conv2d_v2_base_tiling_fast_tiling.cpp:310][OPS_NN][Conv2dOpTilingSetShape][3323863] OpName:[conv_1] ExtendConv2D AscendC: api got: singleCoreCi: 64, singleCorekh: 3, singleCorekw: 3
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.416 [conv2d_v2_base_tiling_fast_tiling.cpp:50][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling from fast_tiling.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.447 [conv2d_v2_base_tiling_fast_tiling.cpp:72][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling from HWmode formulaic algorithm, belong to fast_tiling.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.450 [conv2d_v2_base_tiling_fast_tiling.cpp:75][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: batchDim / hoDim / woDim / nDim / groupDim / minCost: 1, 2, 1, 4, 1, 10375.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.457 [conv2d_v2_base_tiling_print.cpp:46][OPS_NN][PrintOpTilingData][3323863] OpName:[conv_1] ExtendConv2D AscendC: ops tilingdata: hin: 7, win: 1792, hout: 5, wout: 1790, batch: 1, cin: 64, cout: 64, kh: 3, kw: 3, batchDim: 1, hoDim: 2, woDim: 1, nDim: 4, strideH: 1, strideW: 1, dilationH: 1, dilationW: 1, padTop: 0, padLeft: 0, groups: 1, cinOpt: 0, coutOpt: 0, groupOpt: 0, enlarge: 0, groupDim: 1, hasBias: 1
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.466 [conv2d_v2_base_tiling_fast_tiling.cpp:235][OPS_NN][Conv2dApiTilingSetShape][3323863] OpName:[conv_1] ExtendConv2D AscendC: api got: orgCo: 64, orgkH: 3, orgkW: 3, orgHi: 7, orgWi: 1792, singleCo: 16, singleHo: 3, singleWo: 1790, singleM: 0, singleGroups: 0, singleGroupOpt: 0, enlarge: 0
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.547 [conv2d_v2_base_tiling_tilingkey.cpp:425][OPS_NN][SetTilingKey][3323863] OpName:[conv_1] ExtendConv2D AscendC: c04 mode status is: 0
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.551 [conv2d_v2_base_tiling_tilingkey.cpp:428][OPS_NN][SetTilingKey][3323863] OpName:[conv_1] ExtendConv2D AscendC: tiling key: 2883776. fmpTiling[0], weightTiling[0], l1PingPong[0], l0PingPong[3], outputOrder[0], iterOrder[0], groupType[0], enableSmallChannel[0], weightUbTrans[0], fmapCppyMode[0], innerBatch[0], batchOne[1], noPad[1], smallWeight[2], smallKernel[0].
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.564 [conv2d_v2_base_tiling_print.cpp:164][OPS_NN][PrintLibApiTilingData][3323863] OpName:[conv_1] ExtendConv2D AscendC: api tilingdata: singleCoreHo: 3, singleCoreWo: 1790, singleCoreBatch: 1, orgHi: 7, orgWi: 1792, orgHo: 5, orgWo: 1790, groups: 1, orgCi: 64, orgCo: 64, kernelH: 3, kernelW: 3, singleCoreCi: 64, singleCoreCo: 16, hoL1: 3, woL1: 1792, kAL1: 576, kBL1: 576, nBL1: 16, hoL0: 1, woL0: 1024, kL0: 32, nL0: 16, pBufferFlag: 7, multiNBL1: 1, strideH: 1, strideW: 1, dilationH: 1, dilationW: 1, padTop: 0, padBottom: 0, padLeft: 0, padRight: 0, aL1SpaceSize: 573440, singleCoreGroups: 0, singleCoreGroupOpt: 0, enlarge: 0, bUbNStep: 0, iterateMNOrder: 0, biasFullLoadFlag: 1, fixpParamsFullLoadFlag: 1, hf32Enable: 0, hf32TransMode: 0, hasBias: 1, hasScale: 1, offsetx: 15, roundMode: 1, innerBatch: 1
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.567 [conv2d_v2_base_tiling_print.cpp:165][OPS_NN][PrintLibApiTilingData][3323863] OpName:[conv_1] bUbKStep: 0, orgHixWi: 12544, kernelHxkernelW: 9, kernelHxkernelWxkernelD: 9, cinAInCore: 64, cinATailInCore: 64, cinBInCore: 64, cinBTailInCore: 64, mStep: 1024, kStep: 1, nStep: 1, fmapKStride: 64, weightKStride: 1, cinOffsetBlockInGM: 802816, coutOffsetBlock: 576, nL1DivBlockSize: 1, dualOutput: 0, quantMode0: 2, reluMode0: 1, clipMode0: 0, quantMode1: 0, reluMode1: 0, clipMode1: 0, khL1: 0, kwL1: 0, khUb: 0, kwUb: 0, unionDataXt: 271593537
Describe the expected behavior / 预期结果 (Mandatory / 必填)
走hw合轴的m模式，所有核都绑在m轴上
Related log / screenshot / 日志 / 截图 (Mandatory / 必填)
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.401 [conv2d_v2_base_tiling_fast_tiling.cpp:138][OPS_NN][GetTilingSplitMode][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling by check HWMode first with orgWo 1790, basicBlockM 1024, basicBlockN 64.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.412 [conv2d_v2_base_tiling_fast_tiling.cpp:310][OPS_NN][Conv2dOpTilingSetShape][3323863] OpName:[conv_1] ExtendConv2D AscendC: api got: singleCoreCi: 64, singleCorekh: 3, singleCorekw: 3
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.416 [conv2d_v2_base_tiling_fast_tiling.cpp:50][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling from fast_tiling.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.447 [conv2d_v2_base_tiling_fast_tiling.cpp:72][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: get tiling from HWmode formulaic algorithm, belong to fast_tiling.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.450 [conv2d_v2_base_tiling_fast_tiling.cpp:75][OPS_NN][GetTilingFromFastTiling][3323863] OpName:[conv_1] ExtendConv2D AscendC: batchDim / hoDim / woDim / nDim / groupDim / minCost: 1, 2, 1, 4, 1, 10375.
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.457 [conv2d_v2_base_tiling_print.cpp:46][OPS_NN][PrintOpTilingData][3323863] OpName:[conv_1] ExtendConv2D AscendC: ops tilingdata: hin: 7, win: 1792, hout: 5, wout: 1790, batch: 1, cin: 64, cout: 64, kh: 3, kw: 3, batchDim: 1, hoDim: 2, woDim: 1, nDim: 4, strideH: 1, strideW: 1, dilationH: 1, dilationW: 1, padTop: 0, padLeft: 0, groups: 1, cinOpt: 0, coutOpt: 0, groupOpt: 0, enlarge: 0, groupDim: 1, hasBias: 1
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.466 [conv2d_v2_base_tiling_fast_tiling.cpp:235][OPS_NN][Conv2dApiTilingSetShape][3323863] OpName:[conv_1] ExtendConv2D AscendC: api got: orgCo: 64, orgkH: 3, orgkW: 3, orgHi: 7, orgWi: 1792, singleCo: 16, singleHo: 3, singleWo: 1790, singleM: 0, singleGroups: 0, singleGroupOpt: 0, enlarge: 0
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.547 [conv2d_v2_base_tiling_tilingkey.cpp:425][OPS_NN][SetTilingKey][3323863] OpName:[conv_1] ExtendConv2D AscendC: c04 mode status is: 0
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.551 [conv2d_v2_base_tiling_tilingkey.cpp:428][OPS_NN][SetTilingKey][3323863] OpName:[conv_1] ExtendConv2D AscendC: tiling key: 2883776. fmpTiling[0], weightTiling[0], l1PingPong[0], l0PingPong[3], outputOrder[0], iterOrder[0], groupType[0], enableSmallChannel[0], weightUbTrans[0], fmapCppyMode[0], innerBatch[0], batchOne[1], noPad[1], smallWeight[2], smallKernel[0].
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.564 [conv2d_v2_base_tiling_print.cpp:164][OPS_NN][PrintLibApiTilingData][3323863] OpName:[conv_1] ExtendConv2D AscendC: api tilingdata: singleCoreHo: 3, singleCoreWo: 1790, singleCoreBatch: 1, orgHi: 7, orgWi: 1792, orgHo: 5, orgWo: 1790, groups: 1, orgCi: 64, orgCo: 64, kernelH: 3, kernelW: 3, singleCoreCi: 64, singleCoreCo: 16, hoL1: 3, woL1: 1792, kAL1: 576, kBL1: 576, nBL1: 16, hoL0: 1, woL0: 1024, kL0: 32, nL0: 16, pBufferFlag: 7, multiNBL1: 1, strideH: 1, strideW: 1, dilationH: 1, dilationW: 1, padTop: 0, padBottom: 0, padLeft: 0, padRight: 0, aL1SpaceSize: 573440, singleCoreGroups: 0, singleCoreGroupOpt: 0, enlarge: 0, bUbNStep: 0, iterateMNOrder: 0, biasFullLoadFlag: 1, fixpParamsFullLoadFlag: 1, hf32Enable: 0, hf32TransMode: 0, hasBias: 1, hasScale: 1, offsetx: 15, roundMode: 1, innerBatch: 1
[DEBUG] OP(3323863,python3):2026-06-08-17:21:28.220.567 [conv2d_v2_base_tiling_print.cpp:165][OPS_NN][PrintLibApiTilingData][3323863] OpName:[conv_1] bUbKStep: 0, orgHixWi: 12544, kernelHxkernelW: 9, kernelHxkernelWxkernelD: 9, cinAInCore: 64, cinATailInCore: 64, cinBInCore: 64, cinBTailInCore: 64, mStep: 1024, kStep: 1, nStep: 1, fmapKStride: 64, weightKStride: 1, cinOffsetBlockInGM: 802816, coutOffsetBlock: 576, nL1DivBlockSize: 1, dualOutput: 0, quantMode0: 2, reluMode0: 1, clipMode0: 0, quantMode1: 0, reluMode1: 0, clipMode1: 0, khL1: 0, kwL1: 0, khUb: 0, kwUb: 0, unionDataXt: 271593537
Special notes for this issue/备注 (Optional / 选填)