Unification of BF16 enablement process

[wozna]

PR types

Function optimization

PR changes

Others

Describe

This PR unifies the addition of bfloat16 so that the process of adding quantize and dequantize reorders is similar to the quantization process. This way, pass cpu_quantize_squash_pass.cc can also be used in the bfloat16 process.

Therefore, the following changes have been added:

• cpu_bfloat16_pass.cc - adds quantize and dequantize before and after each operator which is marked as bfloat16
• cpu_quantzie_squash_pass.cc
• checks where we have dequantize and quantize, then inserts requantize` if scales are different or removes both operators when scale is the same (the latter is the most common in the case of bfloat16)
• removes dequantize op if it occurs before operators that has force _fp32_output attribute

Additionally, to cpu_quantzie_squash_pass.cc has been added:

• squash, where Scale operator and Quantize operator are merged
• squash where if we have Quantize operator before Conv2d bfloat16 op then remove Quantize because the input reorder can be executed inside the kernel

[paddle-bot-old]

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

[jczaja]

LGTM