ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-07-01 04:07:56 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add bwd wight configs.

Browse Source
This commit is contained in:
Ville Pietilä
2026-02-03 08:54:05 -05:00
parent 8e8e8b0216
commit b131d59347
6 changed files with 1037 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

233
experimental/grouped_convolution_tile_instances/configs/bwd_weight/profiler/ndhwgc_bf16.conf Normal file
View File

@@ -0,0 +1,233 @@
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Default, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Default, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Default, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Default, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Default, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Default, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Default, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Default, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Default, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Filter1x1Stride1Pad0, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 4, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 4, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 2, 8, 2, 8, 1, 1, 2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 256, 32, Default, 8, 1, 16, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 128, 32, Default, 8, 1, 8, 4, 8, 4, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>

233
experimental/grouped_convolution_tile_instances/configs/bwd_weight/profiler/ndhwgc_fp16.conf Normal file
View File

@@ -0,0 +1,233 @@
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Default, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Default, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Default, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Default, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Default, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Default, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Default, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Default, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Default, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Filter1x1Stride1Pad0, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 4, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 4, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 2, 8, 2, 8, 1, 1, 2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 256, 32, Default, 8, 1, 16, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 128, 32, Default, 8, 1, 8, 4, 8, 4, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>

44
experimental/grouped_convolution_tile_instances/configs/bwd_weight/profiler/ndhwgc_fp32.conf Normal file
View File

@@ -0,0 +1,44 @@
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 4, 2, 2, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 4, 2, 1, 4, 4, 1, 1, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 4, 1, 2, 1, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 4, 4, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Default, 4, 2, 4, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Default, 4, 4, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Default, 4, 2, 2, 4, 2, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Default, 4, 2, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Default, 4, 2, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 4, 2, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Default, 4, 2, 1, 4, 2, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Default, 4, 1, 2, 4, 1, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 4, 2, 1, 4, 4, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Default, 4, 1, 2, 4, 1, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Default, 4, 2, 1, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 4, 1, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 4, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 1, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 4, 2, 2, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 4, 1, 1, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 4, 1, 2, 1, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Filter1x1Stride1Pad0, 4, 4, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Filter1x1Stride1Pad0, 4, 2, 4, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Filter1x1Stride1Pad0, 4, 4, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 2, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 2, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Filter1x1Stride1Pad0, 4, 1, 2, 4, 1, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 4, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Filter1x1Stride1Pad0, 4, 1, 2, 4, 1, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 4, 1, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 4, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>

240
experimental/grouped_convolution_tile_instances/configs/bwd_weight/profiler/nhwgc_bf16.conf Normal file
View File

@@ -0,0 +1,240 @@
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Default, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Default, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Default, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Default, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Default, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Default, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Default, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Default, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Default, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Filter1x1Stride1Pad0, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<128, 16, 32, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<128, 16, 64, 64, Default, 8, 1, 2, 8, 8, 8, 8, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 256, 32, 64, Default, 8, 2, 1, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 128, 64, Default, 8, 1, 2, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 64, 64, Default, 8, 2, 2, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 64, 64, Default, 8, 2, 2, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 64, 64, Default, 8, 2, 2, 2, 1, 2, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 4, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 4, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 2, 8, 2, 8, 1, 1, 2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 256, 32, Default, 8, 1, 16, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 128, 32, Default, 8, 1, 8, 4, 8, 4, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>

243
experimental/grouped_convolution_tile_instances/configs/bwd_weight/profiler/nhwgc_fp16.conf Normal file
View File

@@ -0,0 +1,243 @@
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Default, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Default, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Default, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Default, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Default, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Default, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Default, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Default, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Default, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 2, 4, 2, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 2, 1, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 2, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Filter1x1Stride1Pad0, 8, 2, 4, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Filter1x1Stride1Pad0, 8, 4, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 8, 2, 2, 8, 4, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 2, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 1, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 1, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Filter1x1Stride1Pad0, 8, 2, 1, 8, 4, 8, 2, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 8, 1, 2, 8, 2, 8, 4, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<128, 16, 32, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<128, 16, 32, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<128, 16, 64, 64, Default, 8, 1, 2, 8, 8, 8, 8, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<128, 16, 64, 64, Default, 8, 1, 2, 8, 8, 8, 8, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 256, 32, 64, Default, 8, 2, 1, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 128, 64, Default, 8, 1, 2, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 128, 64, Default, 8, 1, 2, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 64, 64, Default, 8, 2, 2, 8, 8, 8, 8, 1, 1, 8>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 64, 64, Default, 8, 2, 2, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3_DirectLoad<256, 64, 64, 64, Default, 8, 2, 2, 2, 1, 2, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 32, 32, Filter1x1Stride1Pad0, 8, 1, 1, 2, 2, 2, 2, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 64, 32, Filter1x1Stride1Pad0, 8, 1, 2, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 32, 128, 32, Filter1x1Stride1Pad0, 8, 1, 4, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 32, 32, Filter1x1Stride1Pad0, 8, 2, 1, 4, 4, 4, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 128, 32, 32, Filter1x1Stride1Pad0, 8, 4, 1, 8, 8, 8, 8, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 80, 32, Filter1x1Stride1Pad0, 8, 4, 5, 4, 4, 5, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 64, 112, 32, Filter1x1Stride1Pad0, 8, 4, 7, 4, 4, 7, 4, 1, 1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 8, 8, 8, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 4, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 4, 8, 1, 1, 4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 2, 8, 2, 8, 1, 1, 2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<256, 64, 64, 64, Default, 8, 1, 1, 1, 8, 1, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 256, 32, Default, 8, 1, 16, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 128, 32, Default, 8, 1, 8, 4, 8, 4, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 64, 32, Default, 8, 1, 2, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 128, 32, Default, 8, 1, 4, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 32, 32, 32, Default, 8, 1, 1, 2, 2, 2, 2, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 2>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 32, 32, Default, 8, 2, 1, 4, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 4>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 128, 32, 32, Default, 8, 4, 1, 8, 8, 8, 8, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 8>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 48, 64, 32, Default, 8, 3, 4, 3, 4, 4, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 48, 32, Default, 8, 4, 3, 4, 4, 3, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 80, 32, Default, 8, 4, 5, 4, 4, 5, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 112, 32, Default, 8, 4, 7, 4, 4, 7, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle<64, 64, 208, 32, Default, 8, 4, 13, 4, 4, 13, 4, 1, 1, 1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v4, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x256x32, WaveTile: 32x32, WaveMap: 4x4, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v3, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v5, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 128x128x64, WaveTile: 32x32, WaveMap: 2x2, VmemReadVec: 8x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<Default, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x4, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x8, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x2, BlkGemmPipelineScheduler: Interwave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 8x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 4x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 2x1, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v1, BlkGemmPipelinePrefetchStages: 1>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 256x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 128x16x64, WaveTile: 16x16, WaveMap: 4x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 64x16x64, WaveTile: 16x16, WaveMap: 2x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 32x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 64, BlkTile: 16x16x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x32x64, WaveTile: 16x16, WaveMap: 1x1, VmemReadVec: 1x4, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 3>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x64x64, WaveTile: 16x16, WaveMap: 1x2, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 128, BlkTile: 16x128x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x8, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>
DeviceGroupedConvBwdWeight_Explicit_Xdl<DeviceBatchedGemmXdlUniversal<MNKPadding, CRR> BlkSize: 256, BlkTile: 16x256x64, WaveTile: 16x16, WaveMap: 1x4, VmemReadVec: 1x2, BlkGemmPipelineScheduler: Intrawave, BlkGemmPipelineVersion: v2, BlkGemmPipelinePrefetchStages: 2>

44
experimental/grouped_convolution_tile_instances/configs/bwd_weight/profiler/nhwgc_fp32.conf Normal file
View File

@@ -0,0 +1,44 @@
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 4, 2, 2, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 4, 2, 1, 4, 4, 1, 1, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 4, 1, 2, 1, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Default, 4, 4, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Default, 4, 2, 4, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Default, 4, 4, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Default, 4, 2, 2, 4, 2, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Default, 4, 2, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Default, 4, 2, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Default, 4, 2, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Default, 4, 2, 1, 4, 2, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Default, 4, 1, 2, 4, 1, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Default, 4, 2, 1, 4, 4, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Default, 4, 1, 2, 4, 1, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Default, 4, 2, 1, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Default, 4, 1, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 4, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 1, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 4, 2, 2, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 4, 1, 1, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 4, 1, 2, 1, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 256, 128, 4, Filter1x1Stride1Pad0, 4, 4, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 256, 4, Filter1x1Stride1Pad0, 4, 2, 4, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 128, 4, Filter1x1Stride1Pad0, 4, 4, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 128, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 2, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 64, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 64, 128, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 64, 4, Filter1x1Stride1Pad0, 4, 2, 2, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 128, 64, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 2, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 128, 4, Filter1x1Stride1Pad0, 4, 1, 2, 4, 1, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 128, 32, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 4, 4, 1, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<128, 32, 128, 4, Filter1x1Stride1Pad0, 4, 1, 2, 4, 1, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 64, 32, 4, Filter1x1Stride1Pad0, 4, 2, 1, 4, 4, 4, 2, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<64, 32, 64, 4, Filter1x1Stride1Pad0, 4, 1, 2, 4, 2, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 4, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 4, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 4, 4, 1, 1, 4>
DeviceGroupedConvBwdWeight_Xdl_CShuffle<256, 64, 64, 8, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 1, 4, 1, 1, 1>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Default, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>
DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<64, 16, 16, 32, Filter1x1Stride1Pad0, 8, 1, 1, 1, 4, 1, 4, 1, 1, 2>