TemporaryAllocator is a singleton used for allocating memory for Cudnn. Since it is a singleton, we can delete it for better performance in memory.

We replace TemporaryAllocator by CUDADeviceContextAllocator and CUDADeviceContextAllocation, which uses stream callback to delete the memory allocated for the stream to avoid singleton.

Also added data_feed_proto to operator to fix CI in CPU compilation

* add pybind interface to get all inplace ops, test=develop

* enhance OpTest to check whether the consistency of operator when using and not using inplace, test=develop

* handle corner cases in op_test, test=develop

* support outputs without tensor holder_, like XShape in reshape_op, test=develop

* fix bug, some op has GradOpMaker, but actually no grad_op in OpInfoMap, test=develop

* use reshape_grad instead of reshape in FlattenGradOp, test=develop

* fix error debug dims info for variables like XShape, test=develop

* change computational order in sum_op to relieve computation difference using inplace, test=develop

* add inplace_atol to check group_norm, and skip inplace_grad for mkldnn, test=develop

* follow sneaxiy's comments, test=develop

* remove unused DefaultGradOpDescMaker in mkldnn op, test=develop

* feature/buffer_shared_inplace, test=develop

* refine code, test=develop

* fix elementwise_add op cpu inplace and sum inplace bug, test=develop

* add unittest and debug log, test=develop

* fix parallel_executor scope bug, polish code, test=develop

* fix sum op, activation op, single_in_place_inference bug, test=develop

* remove kLocalExecScopeName, test=develop

* fix unittest,test=develop

* fix out_var first version bug, test=develop

* follow comments,test=develop

* Update backward.py:
     - If there is no input grad var in all outputs of previous ops, do not append this op into graph.
     - Only apply this stragety when double backward.
* Update some double backward op.
* Update sum_op to judge whether a tensor is empty by numel or IsInitialized().

* Optimize the cuda implementation of sum_op, which add two lod_tensors inplace.
test=develop

* Use eigen to add to tensors.
test=develop

* optimize sum op

fuse multi eigen kernel calls into one cuda kernel.
refine code

test=develop.
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* Refine code.

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* Refine code according to comments.

test=develop

* refine code

delete sum_op_gpu.h
test=develop

* Fix test error.

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* refine code in format.

test=develop.

* refine code

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* refine code

test=develop
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

The macro should be defined by compiler rather than by source.

test=develop

* add fp16 backward support
test=develop

* add sum_op fp16 test

* disable test_dist_save_load
test=develop

* add check_grad for sum

* add unit test for softmax_grad fp16
test=develop

* add scale_op unit test

* add mul_grad_op unit test for fp16

* add cross_entropy_grad and eman_grad unit test for fp16
test=develop

* fix cross_entropy unit test

* add pool2d fp16 unit test

* refine conv2d fp16 unit test
test=develop

* refine activation unit test
test=develop

* fix ci
test=develop

* follow zhihong's comment, copy from https://github.com/PaddlePaddle/Paddle/pull/12796
test=develop

This reverts commit bf3c3496.

* "cherry picked operators changes"

* "remove duplicated code"

* "add constant setter"

* "add get expected kernel"

* "fix ci"

* "add fill constant"

There are mainly following fixes:

- take `DeviceContext` as the template parameter of math functors and OpKernel instead of `Place`
- remove `eigen_device` interface in base class  `DeviceContext`
- remove `GetEigenDevice` interface in `ExecutionContext` and base class `DeviceContext`
- remove unused `platform::EigenDeviceConverter`
- rename `REGISTER_OP_GPU_KERNEL` to `REGISTER_OP_CUDA_KERNEL`
- rename `USE_GPU_ONLY_OP` to `USE_CUDA_ONLY_OP`

* Support int64 for sum op

* Refine code

* Simplize Gradient Check

* Stash

* Extract apply_backward_pass to backward.py

Rename apply_backward_pass to append_backward_ops

* Use graph API to check gradient

* Fix ci

* Fix CI

* Fix backward for double precision

* Stash

* Fix CI

* Fix ci

* Ignore GRU test

* Ignore xe op

* Fix CI

* Fix softmax with xe gradient

The correct equation should be IG = OG * (d_softmax_with_xe())

* Fix typo

* Fix merge error

* Disable LRN

Make implement an operator less noisy.

* a simplest SGD op