Improve fast_tensor_util for bfloat16

[yongtang]

In #19212, improvement has been done to speed up the fast_tensor_util for float16. As both float16
and bfloat16 uses the same size, bfloat16 could be improved as well. This fix speeds up bfloat16 in a similiar fashion as float16.

This fix is related to #19212.

Signed-off-by: Yong Tang yong.tang.github@outlook.com

[tensorflowbutler]

Nagging Reviewer @rmlarsen: It has been 14 days with no activity and the awaiting review label was assigned. Can you please take a look?

[tensorflowbutler]

Nagging Reviewer @rmlarsen: It has been 29 days with no activity and the awaiting review label was assigned. Can you please take a look?

[tensorflowbutler]

Nagging Reviewer @rmlarsen: It has been 44 days with no activity and the awaiting review label was assigned. Can you please take a look?

[rmlarsen]

@yongtang is this covered by existing unit tests?

[yongtang]

@rmlarsen Yes the related unit test is in:

tensorflow/tensorflow/python/framework/tensor_util_test.py

Lines 238 to 257 in 6d8bc3c

	def testBfloat16(self):
	test_type = dtypes.bfloat16.as_numpy_dtype
	t = tensor_util.make_tensor_proto(np.array([10.0, 20.0], dtype=test_type))
	# 10.0: 16672 = 010000010(130) 0100000: (1+0/2+1/4) * 2^(130-127)
	# 20.0: 16800 = 010000011(131) 0100000: (1+0/2+1/4) * 2^(131-127)
	self.assertProtoEquals("""
	dtype: DT_BFLOAT16
	tensor_shape {
	dim {
	size: 2
	}
	}
	half_val: 16672
	half_val: 16800
	""", t)
	a = tensor_util.MakeNdarray(t)
	self.assertEquals(test_type, a.dtype)
	self.assertAllClose(np.array([10.0, 20.0], dtype=test_type), a)