fix bugs of tipc, test=kunlun (#46540)

* migrate sigmoid with cross entropy, and tile xpu kernels to phi, test=kunlun

* migrate add_n kernep to phi, test=kunlun

* fix bugs of tipc, test=kunlun

paddle/fluid/framework/tensor_util.h

@@ -189,6 +189,11 @@ void TensorFromArray(const T* src,
         size,
         reinterpret_cast<const platform::CustomDeviceContext&>(ctx).stream());
   }
+#endif
+#ifdef PADDLE_WITH_XPU
+  else if (platform::is_xpu_place(dst_place)) {  // NOLINT
+    memory::Copy(dst_place, dst_ptr, src_place, src_ptr, size);
+  }
 #endif
   else {  // NOLINT
     PADDLE_THROW(platform::errors::Unimplemented(

paddle/fluid/platform/device/xpu/xpu2_op_list.h

@@ -471,6 +471,7 @@ XPUOpMap& get_kl2_ops() {
                      pOpKernelType(vartype::INT64, XPUPlace())})},
       {"scatter",
        XPUKernelSet({pOpKernelType(vartype::INT64, XPUPlace()),
+                     pOpKernelType(vartype::INT32, XPUPlace()),
                      pOpKernelType(vartype::FP32, XPUPlace())})},
       {"sampling_id",
        XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),

paddle/phi/kernels/assign_kernel.cc

@@ -174,4 +174,12 @@ PD_REGISTER_GENERAL_KERNEL(assign_array,
                            ALL_LAYOUT,
                            phi::AssignArrayKernel<phi::XPUContext>,
                            ALL_DTYPE) {}
+PD_REGISTER_KERNEL(assign_value,
+                   XPU,
+                   ALL_LAYOUT,
+                   phi::AssignValueKernel,
+                   bool,
+                   int,
+                   float,
+                   int64_t) {}
 #endif

paddle/phi/kernels/batch_norm_kernel.cc

@@ -88,3 +88,7 @@ PD_REGISTER_KERNEL(batch_norm_infer,
                    float,
                    phi::dtype::float16) {}
 #endif
+#ifdef PADDLE_WITH_XPU
+PD_REGISTER_KERNEL(
+    batch_norm_infer, XPU, ALL_LAYOUT, phi::BatchNormInferKernel, float) {}
+#endif

paddle/phi/kernels/empty_kernel.cc

@@ -126,4 +126,19 @@ PD_REGISTER_KERNEL(empty,
                    int64_t,
                    bool,
                    phi::dtype::float16) {}
+PD_REGISTER_KERNEL(empty_like,
+                   XPU,
+                   ALL_LAYOUT,
+                   phi::EmptyLikeKernel,
+                   float,
+                   double,
+                   int8_t,
+                   uint8_t,
+                   int16_t,
+                   int,
+                   int64_t,
+                   bool,
+                   phi::dtype::float16) {
+  kernel->InputAt(0).SetBackend(phi::Backend::ALL_BACKEND);
+}
 #endif

paddle/phi/kernels/xpu/stack_grad_kernel.cc

@@ -28,7 +28,7 @@ void StackGradKernel(const Context& dev_ctx,
   auto outs = x_grad;
   auto dy_dims = out.dims();
 
-  if (axis < 0) axis += dy_dims.size() + 1;
+  if (axis < 0) axis += dy_dims.size();
   auto dy_shape = phi::vectorize<int>(dy_dims);
 
   std::vector<int> dx_dims_list(x_grad.size(), 1);

paddle/phi/kernels/xpu/top_k_kernel.cc

@@ -135,11 +135,11 @@ void TopkKernel(const Context& dev_ctx,
     // Transpose back to original dims
     std::vector<int> trans_back_axes;
     for (int i = 0; i < axis; i++) {
-      trans_axes.emplace_back(i);
+      trans_back_axes.emplace_back(i);
     }
-    trans_axes.emplace_back(trans_out_dims.size() - 1);
+    trans_back_axes.emplace_back(trans_out_dims.size() - 1);
     for (int i = axis; i < trans_out_dims.size() - 1; i++) {
-      trans_axes.emplace_back(i);
+      trans_back_axes.emplace_back(i);
     }
 
     std::vector<int> trans_out_shape_host(trans_back_axes.size(), 0);