Fix bug in elementwise_mul/div_grad when inplace strategy (#38840)

* fix bug when inplace strategy

* fix

* fix

* fix

* fix

* fix

paddle/fluid/operators/elementwise/elementwise_div_op.cu

@@ -31,20 +31,10 @@ ElementwiseDivGrad(const framework::ExecutionContext& ctx,
   const auto& dev_ctx = ctx.template device_context<DeviceContext>();
   const auto place = ctx.GetPlace();
   if (dx != nullptr && dy != nullptr) {
-    dx->mutable_data<T>(place);
-    if (dx->IsSharedBufferWith(*dout)) {
-      dx->clear();
-      dx->mutable_data<T>(x->dims(), place);
-    }
     std::vector<const framework::Tensor*> ins = {dout, out, y};
     GetGradXAndYOut<ElementwiseType::kTernary, T>(
         dev_ctx, place, axis, ins, dout, dx, dy, DivGradXYFunctor<T, T>());
   } else if (dx != nullptr && dy == nullptr) {
-    dx->mutable_data<T>(place);
-    if (dx->IsSharedBufferWith(*dout)) {
-      dx->clear();
-      dx->mutable_data<T>(x->dims(), place);
-    }
     std::vector<const framework::Tensor*> ins = {dout, y};
     GetGradXOrYOut<ElementwiseType::kBinary, T>(dev_ctx, place, axis, ins, dout,
                                                 dx, DivGradXFunctor<T>());

paddle/fluid/operators/elementwise/elementwise_mul_op.cu

@@ -74,20 +74,10 @@ ElementwiseMulGrad(const framework::ExecutionContext& ctx,
   const auto place = ctx.GetPlace();
 
   if (dx != nullptr && dy != nullptr) {
-    dx->mutable_data<T>(place);
-    if (dx->IsSharedBufferWith(*dout)) {
-      dx->clear();
-      dx->mutable_data<T>(x->dims(), place);
-    }
     std::vector<const framework::Tensor*> ins = {dout, y, x};
-    GetGradXAndYOut<ElementwiseType::kBinary, T>(
+    GetGradXAndYOut<ElementwiseType::kTernary, T>(
         dev_ctx, place, axis, ins, dout, dx, dy, MulGradXYFunctor<T, T>());
   } else if (dx != nullptr && dy == nullptr) {
-    dx->mutable_data<T>(place);
-    if (dx->IsSharedBufferWith(*dout)) {
-      dx->clear();
-      dx->mutable_data<T>(x->dims(), place);
-    }
     std::vector<const framework::Tensor*> ins = {dout, y};
     GetGradXOrYOut<ElementwiseType::kBinary, T>(dev_ctx, place, axis, ins, dout,
                                                 dx, MulGradFunctor<T>());

paddle/fluid/operators/elementwise/elementwise_op_function.h

@@ -2575,6 +2575,7 @@ void GetGradXAndYOut(const platform::CUDADeviceContext &dev_ctx,
                      framework::Tensor *dy, Functor func) {
   framework::Tensor tmp_dx;
   framework::Tensor tmp_dy;
+  dx->mutable_data<T>(place);
   dy->mutable_data<T>(place);
   std::vector<framework::Tensor *> outs;
   if (dx->dims() == dout->dims() && dy->dims() == dout->dims()) {