Matmul_v2 broadcasting mkldnn fuse passes

[baoachun]

Hi @lidanqing-intel @jczaja, this AR has a higher priority, and ends on the 29th. Thank you very much!

[paddle-bot-old]

您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

[baoachun]

Hi @lidanqing-intel @jakpiase , I am wondering whether we can add a new pass to map matmul_v2 to v1 to reduce the modification?

[jakpiase]

Hi, @baoachun, do mean adding a pass that will convert "matmul_v1" into "matmul_v2"? If yes, then I think that it can be done, but only if alpha(output scale) parameter would be equal to 1.0, otherwise matmul_v2 may not be compatible since it does not have that parameter. This additional pass will be added by your team, or it should be added by mine? Does it have the same deadline as adding other fuses(29.10.2021)?

[baoachun]

Hi, @jakpiase, my colleague will submit the pass of v2 mapping v1 today, but it only supports scenes without broadcast. Scenes that cannot be mapped by broadcast are not necessary to be resolved before the deadline. We need to ensure that the model in this mapping scenario can be inferred successfully and the related passes can work normally. In addition, yesterday PMO requested that cherry-pick codes be completely disallowed starting on the 25th. Thank you for your hard support!

[jakpiase]

@baoachun So should I create a cherry-pick PR of matmul_v2+transpose+reshape fuse pass now(#36481 )?

[baoachun]

@jakpiase I will ask the leader’s opinion and reply to you on Monday.
In addition, according to my colleague's pr, v2 has been mapped to v1, but ernie inference will have an inconsistent dimension error, when mkldnn is turned on. Can you help me see where the problem might be? The link to pr is: https://github.com/PaddlePaddle/Paddle/pull/36652/files

I did find that the two input shapes of the matmul operator do not match, the correct ones are:

dim_a.height_: 115
dim_a.width_: 64
dim_a.batch_size_: 32

dim_b.height_: 64
dim_b.width_: 115
dim_b.batch_size_: 32

the wrong ones are:

dim_a.height_: 115
dim_a.width_: 115
dim_a.batch_size_: 32

dim_b.height_: 115
dim_b.width_: 1024
dim_b.batch_size_: 2

[jakpiase]

@baoachun Ok, I'll try to find out why is it crashing in that PR