revert_matmulv2_npu (#33014)

paddle/fluid/operators/matmul_v2_op_npu.cc

@@ -135,21 +135,8 @@ class MatMulV2GradNPUKernel : public framework::OpKernel<T> {
         }
         if (dy) {
           dy->mutable_data<T>(ctx.GetPlace());
-          framework::Tensor dout_;
-          TensorCopySync(*dout, ctx.GetPlace(), &dout_);
-          std::vector<int> vec_dim = framework::vectorize<int>(dout_.dims());
-          std::vector<int> vec_dim_v{vec_dim[0] * vec_dim[1], vec_dim[2]};
-          dout_.Resize(framework::make_ddim(vec_dim_v));
-
-          framework::Tensor x_;
-          TensorCopySync(*x, ctx.GetPlace(), &x_);
-          std::vector<int> vec_dim_x = framework::vectorize<int>(x_.dims());
-          std::vector<int> vec_dim_x_v{vec_dim_x[0] * vec_dim_x[1],
-                                       vec_dim_x[2]};
-          x_.Resize(framework::make_ddim(vec_dim_x_v));
-          auto runner_dy =
-              NpuOpRunner("MatMul", {x_, dout_}, {*dy},
-                          {{"transpose_x1", true}, {"transpose_x2", false}});
+          auto runner_dy = NpuOpRunner("BatchMatMul", {*x, *dout}, {*dy},
+                                       {{"adj_x1", true}, {"adj_x2", false}});
           runner_dy.Run(stream);
         }
       }