Performance Signal Detected by TorchBench CI on '1.13.0.dev20220718+cu113'

[github-actions]

TorchBench CI has detected a performance signal.

Base PyTorch version: 1.13.0.dev20220717+cu113

Base PyTorch commit: 0b5b10002af2dce031d4fad337728f81c58a8b10

Affected PyTorch version: 1.13.0.dev20220718+cu113

Affected PyTorch commit: 471397d0ee27b61ae1a088ece9f3ead1562152e0

Affected Tests:

• test_eval[timm_efficientnet-cpu-jit]: +457.65788%
• test_eval[timm_efficientnet-cpu-eager]: +387.80496%

cc @xuzhao9

[xuzhao9]

Bisection workflow: https://github.com/pytorch/benchmark/actions/runs/2693627667

Result:

{
  "start": "0b5b10002af2dce031d4fad337728f81c58a8b10",
  "end": "471397d0ee27b61ae1a088ece9f3ead1562152e0",
  "threshold": 7,
  "timeout": 120,
  "torchbench_branch": "v1.0",
  "result": [
    {
      "commit1": "95c148e502b",
      "commit1_time": "2022-07-17 05:49:39 +0000",
      "commit1_digest": {
        "test_eval[timm_efficientnet-cpu-jit]": 0.032386388344457374,
        "test_eval[timm_efficientnet-cpu-eager]": 0.037657551842534706
      },
      "commit2": "4655c3bace4",
      "commit2_time": "2022-07-17 14:58:26 +0000",
      "commit2_digest": {
        "test_eval[timm_efficientnet-cpu-jit]": 0.17101764229009858,
        "test_eval[timm_efficientnet-cpu-eager]": 0.18220487465150653
      }
    }
  ]
}

Root cause commit: pytorch/pytorch@4655c3bace4

@ngimel @malfet @yanbing-j Can you help take a look?

[ngimel]

@yanbing-j can conditions to disable oneDNN be relaxed to not regress existing perf?

[yanbing-j]

Hi @ngimel @xuzhao9 @malfet , current solution needs more check before running into oneDNN Conv, which causes the perf downgrade. And we are trying to mitigate the regression in other approaches.
You may revert this PR first, I will create a new PR to avoid the benchmark regression. Thank you!

[xuzhao9]

Hi @ngimel @xuzhao9 @malfet , current solution needs more check before running into oneDNN Conv, which causes the perf downgrade. And we are trying to mitigate the regression in other approaches. You may revert this PR first, I will create a new PR to avoid the benchmark regression. Thank you!

Reverted. Can you please run the TorchBench CI on the new PR to verify that it doesn't introduce regression (similar to pytorch/pytorch#65839)? Thanks!

[xuzhao9]

Closed as we have reverted the diff.