[BYOC][DNNL] Enable layer normalization in DNNL byoc.

[billishyahao]

This patch is to enable layer normalization in DNNL BYOC by providing an out-of-box rewrite pattern for combining the operators into a single relay layer normalization operator as well as its implementation in dnnl json codegen.

After applying the rewrite pattern, we will observe the following dnnl function:

6202 def @tvmgen_default_dnnl_main_108(%dnnl_108_i0: Tensor[(1, 784, 128), float32], Inline=1, Compiler="dnnl", global_symbol="tvmgen_default_dnnl_main_108", Primitive=1) -> Tensor[(1, 784, 128), float32] {
6203   nn.layer_norm(%dnnl_108_i0, meta[relay.Constant][56] /* ty=Tensor[(128), float32] */, meta[relay.Constant][57] /* ty=Tensor[(128), float32] */) /* ty=Tensor[(1,  784, 128), float32] */
6204 }

Once you enable DNNL_VERBOSE flag, more informations are shown in log file as below:

onednn_verbose,exec,cpu,layer_normalization,simple_layer_normalization:any,forward_inference,data_f32::blocked:abc:f0 stats_undef::undef::f0 diff_undef::undef::f0,,flags:CH,1x784x128,0.0551758

With this patch, I benchmarked the inference performance of a kind of vision-tranformer called PCPVT (https://arxiv.org/abs/2104.13840) on ICX-8352Y. It gains up to 1.18X boost. Here is some boost data:

32 cores	Latency
baseline byoc	11.45ms
byoc w/ patch	9.68ms

Thanks for contributing to TVM! Please refer to guideline https://tvm.apache.org/docs/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.

[crazydemo]

Thanks for your contribution for BYOC-DNNL. And my suggestions are listed below:

1. I wonder if we can get better performance via running layernorm on dnnl codegen than running consecutive ops on native codegen. Could you please provide some performance numbers?
2. Lint has failed. Please run task_lint.sh to check the code style.
3. UT is required. You can add your test cases in tests/python/contrib/test_dnnl.py to ensure the functionality of the enabled ops.

[apeskov]

original "nn.layer_norm" has not only epsilon argument. At least axis, center and scale. By this code you assume that they always equal axis = -1, center=true and scale=true.

Could you please add support of all attributes or verify their values on codegen stage.

[billishyahao]

I add ICHECK to this case. And will update later for supporting all other attributes.

[apeskov]

@crazydemo Answering your question about performance.

I wonder if we can get better performance via running layernorm on dnnl codegen than running consecutive ops on native codegen. Could you please provide some performance numbers?

Yes, there is performance benefit. At least they use different memory access approach. Consecutive ops with llvm codegen will produce sequence of fused kernel like next:

• mean. One pass through src tensor
• sub. One pass through src and dst tensor
• power + mean. One pass through src
• add + sqrt + div + mul + add. One pass through src and dst.

Totally we have 6 times traversing through data tensor for TVM codegen. DNNL implement it as single kernel and do only 4 passes through memory buffers (or 3 in case of in place memory).

In case of multi core system(xeon servers and other) normalise op is memory bound. And reduction of memory access becomes more important.

[billishyahao]

Hi @apeskov , please take a look at the latest version here. Feel free to comment more.

[billishyahao]

Hi @masahi Please take a look.

[billishyahao]

Hi @comaniac , @trevor-m , @mbaret , Could you take a look at this pr? Thanks!

[masahi]

You need to resolve the conflict in test_dnnl.py, but since it will be modified in #11513, let's merge #11513 first.

[billishyahao]

You need to resolve the conflict in test_dnnl.py, but since it will be modified in #11513, let's merge #11513 first.

Hi @masahi , Thanks for pointing out this. I have resolved the conflict in test_dnnl.py.