[AutoParallel-Performance] AMP Flag Memcpy support newexe Overlap (#49219)

* memcpy overlap

* memcpy newexe

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

@@ -320,6 +320,11 @@ OpFuncType AnalyseOpFuncType(const OpFuncNode& op_func_node,
     return OpFuncType::kGpuSync;
   }
 
+  // for memcpy explicitly called by user
+  if (platform::is_gpu_place(place) && op->Type() == interpreter::kMemcpyD2H) {
+    return OpFuncType::kGpuSync;
+  }
+
   if (op->Type() == "shape") {
     return OpFuncType::kGpuSync;
   }

python/paddle/distributed/passes/auto_parallel_fp16.py

@@ -663,8 +663,6 @@ def _insert_memcopy(block, idx, src_var, dist_context, direction="D2H"):
     # TODO to support CUDAPinned/NPU/XPU Places
     if direction == "D2H":
         dst_place_type = 0
-    elif direction == "D2H":
-        dst_place_type = 1
     else:
         raise NotImplementedError(
             "direction [{}] is not supported yet.".format(direction)
@@ -673,7 +671,7 @@ def _insert_memcopy(block, idx, src_var, dist_context, direction="D2H"):
     attrs = {'dst_place_type': dst_place_type}
     new_op = block._insert_op_without_sync(
         index=idx,
-        type='memcpy',
+        type='memcpy_d2h',
         inputs={'X': [src_var]},
         outputs={'Out': [output_var]},
         attrs=attrs,