[Ansor Turning on llvm-bdver4]No latency for Wingrad Conv on AMD

[YueYue]

When I run the tune_newwork_x86.py, there are some question, is AMD (bdvert4) is not work?

I want to triger the winograd conv, so the ConvertLayout transform pass(NHWC Layout wanted) is added in my reley graph. And from the ir and task log, winograd is trigered. And I tried on Intel (target = tvm.target.Target(“llvm -mcpu=skylake-avx512”)), the latency shows normal, and faster than openvino.

But I tried on AMD(model name : AMD EPYC 7K62 48-Core Processor)(llvm -mcpu=bdver4"), the latency did not show, and the tuning time is longer…, so I changed the target as “target = tvm.target.Target(“llvm”)”, the conv latency shows normal.

So I want to know, if I use the “target = tvm.target.Target(“llvm”)”, the tuning results is correct ? And Why llvm-bdver4 is not work?

target = tvm.target.Target(“llvm -mcpu=bdver4”), I wait long time, but it still not work …

2.target = tvm.target.Target(“llvm”)

Which Target should I use ? bdver4 is not work, znver2 works normal

[merrymercy]

This problem is not related to Ansor. It seems your target string is wrong, but I don’t know the right target string for your CPU either. You can try some target strings in a simple task (e.g., compile a vector add or simple matmul) first.