tenstorrent / tt-metal

:metal: TT-NN operator library, and TT-Metalium low level kernel programming model.
Apache License 2.0
465 stars 73 forks source link

Unable to see width_shard in perf_sheet #12930

Open punithsekar opened 1 month ago

punithsekar commented 1 month ago

Describe the bug Unable to see width_sharded in perf while using bunch of conv's .

Note: In unit test Width_sharding conv we are able to see width_sharded in Perf sheet.

To Reproduce Steps to reproduce the behavior:

  1. checkout to branch punith/yolov_ws_perf
  2. Run command ./tt_metal/tools/profiler/profile_this.py -n yolov4_reshard_eliminate -c "pytest /home/ubuntu/punith/tt-metal/tests/ttnn/integration_tests/yolov4/test_ttnn_neck.py"

Please complete the following environment information:

Additional context I am attaching the perf sheet, width_shard_check_perf.csv

punithsekar commented 1 month ago

fyi @saichandax

dvartaniansTT commented 2 weeks ago

@punithsekar can you allaborate, which op is width shared and where you expect to see it in the sheet and you dont?

punithsekar commented 2 weeks ago

In the commit mentioned in the description, the neck.py file contains one convolution with width_sharding layout passed to convolution but in the attached perf sheet, we are getting Block_sharded layout instead of Width_sharded.

In the perf sheet for OptimizedConvNew row, WIDTH_SHARDED is the expected for the OUTPUT_0_MEMORY column.

Note: In the latest main I have not checked. This info is as of sept 20.