* add relu user op

* rename kernel

* add python ifs nn.relu

* test relu fwd ok

* of format

* add namespace

* modify relu test file

* modify test_activations.py

* rm useless file

* fix

* update python interface

* fix matmul test

* update python api and test file

* SetInplaceProposalFn

* s->s

* AlwaysComputeWhenAllOutputsEmpty

* Add more log when finding kernel registration error

* Fix little bugs

* Fix the grammar of log

* Update log

* fix a bug in user_op reduce_sum InferSbp

* fix a bug in reduce_sum

* fix bugs in test_reduce_ops.py/reduce_ops.py

* add interface

* auto clang-format

* fix

* fix merge error

* rename IsSkippable to IsComputeSkippableWhenEmpty

* update reduce kernel

* rename IsComputeSkippableWhenEmpty() to AlwaysSkipWhenOutputEmpty()

* rename

* update

* fix format

* quick_fix_reduce_sum

* fix

* user pad op

* check sbp

* fix pr

* fix constant value

* try fix test

* fix pad grad

* update kernel impl

* fix format

* fix merge

* fix state

* fix state

* Update oneflow/customized/kernels/pad_kernel.cpp

fix

Co-Authored-By: Juncheng <liujuncheng1022@gmail.com>

* # This is a combination of 2 commits.
# The first commit's message is:

fix state

# This is the 2nd commit message:

fix

* fix state

Co-authored-by: james plur <jamesplur7@gmail.com>
Co-authored-by: Juncheng <liujuncheng1022@gmail.com>

* Fix compiling error on gcc 5.4 ubuntu 16.04, reference: https://gcc.gnu.org/bugzilla/show_bug.cgi\?id\=65899

* Add some comments

* Add reference link

* fix implicit conversion

* update

* format scalar add

* quick fix

* rm comment line

* support user op in ci

* fix test

* refine script

* fix syntax

* add name scope to user_op_builder

* hack this problem and backbone + fpn can run normally

* Extract the common compare_with_tf function and update three tests accordingly

* Seperate compare_with_tensorflow into small functions

* Replace flow_initializer with minval and maxval, try to generate numpy random array

* Reorder arguments of Run*Op

* Assert that all values in arg_dict have list type

* Fix no backward op error by adding a zero variable

* Update sqrt test

* Use zeros_initializer

* Format test_util.py

* Rename compare_with_tensorflow to CompareOpWithTensorFlow

* Fix test_malmul broken test

* Change param dict to unpackged param list

* Revert "Fix test_malmul broken test"

This reverts commit e85fa1568ea18cfb71555453278fe4b65deda9d5.

* check in

* fix indent

* fix case

* fix indent

* fix indent

* rm bulletin

* proper title for indent

* fix case and move cmake instruction up

* add title font for xla and trt

* rm cd build &&

* consitent 3rd party cmd

* check in

* fix indent

* fix case

* fix indent

* fix indent

* rm bulletin

* proper title for indent

* fix sbp bugs

* rm prelu

* add check

* user pad op

* check sbp

* fix pr

* fix constant value

* try fix test

* fix pad grad

* update kernel impl

Co-authored-by: james plur <jamesplur7@gmail.com>

* rm l2normalize

* add l2normalize user op

* address review

* rm xpuL2NormalizeKernel()

* fix test

* update interface

* naive 3 stage build

* fix docker authentication error

* minor fix

* Update docker/ci/base/Dockerfile

Co-Authored-By: Jin Yi <xxmyj.k@gmail.com>

* use pre build cmake

* aggressive merge yum cmds

* fix cmake

* dedicate 3rd party build

* rm base

* fix path

* reorder

* seperate stages

* try shrink

Co-authored-by: Jin Yi <xxmyj.k@gmail.com>

* argmax

* refine by review

* update interface

* cpu argmax

* ARG_MAX -> ARGMAX

* update interface

* set_memory_growth

* update by review

* update by review

* unsuccessed compiling

* fix

* gather_nd_op

* gather_nd op & kernel

* local_scatter_nd_upate autograd

* minor fix

* rm old version

* gather nd

* scatter nd

* gen grad

* simple test

* atomic add

* device() -> device_type()

* cpu kernel

* where_op

* forward

* modify by update

* revert

* fix typo

* where_grad

* test where

* fix bug

* more test

* check tmp null

* fix typo

* gen grad

* test grad

* where sbp

* fix where sbp

* complete sbp and batch axis

* elem-wise where

* update test

* fix where grad

* update test

* clear session for test_broadcast_to_compatible_with

* where decay to argwhere when x and y are none

* update test

* calc where grad using zero_like + where

* seperate kernel and kernel_util

* address review

* NaiveInferBatchAxis

* default no arguments constructor is not needed

* modify by update

* update test