[inductor][cpu]some models fp32 static shape default wrapper single thread performance regression #117740
Description
🐛 Describe the bug
Performance regression in 2023-09-03:
| name | batch_size_new | speed_up_new | inductor_new | eager_new | compilation_latency_new | batch_size_old | speed_up_old | inductor_old | eager_old | compilation_latency_old | Ratio Speedup(New/old) | Eager Ratio(old/new) | Inductor Ratio(old/new) | Compilation_latency_Ratio(old/new) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| basic_gnn_gcn | 1 | 1.453264 | 0.07012852100000001 | 0.10191525494254401 | 16.308482 | 1 | 4.18351 | 0.024277587 | 0.10156552799037 | 12.917972 | 0.35 | 1.0 | 0.35 | 0.79 |
| basic_gnn_gin | 1 | 1.102929 | 0.053309631 | 0.058796738009199004 | 16.217177 | 1 | 3.911906 | 0.016065617 | 0.06284718353600201 | 12.840029 | 0.28 | 1.07 | 0.3 | 0.79 |
| basic_gnn_sage | 1 | 1.069263 | 0.055456228 | 0.059297292719964007 | 14.707836 | 1 | 3.680068 | 0.016367606 | 0.060233903077207995 | 11.36611 | 0.29 | 1.02 | 0.3 | 0.77 |
SW info
| SW | Nightly commit | Main commit |
|---|---|---|
| Pytorch | ea7e229 | |
| Torchbench | / | 770d5cf7 |
| torchaudio | dc83b38 | |
| torchtext | c11d758 | |
| torchvision | 58366ab | |
| torchdata | 1d231d1 | |
| dynamo_benchmarks | f228c8b | / |
Repro:
bash inductor_single_test.sh single inference performance torchbench model float32 first static default 0
Suspected guilty commit: cbf7c91
torchbench-basic_gnn_sage-inference-float32-static-default-performance-single-drop_guilty_commit.log
cc @ezyang @msaroufim @bdhirsh @anijain2305 @zou3519 @WeizhuoZhang-intel @chuanqi129