flags to choose kp kernel (#38455)

This PR adds runtime flags run_kp_kernel, which choose which op to run for xpu2. There are two: dynamic linked and built from kp.

paddle/fluid/framework/operator.cc

@@ -56,6 +56,7 @@ DECLARE_bool(enable_unused_var_check);
 PADDLE_DEFINE_EXPORTED_int32(inner_op_parallelism, 0,
                              "number of threads for inner op");
 DECLARE_bool(run_pten_kernel);
+DECLARE_bool(run_kp_kernel);
 
 namespace paddle {
 namespace framework {

paddle/fluid/imperative/prepared_operator.cc

@@ -29,6 +29,7 @@
 DECLARE_bool(check_nan_inf);
 DECLARE_bool(run_pten_kernel);
 DECLARE_bool(benchmark);
+DECLARE_bool(run_kp_kernel);
 
 namespace paddle {
 namespace imperative {

paddle/fluid/platform/flags.cc

@@ -694,6 +694,18 @@ PADDLE_DEFINE_EXPORTED_bool(
 PADDLE_DEFINE_EXPORTED_bool(run_pten_kernel, true,
                             "It controls whether to use pten kernel");
 
+/**
+ * Pt kernel related FLAG
+ * Name: FLAGS_run_kp_kernel
+ * Since Version: 2.3.0
+ * Value Range: bool, default=false
+ * Example: FLAGS_run_kp_kernel=true would use the kp kernel to compute in
+ * the Op for XPU2.
+ * Note:
+ */
+PADDLE_DEFINE_EXPORTED_bool(run_kp_kernel, true,
+                            "It controls whether to use kp kernel for xpu2");
+
 /**
  * Distributed related FLAG
  * Name: FLAGS_allreduce_record_one_event