Potential performance regression on RK3399 Mali GPU

Hi all,

I tried to benchmark ResNet50 on rk3399 with newest tvm but could not obtain the results shown here https://github.com/apache/incubator-tvm/wiki/Benchmark. As I further look into the issue, I found that rk3399 performance was consistent with the wiki page’s benchmark until commit 4273e461. The compilation of the model then would fail after this commit until a fix from commit f5f2fee, which unfortunately showed a performance regression. Can anyone encounter the same problem or explain this issue?

Thank you very much.

@tqchen It looks like this is related to the new simplifier? @yma10 Can you run a benchmark for a single conv2d workload and print the low ir from before/after this commit?