FP32 latency growth for ERNIE model
Created by: pmajchrzak
Using application:
cd /root/models/benchmark/Inference/c++/ernie
KMP_AFFINITY=granularity=fine,compact,1,0 KMP_BLOCKTIME=1 ./run.sh -1 1 /data/PaddlePaddle/trained_models/ERNIE/ernie_fp32_model /data/PaddlePaddle/datasets/ernie/Ernie_dataset/1.8w.bs1
which is get from the repository: https://github.com/PaddlePaddle/benchmark.git
for measuring latency it can be observed that latency grows significant (so performance drops) and permanently starting from the commit: f9066e6a
This observation is for fp32 only
There were more latency growths before but from the mentioned commit this not getting better