[oneDNN] Fix to elementwise_add grad (#24639)

paddle/fluid/operators/elementwise/elementwise_op.h

@@ -104,7 +104,7 @@ class ElementwiseOp : public framework::OperatorWithKernel {
       int axis = ctx.Attr<int>("axis");
       int rankdiff = ctx.Input<Tensor>("X")->dims().size() -
                      ctx.Input<Tensor>("Y")->dims().size();
-      return (axis == -1) || (axis == rankdiff);
+      return (rankdiff == 0) || (axis == -1) || (axis == rankdiff);
     };
 
     if (platform::CanMKLDNNBeUsed(ctx) &&
@@ -243,9 +243,7 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
 #ifdef PADDLE_WITH_MKLDNN
     // If broadcasting is needed, use native implementation
     auto CanMKLDNNElementwiseAddGradBeUsed = [&]() {
-      auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
+      return (ctx.Input<Tensor>("X")->dims() == ctx.Input<Tensor>("Y")->dims());
-      auto dy = ctx.Output<Tensor>(framework::GradVarName("Y"));
-      return (dx != nullptr && dy != nullptr && dx->dims() == dy->dims());
     };
 
     if (platform::CanMKLDNNBeUsed(ctx) &&

paddle/fluid/operators/elementwise/mkldnn/elementwise_add_mkldnn_op.cc

@@ -85,6 +85,7 @@ class EltwiseAddMKLDNNGradKernel : public ElemwiseGradKernel<T> {
       in->set_format(out->format());
     };
 
+    // TODO(jczaja): Double check if vcopy works for blocked data
     auto blas = math::GetBlas<paddle::platform::CPUDeviceContext, T>(ctx);
     if (dx) {
       blas.VCOPY(dout->numel(), dout->data<T>(),