Added cast op oneDNN kernel for bf16/fp32 datatypes casting(FWD/BWD)

[jakpiase]

PR types

New features

PR changes

OPs

Describe

Added cast op for 4 types of casts (fp32->fp32, fp32->bf16, bf16->fp32, bf16->bf16) for enabling Word2Vec model. OneDNN kernel can be used both in FWD and BWD.

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[jczaja]

LGTM

[lidanqing-intel]

@jakpiase Hi Please rename the PR title or make it more clear because I will use these titles for release note.
I notice another cast op FP32/BF16 PR has been merged. What are the difference

[arlesniak]

LGTM

[arlesniak]

LGTM

[jakpiase]

@luotao1 Could you please start your review?

[luotao1]

Paddle/python/paddle/fluid/tests/unittests/op_test.py

Lines 1076 to 1090 in 394d8d4

	def check_output_with_place(self,
	place,
	atol=0,
	no_check_set=None,
	equal_nan=False,
	check_dygraph=True,
	inplace_atol=None):
	self.infer_dtype_from_inputs_outputs(self.inputs, self.outputs)
	if self.dtype == np.float64 and \
	self.op_type not in op_threshold_white_list.NEED_FIX_FP64_CHECK_OUTPUT_THRESHOLD_OP_LIST:
	atol = 0
	if self.is_bfloat16_op():
	check_dygraph = False
	if hasattr(self, 'force_fp32_output') and getattr(

since you add check_dygraph = False when bf16_op in check_output_with_place, could you add same code in check_grad_with_place? Thus, you can decrease one approve

@jakpiase @lidanqing-intel @jczaja

@wzzju Please note this.