Optimize x86 conv3d_ndhwc using data packing approach.

[alexgl-github]

Add tuneable conv3d_ndhwc schedule

Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers by @ them in the pull request thread.

[alexgl-github]

@anijain2305 Please take a look

[kevinthesun]

Thank you for this work! It would be great if you can provide benchmarking data comparing tvm conv3d performance VS existing solution(tensorflow + mkldnn?) to see where current implementation stands.

[anijain2305]

Minor comments.

Overall looks good to me. Currently, you have used NCHW schedule of Conv2D. you should also try Conv2d NCHWc schedule. That schedule gives best performance for Conv2D, and has potential here as well.

https://github.com/apache/incubator-tvm/blob/master/topi/python/topi/nn/conv2d.py#L421

[anijain2305]

Good to go from my side! If you can, please add some TF comparison.

[alexgl-github]

Thank you for this work! It would be great if you can provide benchmarking data comparing tvm conv3d performance VS existing solution(tensorflow + mkldnn?) to see where current implementation stands.

@kevinthesun
Below is benchmark results for certain data/kernel combinations, run on 2 core Intel "ivybridge" vs TF 1.15 + mkldnn. X: value means speedup (or slowdown) of TVM model vs same TF model

TVM: 0.007 sec; X: 1.645; TF: 0.011 sec; input_shape=(1, 16, 256, 256, 1) ; kernel_shape=(1, 3, 3, 1, 8)
TVM: 0.019 sec; X: 1.218; TF: 0.023 sec; input_shape=(1, 16, 256, 256, 1) ; kernel_shape=(1, 7, 7, 1, 8)
TVM: 0.054 sec; X: 1.490; TF: 0.080 sec; input_shape=(1, 16, 256, 256, 8) ; kernel_shape=(1, 3, 3, 8, 16)
TVM: 0.262 sec; X: 0.869; TF: 0.228 sec; input_shape=(1, 16, 256, 256, 8) ; kernel_shape=(1, 7, 7, 8, 16)
TVM: 0.013 sec; X: 1.290; TF: 0.016 sec; input_shape=(1, 16, 256, 256, 1) ; kernel_shape=(3, 3, 3, 1, 8)
TVM: 0.114 sec; X: 1.148; TF: 0.131 sec; input_shape=(1, 16, 256, 256, 1) ; kernel_shape=(7, 7, 7, 1, 8)
TVM: 0.146 sec; X: 1.058; TF: 0.154 sec; input_shape=(1, 16, 256, 256, 8) ; kernel_shape=(3, 3, 3, 8, 16)
TVM: 2.432 sec; X: 0.591; TF: 1.436 sec; input_shape=(1, 16, 256, 256, 8) ; kernel_shape=(7, 7, 7, 8, 16)

[kevinthesun]

LGTM. As suggestion from @anijain2305, we might also want to try similar data layout as conv2d_NCHWc to see whether we can get more performance improvement.

[kevinthesun]

Thanks @alexgl-github @anijain2305