Created by: bingyanghuang
#15032 (closed) reported that recently, DAM model CI test frequently hang with the test compare_mkldnn. This PR is for disabling the conv3d mkldnn kernel when test compare_mkldnn and profile_mkldnn. This test is also aim to help PR 15032 narrow down the cause.
| Platform: | Intel(R) Xeon(R) CPU E5-2650 v4 @ 2.20GHz |
|---|---|
| Core: | 24 |
| Turbo boost: | Enable |
| Env Config: | OMP_NUM_THREADS=1 |
| Command line: | ./paddle/fluid/inference/tests/api/test_analyzer_small_dam --infer_model=third_party/inference_demo/small_dam/model/ --infer_data=third_party/inference_demo/small_dam/data.txt --gtest_filter=Analyzer_dam.profile_mkldnn --paddle_num_threads=1 --repeat=10 --batch_size=300 --max_turn_num=1 --test_all_data |
| repeat: | 10 |
| commit id: | 3f815e07 |
| Average latency of each sample | MKLML | MKLDNN conv3d | MKLDNN disable conv3d |
|---|---|---|---|
| BS=300 (ms) | 7.75 | 6.94 | 8.15 |
MKLML: Run with --gtest_filter=Analyzer_dam.profile
MKLDNN conv3d : Run with original small dam with --gtest_filter=Analyzer_dam.profile_mkldnn , it will only enable conv3d mkldnn kernel
MKLDNN disable conv3d: Run with modified small dam with --gtest_filter=Analyzer_dam.profile_mkldnn， it will enable all mkldnn supported op except the conv3d.