* Remove output arguments from functions.
Replace pointers with references

* Name used bool flags

* Reorder functions

* Enable bfloat16 data type

* Give declarations some space

* Style

* Style

1. test_parallel_executor_seresnext_base_gpu failed on 2 P100 GPUs with `470.82` driver.
```
======================================================================
FAIL: test_seresnext_with_learning_rate_decay (test_parallel_executor_seresnext_base_gpu.TestResnetGPU)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/paddle/paddle/build/python/paddle/fluid/tests/unittests/test_parallel_executor_seresnext_base_gpu.py", line 32, in test_seresnext_with_learning_rate_decay
    self._compare_result_with_origin_model(
  File "/opt/paddle/paddle/build/python/paddle/fluid/tests/unittests/seresnext_test_base.py", line 56, in _compare_result_with_origin_model
    self.assertAlmostEquals(
AssertionError: 6.8825445 != 6.882531 within 1e-05 delta (1.335144e-05 difference)
----------------------------------------------------------------------
```
2. To be more accuracte on evaluating loss convergence, we proposed to apply IOU as metric, instead of comparing first and last loss values.
3. As offline discussion, we also evaluated convergence on P100 and A100 in 1000 interations to make sure this UT have the same convergence property on both devices. The curves are showed below.
![A100-Single, P100-Single and Diff (1)](https://user-images.githubusercontent.com/13541238/175461920-25df6101-6dd8-4387-862c-d1c8e9299c57.png)

* [Sparse]add SparseTensor mv kernel(csr*dense_vec->dence_vec, coo*dense_vec->dense_vec)

* fix CI

* fix merge_grad&push_sparse

* change typo

* fix code format. test=develop

* fix code format. test=develop

* fix code format. test=develop

* fix debug info

* optimize hbm

* fix size_t

* fix size_t

* boost::variant -> paddle::variant

* boost::variant.apply_visit -> paddle::visit

* Update pybind_boost_hraders.h

* Fix CINN compilation errors

* Revert FetchResultType

This reverts commit f8b9073f.

* Polish Layer code

* Refine some api name

* Polish load program code

* Clean smoe include

* Process windows style path

* Move some funtions to utils

* record memory and op supplement info

* update

* update

* fix a bug

* fix memory recording

* fix a bug

* update

* update

* fix a bug

* update

* fix a bug

* fix a bug

* fix a bug

* update dygraph record

* add infer shape record

* fix

* fix

* fix

* add comments

* fix a bug

* fix

* fix

* inplace op without backward supports input of multiple inplace vars

* support inplace without backward in eager mode and fix python_c in eager_fluid

* fix output in yaml

* fix map to const&

* solve conflict v2

* enable Jacobian,Hessian supporting new autograd

* fix prim mode failed in PR-CI-Windows

* add forward_gradients api

* add forward_gradients api

* skip test_autograd_functional_prim in windows ci

* fix test_autograd_funciton_prim timeouot

* remove the block parameter in prim2orig method

* remove duplicate to_tensors code snippet # test=allcases

The paddle-cinn failed recently and block all PRs from merging. We found the key error reason from the log:

![b207bbfd3ebfabb2141a88b85241420b](https://user-images.githubusercontent.com/7913861/175882664-4dd14059-b716-44c4-a975-528de17ace46.png)

The CI environment didn't contain llvm.

We then debug and found https://github.com/PaddlePaddle/Paddle/pull/43534 deleted the llvm due to clang-format upgrade:
![f9d07b8e3c2b2cba7a790716961d98cf](https://user-images.githubusercontent.com/7913861/175883020-1a29c538-e511-4ada-a899-9b3ced3c0ae2.png)

So this PR added the llvm back. However, `clang-llvm-3.8.0` contains `clang-format-3.8`, so we modified the old command `cp -r` to `cp -rn`, which could ignore the installed `clang-format-13`

* make inference_api_test use import_lib + dll to smaller its size;test=document_fix;test=windows_ci_inference

* make place.h not export symbols; change the order of inference test dependencies

* cant delete analysis pass ..in inference_analysis_test or build test_analyzer.exe will failed; temporarily comment test need paddle_inference_c

* add onnx path; cp onnxruntime.dll to tests/api

* make c_ut depend on paddle_inference_c_shared rather than paddle_inference_c

* make c_ut depend .dll in windows and .lib in linux;path add capi_exp

* init context pool by analysis predictor

* make windows and linux both use shared; compile shared even on_infer off

* move case:profile after case:fuse_statis to avoid context pool null problem

* make c_ut depend .dll in windows and .lib in linux;path add capi_exp

* [Dy2Stat]Refactor convert_shape transformer logic

* clean usless unittest

* rename eagerpylayer

* move apis from legacy_api.yaml to api.yaml

* remove comment

* revert argsort

* fix merge conflict

* [CustomDevice]add custom place supports

* sync format

* update pateches

* test=document_fix; add patch file

* Update kaiming.py

* Update initializer.py

* fix doc bug;test=document_fix

* fix doc;test=document_fix

* Update initializer.py

* Update kaiming.py

* for ci;test=document_fix
Co-authored-by: NLigoml <39876205+Ligoml@users.noreply.github.com>

use jinja 2.11.3 in python lower than 3.6.2

* add get_op_names api

* Update pybind.cc

* tmp fix

* init

* compile ok

* compile ok

* add vlogs

* add test

* fix termination error

* add testfile

* add

* fix window compile

* fix window compile

* fix windows compile

* fix windows compile

* fix windows compile

* fix windows compile

* fix windows compile

* fix windows compile

* fix kunlun compile

* fix compilation

* fix compilation

* fix compilation

* tmp fix

* add windows

* add windows

* add more logs

* change timeout to protected

* SB

* add

* add

* fix timeout

* add

* fix test

* fix test

* fix test

* fix ut

* fix ut

* fix ut