Add fusion_gru and multi_gru to PTQ (Post-Training Quantization)

[wozna]

PR types

New features

PR changes

OPs

Describe

This PR:

• adds support for fusion_gru and multi_gru to Post-Training Quantization process
• adds gathering scales based on the max absolute value per output channel for fusion_gru/multi_gru operators
• adds GRU model test for PTQ for with multi_gru_fuse_pass and without it
• fix in the requantize operator so that there is no hard-coded data format but that it is set based on input

[paddle-bot-old]

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

[paddle-bot-old]

Sorry to inform you that 3b67f94's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[wozna]

I have a problem that in PR-CI-Coverage for file paddle/fluid/inference/api/mkldnn_quantizer.cc coverage is 0% even if all UT added in this PR run this code multiple times. I added the same changes in PR #33912 where I added a few prints and made the test not pass, then everything is written in the log and as you can see in the logs in PR-CI-Coverage the tests go through these parts of the code. I still don't know why only this file has line coverage equal to 0.

[lidanqing-intel]

LGTM

[wozna]

The recent changes are related to cmake files. It turned out that mkldnn_quantizer.cc was duplicated. It was added to the analysis_predictor library and then it was separately added to the paddle_inference_shared library where the analysis_predictor library was added as a dependency. That's why I left this file only in analysis_predictor lib.