Fix scatter_op fp16 perf problem. (#38499)

* Fix scatter_op fp16 perf problem.

* Add scatter into black list.

* Add scatter into black list for dygraph.

python/paddle/fluid/contrib/mixed_precision/fp16_lists.py

@@ -18,7 +18,9 @@ from ... import core
 __all__ = ["CustomOpLists", "AutoMixedPrecisionLists"]
 
 # lookup_table fp16 is slower than fp32, though fp16 is supported.
-_extra_unsupported_fp16_list = {'lookup_table', 'lookup_table_v2'}
+_extra_unsupported_fp16_list = {
+    'lookup_table', 'lookup_table_v2', 'scatter', 'scatter_grad'
+}
 
 
 class AutoMixedPrecisionLists(object):

python/paddle/fluid/dygraph/amp/auto_cast.py

@@ -71,7 +71,9 @@ AMP_RELATED_FLAGS_SETTING = {
 }
 
 PURE_FP16_WHITE_LIST = {' '}
-PURE_FP16_BLACK_LIST = {'lookup_table', 'lookup_table_v2'}
+PURE_FP16_BLACK_LIST = {
+    'lookup_table', 'lookup_table_v2', 'scatter', 'scatter_grad'
+}
 
 
 #NOTE(zhiqiu): similar as paddle.fluid.contrib.mixed_precision.fp16_lists.AutoMixedPrecisionLists._update_list