Conv1D is slow

[eric-haibin-lin]

The Conv1D block uses MXNet's implementation for Convolution, because CUDNN only implements 2-D convolution. In the operator, we can try to optimize the performance by reshaping the inputs to 4D, and kernels to 2D, so that CUDNN kernel can be used.

[TaoLv]

Yes. By doing this, conv1d can also benefit from mkldnn conv2d on cpu side. @pengzhao-intel

[pengzhao-intel]

In our local case, it can be 10x faster :) We can provide the PR for CPU side @jinhuang415

[eric-haibin-lin]

That will be great!

[vandanavk]

@pengzhao-intel @jinhuang415 is there a PR for this issue?

[pengzhao-intel]

@vandanavk Not yet, the 1D conv is still under the development.
We will file PR and merge it into MXNet in the next MKL-DNN release (0.17).
oneapi-src/oneDNN@a1204e4

[pengzhao-intel]

@xinyu-intel please check the latest MKL-DNN and conv1d.
We could start the evaluate and integrate conv1d now.

[pengzhao-intel]

#13530 (comment)

The performance is listed in this PR @eric-haibin-lin