[TOPI] Add dilation argument to conv2d and depthwise_conv2d

[vinx13]

Currently dilate is a separate operation and one has to call dilate + conv2d to compose dilated convolution.
The problems:

• Sometimes it is better to combine both into a single operator because we may want to write the expresion of dilated convolution directly instread of constructing a dilated kernel and then call compute_inline.
• The workload in AutoTVM does not reflect dilation. For example, we will use the same configuration for 5x5 kernel and 3x3 kernel with dilation=2 because they have the same workload.

This PR adds dilation as a new argument to conv2d / depthwise_conv2d. If dilation is used, it will call nn.dilate internally in conv2d (like nn.pad). The workload is also updated. This will invalidate AutoTVM logs generated previously.

cc @merrymercy @tqchen @masahi @nishi-t @anijain2305 @Huyuwei

[merrymercy]

Currently, the log format of autotvm depends on arguments. So every time we change arguments, we will invalidate all old log records. Should we use another design? @tqchen

Since this PR will break old logs, you should provide a log converter for other users.
I can test you converter and update the logs in TopHub.

[tqchen]

I think in this case, adding dilation makes sense, because this makes things consistent with the high level op. We do need to provide a quick upgrade path though, please try make API consistent with high level op https://docs.tvm.ai/langref/relay_op.html#level-2-definitions

[vinx13]

@merrymercy Are there cases where dilation is used in tophub logs? To convert old logs to new workload, I will add a dilation field to all old logs. Since we don't have dilation previously, if dilation is used we will have dilated kernel shape in the workload.

[merrymercy]

No dilation is used in tophub.

[vinx13]

here is the converter for old logs
https://gist.github.com/vinx13/6c9231843a5833f31de7eab938b345b1

[tqchen]

@vinx13 please check the CI errors, then ask for review again

[vinx13]

the test of dilated conv2d nhwc is incorrect, see https://github.com/dmlc/tvm/blob/42dc24a310170577f929f648f477ca2567c8bc9a/topi/tests/python/test_topi_conv2d_nhwc.py#L16
https://github.com/dmlc/tvm/blob/42dc24a310170577f929f648f477ca2567c8bc9a/topi/python/topi/nn/conv2d.py#L277
the dilation should be [dilation_h, dilation_w, 1, 1]

@tqchen How can I update pickle-memorized reference data?

[merrymercy]

Currently x86 backend in under refactor #1993 .
I think it is better to merge that first. So you can stop all your changes on topi/x86/*.

To update a pickle-memorized data, a trick is to rename the string.
https://github.com/dmlc/tvm/blob/42dc24a310170577f929f648f477ca2567c8bc9a/topi/tests/python/test_topi_conv2d_nhwc.py#L23

[vinx13]

@merrymercy previous dilation is incorrect, I have updated to [dilation, dilation, 1, 1] but current fallback schedule generated invalid ptx on cuda

[merrymercy]

The filter is incorrect. The dilation is correct.

[vinx13]

nchw and nhwc have different filter layouts, the filter here is consistent with nn.depthwise_conv2d_nhwc.
https://github.com/dmlc/tvm/blob/b840e9602e357c50124d0c7fb131c52321062570/topi/python/topi/nn/depthwise_conv2d.py#L80

[merrymercy]

Sorry.
NHWC layout doesn't use autotvm template. I think we can just disable this test, since no one uses this layout.. or you can try to fix the manual schedule https://github.com/dmlc/tvm/blob/b840e9602e357c50124d0c7fb131c52321062570/topi/python/topi/cuda/depthwise_conv2d.py#L119

[vinx13]

@merrymercy I have looked into the error in http://ci.tvm.ai:8080/blue/organizations/jenkins/tvm/detail/PR-1970/21/pipeline/232#step-255-log-1230. There is one failing case of nchw conv2d on nvptx because tophub logs are invalidated in this pr. The fallback schedule from fallback_with_reference_log produced wrong conv2d result.

[merrymercy]

There are some errors with nvptx backend. We found some cases before.

The solution can be

1. Update llvm in CI to 6.0
2. Disable that test case (for nvptx)

[vinx13]

There are some errors with nvptx backend. We found some cases before.

The solution can be

1. Update llvm in CI to 6.0
2. Disable that test case (for nvptx)

I tried llvm-6.0 locally but the error still occurs. This error can be fixed by updating tophub logs with dilation arg, or disabling fallback schedule generated from reference log

[merrymercy]

OK. We will first merge #2034, and then merge this PR.

In the meanwhile, you can do update to tophub

1. Use your converter to convert all latest logs under https://github.com/uwsampl/tvm-distro/tree/master/tophub
   • arm_cpu_v0.03 -> arm_cpu_v0.04, cuda_v0.03 -> cuda_v0.04, opencl_v0.01 -> opencl_v0.02, ..., the only special case is vta_v0.01 -> vta_v0.04
   • Note there are also dense layer workloads. So your convert should only convert conv layers.
   • Then I will merge your PR
2. Update tophub versions for all backends in tvm
   • Update them to your version
     https://github.com/dmlc/tvm/blob/b840e9602e357c50124d0c7fb131c52321062570/python/tvm/autotvm/tophub.py#L22-L31
   • Then you should be able to pass your test

[merrymercy]

Can you also help to update the style of cuda winograd and cuda int8 according to #2034 ?
So we can delete all strange logic in things like _conv2d_arg_to_workloads

[tqchen]

Thanks @merrymercy @vinx13 this is now merged