* cherry-pick of DeviceContext Split, test=develop (#23737)

* New feature: thread local allocator, test=develop (#23989)

* add the thread_local_allocator, test=develop

* refactor the thread_local_allocator, test=develop

* provides option setting strategy, test=develop

* add boost dependency to cuda_stream, test=develop

* declare the stream::Priority as enum class, test=develop

* deal with PADDLE_ENFORCE_CUDA_SUCCESS macro in pr #23816

* cherry-pick,Optimize the error messages of paddle CUDA API

* fix the error messages of paddle CUDA API

* Refactoring PADDLE_ENFORCE_CUDA_SUCCESS, and apply to curand/cudnn/cublas/NCCL

* remove build_ex_string

As nccl comm is not created by CUDADeviceContext, it should be destroyed by the creator as the best practice of RAII.

cmake选项中添加了WITH_NCCL，显示指定是否编译NCCL的部分代码，WITH_NCCL默认打开，但如果WITH_GPU为OFF，则关闭WITH_NCCL

添加了PADDLE_WITH_NCCL定义

单机单卡能够关闭NCCL编译，多卡的话需要默认打开NCCL，如果关闭NCCL，则只能使用单卡
Co-authored-by: N石晓伟 <39303645+Shixiaowei02@users.noreply.github.com>

stack's wait cost a lot of cpu time, use cuda kernel to do memory copy
will reduce cpu time.
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* MKLDNN v1.0 rebase to Paddle 1.6
test=develop

* Add hacky paddle::string::to_string() implementation

* vectorize<int64-t>() -> vectorize() cleanup
test=develop

* PADDLE_ENFORCE and void_cast fixes
test=develop

* Rebase changes
test=develop

* Cosmetics
test=develop

* Delete MKL from mkldnn.cmake
test=develop

* CMake debug commands
test=develop

* Delete MKLDNN_VERBOSE and rebase fixes
test=develop

* Rebase fixes
test=develop

* Temporarily disable int8 resnet101 vgg16 and vgg19 tests
test=develop

* Add libmkldnn.so.1 to python setup
test=develop

* Add libmkldnn.so.1 to inference_lib cmake after rebase
test=develop

* Post rebase fixes + FC int8 changes
test=develop

* Fix LRN NHWC
test=develop

* Fix NHWC conv3d
test=develop

* Windows build fix + next conv3d fix
test=develop

* Fix conv2d on AVX2 machines
test=develop

cudaStreamSynchronize randomly hang when used in multi-thread environment, replace it with cudaStreamQuery API on windows

* Improve topk performance.

give 200000 data to compute topk,
before opt: cost 1s
after opt: cost 0.0028s.

* Refine return value.
* Add cuda util funtions.
* Fix ComputeBlockSize bug & refine comments.
Signed-off-by: Nzhaoyuchen <zhaoyuchen01@baidu.com>

* refine reallocate of workspace size, test=develop

* add lock to cudnn handle calls, test=develop

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

Polish fleet API to support cuda collective mode and nccl2 mode

* add config.SetMkldnnCacheCapacity api for mkldnn cache clear strategy

test=develop

* enhance MkldnnPostReset

test=develop

* add comments for mkldnn_cache_capacity field

test=develop

* add mkldnn shapeblob cache clear strategy

test=develop

* refine with comments

test=develop

* make cache clear strategy more safey

test=develop

* add lock for GetShapeBlobSize

test=develop

test=develop

* rename mkldnn set/get_cur_thread_id() to set/get_cur_mkldnn_session_id()

test=develop

* update session id definition and adjust logic for default behavior

test=develop

* reset logic in mkldnn reuse as most of cases work in default.

test=develop

test=develop

* remove nccl dep when the number of GPU is 1
test=develop

* remove attribute in Allocator::Allocate, test=develop

* fix travis ci error, test=develop

* fix cuda/cudnn version detection error, test=develop

* fix again, test=develop

test=develop

test=develop

test=develop

* Support Sync Batch Norm.
* Note, do not enable it in one device.

Usage:

build_strategy = fluid.BuildStrategy()
build_strategy.sync_batch_norm = True
binary = fluid.compiler.CompiledProgram(tp).with_data_parallel(
        loss_name=loss_mean.name,
        build_strategy=build_strategy)

reason: dereferencing smart pointer is the same as the underlying pointer
test=develop

test=develop

test=develop

test=develop
This reverts commit 064512aa.

* remove workspace_handle in conv2d_cudnn
test=develop

* remove workspace_handle
test=develop

* fix bug
test=develop

* make test_conv2d_op SERIAL
test=develop

* save memory in conv_cudnn
test=develop

* enhance thread safety
test=develop

* enhance temporary allocator
test=develop

* Add excess fraction
test=develop

* follow comments
test=develop

* fix bug and code refine
test=develop

* fix memory size check
test=develop

* rename reuse_tmp_allocation_excess_fraction
test=develop

test=develop

test=develop

test=develop