Mxnet group conv not support very well by tvm

hi all , I have a problem, I use tvm to speed up resnet101(mxnet) on 1080ti, I try to use group conv, but after I set group =2, the speed became slower too much. Then I try to speed up resnext101(mxnet). ResNext also became more slower . So , do tvm not support group conv (mxnet ) ?

To get more information , I try to set group = 32, after tvm speed up , it is more faster then group = 2.

if group = 1 , after tvm , process one image cost 6ms, if group = 2 , cost about 200ms, group = 3 cost 17ms. When group =2 , the speed is the slowest.

Can anyone solve the problem?

sorry, group=32, cost 17ms not group=3

maybe group=1 fallback to a dense convolution. group=32 become a depthwise convolution. tvm has tuned optimal schedules for dense conv and depthwise conv

have you tried using AutoTVM to tune these ops? Which platform are you using?

I have tuned , but it wasn’t useful. I use nvidia 1080ti, cuda.

Could you provide your workload (input/kernel shapes, stride, padding)? I’ll take a look