Make OpTest check grad inplace even if forward has no inplace (#19847)

* make OpTest check grad inplace even if forward has no inplace, test=develop

* do not run PE when enable_inplace is False, test=develop

* add conv3d cuda kernel for float16 type, test=develop

* refactor OpTest for inplace, test=develop

* add comments, test=develop

paddle/fluid/operators/conv_cudnn_op.cu.cc

@@ -509,7 +509,8 @@ REGISTER_OP_KERNEL(conv3d, CUDNN, plat::CUDAPlace,
                    paddle::operators::CUDNNConvOpKernel<plat::float16>);
 REGISTER_OP_KERNEL(conv3d_grad, CUDNN, plat::CUDAPlace,
                    paddle::operators::CUDNNConvGradOpKernel<float>,
-                   paddle::operators::CUDNNConvGradOpKernel<double>);
+                   paddle::operators::CUDNNConvGradOpKernel<double>,
+                   paddle::operators::CUDNNConvGradOpKernel<plat::float16>);
 REGISTER_OP_KERNEL(
     conv3d_grad_grad, CUDNN, plat::CUDAPlace,
     paddle::operators::CUDNNConvDoubleGradOpKernel<float>,

paddle/fluid/operators/elementwise/elementwise_mul_op.h

@@ -186,8 +186,7 @@ class ElementwiseMulDoubleGradKernel : public framework::OpKernel<T> {
   }
 };
 
-DECLARE_INPLACE_OP_INFERER(ElementwiseMulDoubleGradOpInplace, {"DDX", "DDOut"},
-                           {"X", framework::GradVarName("X")},
-                           {"Y", framework::GradVarName("Y")});
+DECLARE_INPLACE_OP_INFERER(ElementwiseMulDoubleGradOpInplace, {"DDX", "DDOut"});
+
 }  // namespace operators
 }  // namespace paddle