Skip to content

[Speed]Refine elementwise_mul_op gradient functor #8810

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
merged 1 commit into from
Mar 7, 2018
Merged

[Speed]Refine elementwise_mul_op gradient functor #8810

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
83 changes: 8 additions & 75 deletions paddle/fluid/operators/elementwise_mul_op.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,80 +40,14 @@ class ElementwiseMulKernel : public framework::OpKernel<T> {
};

template <typename T>
struct ElementwiseMulGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);

if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e * y_e;
}

if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = x_e * dz_e;
}
}
struct IdentityGrad_DX {
Copy link
Collaborator

Choose a reason for hiding this comment

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

This is not a IdentityGrad functor.
Please use a better name.

Copy link
Contributor Author

Choose a reason for hiding this comment

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

It has been renamed here.

HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout * y; }
};

template <typename T>
struct ElementwiseMulBroadCastGradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);

auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 2>(1, n))
.broadcast(Eigen::DSizes<int, 2>(pre, 1))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));

if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e * y_e_bcast;
}

if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (x_e * dz_e)
.reshape(Eigen::DSizes<int, 2>(pre, n))
.sum(Eigen::array<int, 1>{{0}});
}
}
struct IdentityGrad_DY {
HOSTDEVICE T operator()(T x, T y, T out, T dout) const { return dout * x; }
Copy link
Collaborator

Choose a reason for hiding this comment

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

This is not a IdentityGrad

};

template <typename T>
struct ElementwiseMulBroadCast2GradFunctor {
template <typename Device, typename X, typename Y, typename Z, typename dX,
typename dY, typename dZ, typename Pre, typename N, typename Post>
void operator()(Device d, X x, Y y, Z z, dX dx, dY dy, dZ dz, Pre pre, N n,
Post post) {
auto x_e = framework::EigenVector<T>::Flatten(*x);
auto y_e = framework::EigenVector<T>::Flatten(*y);
auto dz_e = framework::EigenVector<T>::Flatten(*dz);

auto y_e_bcast = y_e.reshape(Eigen::DSizes<int, 3>(1, n, 1))
.broadcast(Eigen::DSizes<int, 3>(pre, 1, post))
.reshape(Eigen::DSizes<int, 1>(x_e.size()));
if (dx) {
auto dx_e = framework::EigenVector<T>::Flatten(*dx);
dx_e.device(d) = dz_e * y_e_bcast;
}

if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
dy_e.device(d) = (x_e * dz_e)
.reshape(Eigen::DSizes<int, 3>(pre, n, post))
.sum(Eigen::array<int, 2>{{0, 2}});
}
}
};

template <typename DeviceContext, typename T>
class ElementwiseMulGradKernel : public framework::OpKernel<T> {
public:
Expand All @@ -127,12 +61,11 @@ class ElementwiseMulGradKernel : public framework::OpKernel<T> {
auto* dx = ctx.Output<Tensor>(framework::GradVarName("X"));
auto* dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
int axis = ctx.Attr<int>("axis");
ElementwiseGradCompute<DeviceContext, T, ElementwiseMulGradFunctor<T>,
ElementwiseMulBroadCastGradFunctor<T>,
ElementwiseMulBroadCast2GradFunctor<T>>(
ctx, x, y, out, dout, axis, dx, dy);
ElemwiseGradCompute<DeviceContext, T, IdentityGrad_DX<T>,
IdentityGrad_DY<T>>(ctx, *x, *y, *out, *dout, axis, dx,
dy, IdentityGrad_DX<T>(),
IdentityGrad_DY<T>());
}
};

} // namespace operators
} // namespace paddle
2 changes: 1 addition & 1 deletion paddle/fluid/operators/elementwise_op_function.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -301,7 +301,7 @@ struct ElemwiseGradNoBroadcast {
dx_[i] = dx_op_(x_[i], y_[i], out_[i], dout_[i]);
}
if (dy_ != nullptr) {
dy_[i] = dx_op_(x_[i], y_[i], out_[i], dout_[i]);
dy_[i] = dy_op_(x_[i], y_[i], out_[i], dout_[i]);
}
}

Expand Down