Skip to content

use IndexList to improve performance of instance_norm op #25132

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
zhangting2020 merged 3 commits into from
Oct 12, 2020
Merged

use IndexList to improve performance of instance_norm op #25132

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
665329b
use IndexList to improve performance, test=develop
Jun 19, 2020
1636a11
remove EIGEN_HAS_INDEX_LIST, test=develop
Jun 28, 2020
572f972
use IndexList only when EIGEN_HAS_INDEX_LIST is true
Sep 13, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
33 changes: 27 additions & 6 deletions paddle/fluid/operators/instance_norm_op.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -181,10 +181,22 @@ class InstanceNormKernel<platform::CPUDeviceContext, T>
auto &dev_ctx = ctx.template device_context<platform::CPUDeviceContext>();
auto *place = dev_ctx.eigen_device();

Eigen::DSizes<int, 2> shape(NxC, sample_size);
// Once eigen on Windows is updated, the if branch can be removed.
#ifndef EIGEN_HAS_INDEX_LIST
Eigen::DSizes<int, 2> bcast(1, sample_size);
Eigen::DSizes<int, 2> C_shape(C, 1);
Eigen::DSizes<int, 2> NxC_shape(NxC, 1);
Eigen::DSizes<int, 2> shape(NxC, sample_size);
Eigen::DSizes<int, 1> rdims(1);
#else
Copy link
Contributor

@luotao1 luotao1 Jun 19, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

可以直接替换成IndexList么？

Copy link
Contributor Author

@zhangting2020 zhangting2020 Oct 12, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

目前由于windows上eigen版本未升级，因此直接替换为IndexList会导致编译错误。待windows也升级了eigen后，可以直接使用IndexList

Eigen::IndexList<Eigen::type2index<1>, int> bcast;
bcast.set(1, sample_size);
Eigen::IndexList<int, Eigen::type2index<1>> C_shape;
C_shape.set(0, C);
Eigen::IndexList<int, Eigen::type2index<1>> NxC_shape;
NxC_shape.set(0, NxC);
Eigen::IndexList<Eigen::type2index<1>> rdims;
#endif

math::SetConstant<platform::CPUDeviceContext, T> set_constant;

Expand All @@ -201,8 +213,6 @@ class InstanceNormKernel<platform::CPUDeviceContext, T>
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto x_arr = x_e.reshape(shape);

Eigen::DSizes<int, 1> rdims(1);

saved_mean_e.device(*place) = x_arr.mean(rdims);
auto saved_variance_arr =
(x_arr - saved_mean_e.broadcast(bcast)).square().mean(rdims) + epsilon;
Expand Down Expand Up @@ -316,14 +326,25 @@ class InstanceNormGradKernel<platform::CPUDeviceContext, T>
auto &dev_ctx = ctx.template device_context<platform::CPUDeviceContext>();
auto *place = dev_ctx.eigen_device();

Eigen::DSizes<int, 2> rshape(NxC, sample_size);
Eigen::DSizes<int, 2> param_shape(N, C);
Eigen::DSizes<int, 2> shape(NxC, sample_size);
#ifndef EIGEN_HAS_INDEX_LIST
Eigen::DSizes<int, 1> rdims(0);
Eigen::DSizes<int, 1> mean_rdims(1);
Eigen::DSizes<int, 2> rshape(NxC, sample_size);
Eigen::DSizes<int, 2> bcast(1, sample_size);
Eigen::DSizes<int, 2> C_shape(C, 1);
Eigen::DSizes<int, 2> NxC_shape(NxC, 1);
Eigen::DSizes<int, 2> param_shape(N, C);
Eigen::DSizes<int, 2> shape(NxC, sample_size);
#else
Eigen::IndexList<Eigen::type2index<0>> rdims;
Eigen::IndexList<Eigen::type2index<1>> mean_rdims;
Eigen::IndexList<Eigen::type2index<1>, int> bcast;
bcast.set(1, sample_size);
Eigen::IndexList<int, Eigen::type2index<1>> C_shape;
C_shape.set(0, C);
Eigen::IndexList<int, Eigen::type2index<1>> NxC_shape;
NxC_shape.set(0, NxC);
#endif

math::SetConstant<platform::CPUDeviceContext, T> set_constant;

Expand Down