* is head regst desc where sharing same mem

* fix typo

* feat: add UseCudnnOnGpu in Operator and fix conv op

* feat: add WithCudnn and WithoutCudnn in ConvKernel<kGPU, T>

* feat: add CUDA NCDHWIm2ColGpu kernel and compile done

* refine: rename Im2ColNCDHWGpu() to NCDHWIm2ColGpu()

* fix: reverse update about int32_t to int64_t in BlocksNum4ThreadNum()

* feat: add CUDA NCDHWCol2ImGpu kernel

* refactor: extract InitSharedArrays() for device code

* feat: add CUDA NDHWCIm2ColGpu kernel

* feat: add CUDA NDHWCCol2ImGpu kernel and fix typos

* fix: fix the bug of calc im_offset when NDHWCIm2ColGpu

* refactor: extract Im2ColCalcKernelAndOutIndex() and Im2ColCalcImIndex()

* fix: fix format and the missing shared_im[] parameter in Im2ColCalcImIndex()

* refactor: merge NCDHWCol2ImGpu() and NDHWCCol2ImGpu() into Col2ImGpu()

* refactor: merge NCDHWIm2ColGpu() and NDHWCIm2ColGpu() into Im2ColGpu()

* feat: add class ConvKernelImplByIm2Col between ConvKernelIf and ConvKernel; compile done, to be run

* fix: add explicit template instantiation for ConvKernelUtil

* refine: remove unused class function declaration: KernelInitWithoutCudnn e.g.

* fix(operator/conv_op.cpp): make sure UseCudnnOnGpu() == true when infer cudnn algo

* refine(kernel/conv_kernel.cu): let the gpu kernel function be inside the anoymous namespace

* refactor: add dim_num as the template paramter of Im2ColGpu() and Col2ImGpu()

* refactor: add is_channel_first as the template paramter of Im2ColGpu() and Col2ImGpu()

* refine(kernel/conv_kernel.cu): add  #undef IM2COL_FUNC_CALL

* refine(kernel/conv_kernel.cu): add dim_num as the template parameter of InitSharedMemory()

* fix(kernel/conv_kernel.cu): fix the bug of use col_offset in Im2ColGpu()

* refine regst manager

* refine regst mananger version 2

* remove memInfo

* refine code and reduce rows

* using protobuf message differencer

* explicit

* memory allocator free memory

* add regst desc proto list init for test code

* remove cplusplus_17.h, it's no need now.

* improve make_unique

* make macros together

* remove OF_CALL_ONCE macro

* remove macros in ctrl_server

* -Wno-unused-function

* remove useless header

* refine vgg proto && add report

* refine proto

* add proto for 2 machine

* add proto for 2 machine 4 GPU and measurement of reduce time for 2 machine 8 GPU

* refine protos

* remove comment

* out delay regst

* no delay edge to kMdSave task

* fix bug when elem cnt in model < split num

* add checker

* let opencv depend on libjpeg-turbo

* fix libjpeg-turbo dependency

* remove copy by itor

* use already existed copy method

* delete forward field override and use CopyField instead

* remove needless

* use libjpeg-turbo

* refine format

* parse /proc/meminfo

* CHECK for getMemInfo

* refine strToToken

* for review

* use onefloe cast

* set regst max=min when same stream

* move fix regst num to task node