Add output defs for some kernels (#51333)

paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc

@@ -88,10 +88,6 @@ static std::set<std::string> OpsNeedSetOutputDtypeWhenRegisterPhiKernel = {
     "layer_norm_grad",
     "less_equal",
     "less_than",
-    "lstsq",
-    "lu",
-    "matrix_nms",
-    "matrix_rank_tol",
     "merged_adam",
     "mode",
     "momentum",

paddle/phi/kernels/cpu/lstsq_kernel.cc

@@ -301,4 +301,6 @@ void LstsqKernel(const Context& dev_ctx,
 
 }  // namespace phi
 
-PD_REGISTER_KERNEL(lstsq, CPU, ALL_LAYOUT, phi::LstsqKernel, float, double) {}
+PD_REGISTER_KERNEL(lstsq, CPU, ALL_LAYOUT, phi::LstsqKernel, float, double) {
+  kernel->OutputAt(2).SetDataType(phi::DataType::INT32);
+}

paddle/phi/kernels/cpu/lu_kernel.cc

@@ -73,4 +73,7 @@ void LUKernel(const Context& dev_ctx,
 
 }  // namespace phi
 
-PD_REGISTER_KERNEL(lu, CPU, ALL_LAYOUT, phi::LUKernel, float, double) {}
+PD_REGISTER_KERNEL(lu, CPU, ALL_LAYOUT, phi::LUKernel, float, double) {
+  kernel->OutputAt(1).SetDataType(phi::DataType::INT32);
+  kernel->OutputAt(2).SetDataType(phi::DataType::INT32);
+}

paddle/phi/kernels/cpu/matrix_nms_kernel.cc

@@ -318,4 +318,7 @@ void MatrixNMSKernel(const Context& ctx,
 }  // namespace phi
 
 PD_REGISTER_KERNEL(
-    matrix_nms, CPU, ALL_LAYOUT, phi::MatrixNMSKernel, float, double) {}
+    matrix_nms, CPU, ALL_LAYOUT, phi::MatrixNMSKernel, float, double) {
+  kernel->OutputAt(1).SetDataType(phi::DataType::INT32);
+  kernel->OutputAt(2).SetDataType(phi::DataType::INT32);
+}

paddle/phi/kernels/cpu/matrix_rank_tol_kernel.cc

@@ -175,4 +175,5 @@ void MatrixRankTolKernel(const Context& dev_ctx,
 
 PD_REGISTER_KERNEL(
     matrix_rank_tol, CPU, ALL_LAYOUT, phi::MatrixRankTolKernel, float, double) {
+  kernel->OutputAt(0).SetDataType(phi::DataType::INT64);
 }

paddle/phi/kernels/gpu/lu_kernel.cu

@@ -183,6 +183,9 @@ PD_REGISTER_KERNEL(lu,  // cuda_only
                    ALL_LAYOUT,
                    phi::LUKernel,
                    float,
-                   double) {}
+                   double) {
+  kernel->OutputAt(1).SetDataType(phi::DataType::INT32);
+  kernel->OutputAt(2).SetDataType(phi::DataType::INT32);
+}
 
 #endif  // not PADDLE_WITH_HIP

paddle/phi/kernels/gpu/matrix_rank_tol_kernel.cu

@@ -445,6 +445,8 @@ PD_REGISTER_KERNEL(matrix_rank_tol,  // cuda_only
                    ALL_LAYOUT,
                    phi::MatrixRankTolKernel,
                    float,
-                   double) {}
+                   double) {
+  kernel->OutputAt(0).SetDataType(phi::DataType::INT64);
+}
 
 #endif  // not PADDLE_WITH_HIP