System information

• Have I written custom code (as opposed to using a stock example script provided in TensorFlow):
• yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
• Linux Ubuntu 18.04
• Mobile device (e.g. iPhone 8, Pixel 2, Samsung Galaxy) if the issue happens on mobile device:
• TensorFlow installed from (source or binary):
• binary wheel via PyPI
• TensorFlow version (use command below):
• v2.5.0-0-ga4dfb8d1a71 2.5.0
• Python version:
• 3.6
• CUDA/cuDNN version:
• CUDA 11.2
• GPU model and memory:
• V100 32 GB

Describe the current behavior
When slicing tensors of shape int16 or int32 it takes significant amount of time comparing to float16, float32

type array: <dtype: 'int16'>
took 35.05420684814453

type array: <dtype: 'int32'>
took 22.861242294311523

type array: <dtype: 'int64'>
took 5.330085754394531

type array: <dtype: 'float16'>
took 1.550912857055664

type array: <dtype: 'float32'>
took 2.5637149810791016

type array: <dtype: 'float64'>
took 5.917549133300781

type array: <dtype: 'bool'>
took 1.6639232635498047

Describe the expected behavior
takes around the same time as the float versions

Contributing

• Do you want to contribute a PR? (yes/no): no

when running the code using tf.debugging.set_log_device_placement(True) found that slicing int16 is not using gpu at all:
type array: <dtype: 'int16'>
Executing op Mul in device /job:localhost/replica:0/task:0/device:GPU:0
Executing op StridedSlice in device /job:localhost/replica:0/task:0/device:CPU:0
Executing op StridedSlice in device /job:localhost/replica:0/task:0/device:CPU:0
Executing op StridedSlice in device /job:localhost/replica:0/task:0/device:CPU:0
Executing op StridedSlice in device /job:localhost/replica:0/task:0/device:CPU:0
Executing op StridedSlice in device /job:localhost/replica:0/task:0/device:CPU:0

for int32 I don't understand the reason for such difference comparing to float32

Standalone code to reproduce the issue

import tensorflow as tf
# tf.debugging.set_log_device_placement(True)
import numpy as np
import time

def test_slicing():

    with tf.device('/GPU:0'):
      shape = (52000, 2, 15, 15)
      np_array = np.ones(shape, dtype=np.uint16)

      dtypes= [tf.uint16, tf.int16, tf.int64, tf.float16, tf.float32, tf.float64, tf.bool]
      start = 0
      end = 50000
      for dtype in dtypes:
        print("*****************************************************")

        tf_array = tf.constant(np_array, dtype=dtype) * 1 if dtype != tf.bool else tf.math.logical_and(tf.constant(np_array, dtype=dtype) , tf.constant(np_array, dtype=dtype)) 
        print(f'type array: {tf_array.dtype}')
        
        for i in range(1):
            tf_array = tf.constant(np_array, dtype=dtype) * 1 if dtype != tf.bool else tf.math.logical_and(tf.constant(np_array, dtype=dtype) , tf.constant(np_array, dtype=dtype)) 

            tic = time.time()
            a = tf_array[start:end]
            a[0][0][0][0].numpy()
            toc = time.time()
            print(f'took {(toc - tic)*1000}')

print()          
print(tf.version.GIT_VERSION, tf.version.VERSION)
test_slicing()