elementwise max min #7538
elementwise max min #7538
Conversation
| }; | ||
|
|
||
| template <typename DeviceContext, typename T> | ||
| class ElementwiseMaxKernel : public framework::OpKernel<T> { |
There was a problem hiding this comment.
It seems that Max and Min are almost the same logic. We can reuse the code. Please refer to https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/reduce_op.h#L97 and https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/operators/reduce_op.h#L253
… dev_elementwise_max_min
… dev_elementwise_max_min
| } | ||
| } | ||
| }; | ||
|
|
There was a problem hiding this comment.
@JiayiFeng We can reduce more codes. But we can do it later, if we have more element-wise operators to implement. Anyway, let's find a general pattern and clean code later.
template <typename T>
struct MaxGradFunctor1 {
template <typename Device, typename Dxe, typename Xe, typename Ye, typename Dze>
void operator()(Device d, Dxe dx_e, Xe, x_e, Ye, y_e, Dze, dz_e) {
dx_e.device(d) = (x_e > y_e).template cast<T>() * dz_e;
}
};
template <typename T>
struct MinGradFunctor1 {
template <typename Device, typename Dxe, typename Xe, typename Ye, typename Dze>
void operator()(Device d, Dxe dx_e, Xe, x_e, Ye, y_e, Dze, dz_e) {
dx_e.device(d) = (x_e < y_e).template cast<T>() * dz_e;
}
};
template <typename T, typename Functor1, typename Functor2>
struct ElementwiseMinMaxGradFunctor {
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);
Functor1 functor;
functor(d, dx_e, x_e, y_e, dz_e);
}
if (dy) {
auto dy_e = framework::EigenVector<T>::Flatten(*dy);
Functor2 functor;
functor(d, dy_e, x_e, y_e, dz_e);
}
}
};There was a problem hiding this comment.
@QiJune
I agree. Finding a general pattern in the backward pass is not as easy as that in forwarding pass, for backward op takes Z, dZ, X, Y while forwarding op only takes X and Y.
… dev_elementwise_max_min
solve #7567