Execution times seem to be only CPU times, accelerator times are ignored

• TensorBoard version: 1.13.1
• OS Platform and version: Ubuntu 18.04
• Python version: 3.6

The following example code runs a simple graph on the GPU and records run metadata for a single run. It then writes the metadata to a logdir and uses it to generate profile info:

from datetime import datetime
import numpy as np
import tensorflow as tf

shape = (1024, 48, 48, 3)
input_img = np.ones(shape=shape, dtype=np.float32)
img = tf.placeholder(dtype=tf.float32, shape=shape)
out = tf.nn.conv2d(img, tf.ones(shape=(3, 3, shape[-1], 128)), [1, 1, 1, 1], 'SAME')

with tf.Session(config=tf.ConfigProto(gpu_options=tf.GPUOptions(allow_growth=True))) as sess:
    for _ in range(10): sess.run([out], {img: input_img}) # warm up before profiling
    profiler = tf.profiler.Profiler(sess.graph)
    writer = tf.summary.FileWriter('tensorboard_bug/{}'.format(datetime.now()), sess.graph)
    run_metadata = tf.RunMetadata()
    sess.run([out], {img: input_img}, tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE), run_metadata)
    writer.add_run_metadata(run_metadata, 'step0', 0)
    profiler.add_step(0, run_metadata)
    profiler.profile_name_scope(tf.profiler.ProfileOptionBuilder().build())

My expectation is that the timings in the profile match exactly. However, there is a huge discrepancy. It seems that tensorboard completely ignores the GPU computation time...

This is the profiler output:

Profile:
node name | total execution time | accelerator execution time | cpu execution time
_TFProfRoot (--/14.42ms, --/14.03ms, --/396us)
  Conv2D (14.39ms/14.39ms, 14.03ms/14.03ms, 369us/369us)
  Placeholder (0us/0us, 0us/0us, 0us/0us)
  _arg_Placeholder_0_0 (7us/7us, 0us/0us, 7us/7us)
  _retval_Conv2D_0_0 (3us/3us, 0us/0us, 3us/3us)
  edge_6__arg_Placeholder_0_0 (0us/0us, 0us/0us, 0us/0us)
  edge_7_Conv2D (0us/0us, 0us/0us, 0us/0us)
  ones (17us/17us, 0us/0us, 17us/17us)
    ones/Const (0us/0us, 0us/0us, 0us/0us)
    ones/shape_as_tensor (0us/0us, 0us/0us, 0us/0us)

and this is the corresponding info in tensorboard:

Reproduce the discrepancy with the provided Python snippet, focusing on profiler.profile_name_scope, profiler output, and the TensorBoard profile view. Trace how the recorded accelerator execution time is represented in the profile and make the displayed timings agree with the profiler output.