Changed back to Shard for SplitV to get better performance.

PiperOrigin-RevId: 187257148

diff --git a/tensorflow/core/kernels/split_v_op.cc b/tensorflow/core/kernels/split_v_op.cc
index 16fa890780555583f6ee24da10125d08045d766a..51d96a17b36071d41514e3a33c03ae2c89fc54a0 100644 (file)
--- a/tensorflow/core/kernels/split_v_op.cc
+++ b/tensorflow/core/kernels/split_v_op.cc
@@ -236,8 +236,9 @@ class SplitVOpCPUImpl {
     };
     if (use_parallelism_between_outputs) {
       // Run in parallel, disabling parallelism in functor.
-      context->device()->tensorflow_cpu_worker_threads()->workers->ParallelFor(
-          num_split, input_element_count / num_split, range_output_func);
+      Shard(num_split,
+            context->device()->tensorflow_cpu_worker_threads()->workers,
+            num_split, input_element_count / num_split, range_output_func);
     } else {
       // Run sequentially, but allow internal parallelism in functor.
       range_output_func(0, num_split);