I have two questions:
Question 1:
Figure 12 , in “TVM: An Automated End-to-End Optimizing Compiler for Deep Learning” paper, shows that autotvm can get the better performance than cudnn.
So we wanted to verify it and did the benchmark . but the result show that cudnn had the better performance than autotvm and we set the n_trails = 2000.
This makes me confused.
Can you give me the test code about this. my email is 734843835@qq.com
If you can’t give me the test code, please tell me how did you do the benchmark.
Question 2:
I got PPT “AutoTVM & Device Fleet” from https://sampl.cs.washington.edu/tvmconf/slides/Eddie-Yan-AutoTVMFleet.pdf
Page 33 show the benchmark result about NVIDIA GPU Optimization(GTX 1080Ti)
Can you give me the test code about this benchmark. my email is 734843835@qq.com
If you can’t give me the test code, please tell me how did you do the benchmark.
For AutoTVM and “MXNet + TensorRT 4.0”, Is the same DL model ?
Or “MXNet+ TensorRT 4.0” use the mxnet real model but AutoTVM use the relay model from relay.testing?
Thank you very much.