change min_param_size_to_use_multithread to min_row_size_to_use_multithread

paddle/fluid/framework/operator.cc

@@ -30,7 +30,7 @@ DEFINE_bool(check_nan_inf, false,
             "Checking whether operator produce NAN/INF or not. It will be "
             "extremely slow so please use this flag wisely.");
 DEFINE_int32(inner_op_parallelism, 0, "number of threads for inner op");
-DEFINE_int32(min_param_size_to_use_multithread, 0, "");
+DEFINE_int32(min_row_size_to_use_multithread, 0, "");
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/operator.h

@@ -35,7 +35,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/variant.h"
 
 DECLARE_int32(inner_op_parallelism);
-DECLARE_int32(min_param_size_to_use_multithread);
+DECLARE_int32(min_row_size_to_use_multithread);
 
 namespace paddle {
 namespace framework {

paddle/fluid/operators/optimizers/adam_op.h

@@ -478,12 +478,12 @@ class AdamOpKernel : public framework::OpKernel<T> {
             }
           }
         } else if (FLAGS_inner_op_parallelism > 1 &&
-                   FLAGS_min_param_size_to_use_multithread > 0 &&
-                   param.numel() > FLAGS_min_param_size_to_use_multithread) {
+                   FLAGS_min_row_size_to_use_multithread > 0 &&
+                   param.dims()[0] > FLAGS_min_row_size_to_use_multithread) {
           VLOG(3) << "use multi thread, inner_op_parallelism="
                   << FLAGS_inner_op_parallelism
-                  << " min_param_size_to_use_multithread="
-                  << FLAGS_min_param_size_to_use_multithread;
+                  << " min_row_size_to_use_multithread="
+                  << FLAGS_min_row_size_to_use_multithread;
           if (FLAGS_inner_op_parallelism > 10) {
             LOG(WARNING) << "FLAGS_inner_op_parallelism "
                          << FLAGS_inner_op_parallelism << " is two large!";

python/paddle/fluid/__init__.py

@@ -129,7 +129,7 @@ def __bootstrap__():
         'eager_delete_tensor_gb', 'fast_eager_deletion_mode',
         'allocator_strategy', 'reader_queue_speed_test_mode',
         'print_sub_graph_dir', 'pe_profile_fname', 'warpctc_dir',
-        'inner_op_parallelism', 'min_param_size_to_use_multithread',
+        'inner_op_parallelism', 'min_row_size_to_use_multithread',
         'enable_parallel_graph'
     ]
     if 'Darwin' not in sysstr:

python/paddle/fluid/tests/unittests/CMakeLists.txt

@@ -87,7 +87,7 @@ list(REMOVE_ITEM TEST_OPS test_nearest_interp_op)
 foreach(TEST_OP ${TEST_OPS})
     py_test_modules(${TEST_OP} MODULES ${TEST_OP})
 endforeach(TEST_OP)
-py_test_modules(test_adam_op_multi_thread MODULES test_adam_op ENVS FLAGS_inner_op_parallelism=4 FLAGS_min_param_size_to_use_multithread=2)
+py_test_modules(test_adam_op_multi_thread MODULES test_adam_op ENVS FLAGS_inner_op_parallelism=4 FLAGS_min_row_size_to_use_multithread=2)
 py_test_modules(test_warpctc_op MODULES test_warpctc_op ENVS FLAGS_warpctc_dir=${WARPCTC_LIB_DIR} SERIAL)
 py_test_modules(test_bilinear_interp_op MODULES test_bilinear_interp_op SERIAL)
 py_test_modules(test_nearest_interp_op MODULES test_nearest_interp_op SERIAL)