Skip to content
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

optimize some grad op using where #70992

Open
wants to merge 1 commit into
base: develop
Choose a base branch
from
+7 −3
Open

optimize some grad op using where #70992

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
4 changes: 3 additions & 1 deletion paddle/fluid/primitive/decomp_rule/decomp_rule/composite.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1264,7 +1264,9 @@ Tensor elu_decomp(const Tensor& x, const float alpha) {
zero = full<T>(x_cast.shape(), 0, x_cast.type(), x_cast.place());
tmp_res = alpha * (exp<T>(x_cast) - 1);
}
auto ans = where<T>(x_cast > zero, x_cast, tmp_res);
auto cond = cast<T>(x_cast > zero, x_cast.dtype());
auto ans =
cond * x_cast + (full_scalar<T>(1.0, x_cast.dtype()) - cond) * tmp_res;
return ConvertToOrig<T>(ans, x.dtype());
}

Expand Down
6 changes: 4 additions & 2 deletions paddle/fluid/primitive/decomp_rule/decomp_vjp/details.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1857,7 +1857,6 @@ void hardswish_grad(const Tensor& x, const Tensor& out_grad, Tensor* x_grad) {
t2 = cast<T>(t2, x.dtype());

auto res = out_grad * (t1 * t2 * (x / offset + factor) + one - t2);
// auto res = out_grad * (t1 * t2 * (x / offset + factor) );
set_output<T>(res, x_grad);
}
}
Expand All @@ -1870,7 +1869,10 @@ void leaky_relu_grad(const Tensor& out,
if (x_grad) {
auto zero = full_scalar<T>(0.0, out.dtype());
auto condition = greater_than<T>(out, zero);
auto res = where<T>(condition, out_grad, out_grad * negative_slope);
auto cond_cast = cast<T>(t1, out_grad.dtype());
auto res = cond_cast * out_grad +
(1 - cond_cast) * out_grad *
full_scalar<T>(negative_slope, out_grad.type());
set_output<T>(res, x_grad);
}
}
Expand Down
Loading