Add aten mkldnn batch_norm operator #19206
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,77 @@ | ||
| #include <ATen/ATen.h> | ||
| #include <ATen/Config.h> | ||
| #include <ATen/NativeFunctions.h> | ||
| #include <tuple> | ||
|
|
||
| #if !AT_MKLDNN_ENABLED() | ||
|
|
||
| namespace at { | ||
| namespace native { | ||
|
|
||
| std::tuple<Tensor, Tensor, Tensor> mkldnn_batch_norm( | ||
| const Tensor& self, | ||
| const Tensor& weight, | ||
| const Tensor& bias, | ||
| const Tensor& running_mean, | ||
| const Tensor& running_var, | ||
| bool train, | ||
| double momentum, | ||
| double eps) { | ||
| AT_ERROR("mkldnn_batch_norm: ATen not compiled with MKLDNN support"); | ||
| } | ||
|
|
||
| } // namespace native | ||
| } // namespace at | ||
|
|
||
| #else // AT_MKLDNN_EBABLED | ||
|
|
||
| #include <ATen/native/mkldnn/MKLDNNCommon.h> | ||
|
|
||
| namespace at { | ||
| namespace native { | ||
|
|
||
| std::tuple<Tensor, Tensor, Tensor> mkldnn_batch_norm( | ||
| const Tensor& input, | ||
| const Tensor& weight, | ||
| const Tensor& bias, | ||
| const Tensor& running_mean, | ||
| const Tensor& running_var, | ||
| bool train, | ||
| double momentum, | ||
| double eps) { | ||
| ideep::tensor& x = itensor_from_mkldnn(input); | ||
| ideep::tensor& w = itensor_from_mkldnn(weight); | ||
| ideep::tensor& b = itensor_from_mkldnn(bias); | ||
| ideep::tensor& m = itensor_from_mkldnn(running_mean); | ||
| ideep::tensor& v = itensor_from_mkldnn(running_var); | ||
|
|
||
| ideep::tensor y; | ||
|
|
||
| if (train) { | ||
|
Collaborator
There was a problem hiding this comment. This training support is incomplete. Derivatives.yaml calls native_batch_norm_backward and that one doesn't know what to do with mkldnn tensors. Thus I'd say either assert that train=false or implement native_batch_norm_backward for consistency
Contributor
Author
There was a problem hiding this comment. Makes senses will add an assert |
||
| // TODO: support training | ||
| AT_ERROR("mkldnn_batch_norm: mkldnn training is not supported in yet."); | ||
|
|
||
| // ideep::tensor saved_mean; | ||
| // ideep::tensor saved_var; | ||
| // ideep::batch_normalization_forward_training::compute<AllocForMKLDNN>( | ||
| // x, w, b, y, saved_mean, saved_var, m, v, momentum, eps); | ||
| // return std::make_tuple( | ||
| // new_with_itensor_mkldnn(std::move(y), input.options()), | ||
| // new_with_itensor_mkldnn(std::move(saved_mean), input.options()), | ||
| // new_with_itensor_mkldnn(std::move(saved_var), input.options())); | ||
| } else { | ||
| AT_ASSERTM(input.dim() == 4 || input.dim() == 5, | ||
| "mkldnn_batch_norm: currently mkldnn only support 2d and 3d batchnorm"); | ||
| ideep::batch_normalization_forward_inference::compute<AllocForMKLDNN>( | ||
| x, m, v, w, b, y, eps); | ||
| return std::make_tuple( | ||
| new_with_itensor_mkldnn(std::move(y), input.options()), | ||
| new_with_itensor_mkldnn(ideep::tensor{}, input.options()), | ||
| new_with_itensor_mkldnn(ideep::tensor{}, input.options())); | ||
| } | ||
| } | ||
|
|
||
| } // namespace native | ||
| } // namespace at | ||
|
|
||
| #endif // AT_MKLDNN_EBABLED | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Uh oh!
There was an error while loading. Please reload this page.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@bddppq , I think there should check whether running status is used for training and infernce, i.e. running_mean and running_var are defined or not. Another suggestion is that mkldnn only support batchnorm2d and batchnorm3d, so you shoud make some checks when call mkldnn, perhaps you can see the code as reference. thanks!
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
would mkldnn throw with nice message that it's not supported? in this case the input is already mkldnn tensor so it's better to fail if something is not supported
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@dzhulgakov I'm afraid not. :-( How about we add an assertion here to guarantee 2d or 3d batchnorm here?