[inductor][cpu]rexnet100 dynamic FP32 multiple threads performance regression #115346
Description
🐛 Describe the bug
dynamic_FP32_multiple_perf_regression in 2023-12-03
| 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) |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rexnet_100 | 128 | 2.637637 | 0.127494719 | 0.336284788139003 | 63.194334 | 128 | 3.043876 | 0.110155835 | 0.33530070241646 | 61.086464 | 0.87 | 1 | 0.86 | 0.97 |
SW info
| SW | Nightly commit | Main commit |
|---|---|---|
| Pytorch | 6fc9d89 | |
| Torchbench | / | 99944a2f |
| torchaudio | db62484 | |
| torchtext | b0ebddc | |
| torchvision | c1e2095 | |
| torchdata | 11bb5b8 | |
| dynamo_benchmarks | nightly | / |
Reference SW info (nightly)
Repro
inductor_single_run.sh
bash inductor_single_run.sh multiple/single inference performance timm_models rexnet100 float32 first dynamic default 0
Suspected guilty commit: afbaa0c
timm_models-rexnet_100-inference-float32-dynamic-default-performance-multiple-drop_guilty_commit.log
cc @ezyang @msaroufim @wconstab @bdhirsh @anijain2305 @zou3519 @WeizhuoZhang-intel @chuanqi129
| item | commit |
|---|---|
| torchbench | 94126be6 |
| torch | 2.2.0a0+gitdf0c1b3 |
| torchvision | 0.17.0a0+893b4ab |
| torchtext | 0.16.0a0+b0ebddc |
| torchaudio | 2.1.1+db62484 |
| torchdata | 0.7.0a0+11bb5b8 |
| dynamo_benchmarks | nightly |