* qkv

* ci_test

* add trt LT version helper

* trt8 requires void** to be void* const*

* 1, remove layernorm dynamic fp16; 2, let reshape out in dynamic shape

* remove useless code

* add dynamic layer_norm plugin

* fix bug

* fix numpy.allclose

* fix format

* fix code style

* remove shepe in dynamic shape

* code format

* remove layer norm fp16

* fix format

* implement MHA order same as training

* fix fp16 compile issue on old architecture

* fix format

* fix format

* fix yolobox teller condition

* fix cuda double free bug

* update trt engine addplugin name.

* update

* add anchor generator op plugin

* add anchor generator unit_test

* remove dbg info

* remove redundant line

* replace assertion with paddle enforce

* dynamic plugin replaces assertion with paddle enforce

* anchor generator support dynamic shape on spatial axis

* anchor generator test with fp16, dynamic shape

* add anchor generator test all

* add back main

* reduce test input size to not exceed the timelimit of ci

* change super to InferencePassTest for python2 compatibility

* reuse paddle operator anchor generator

* move creator construct to header with default

* add cuda ifdef

* reduce line

* change super to InferencePassTest for python2 compatibility

* fix anchor generator fp16 serialize setting

* split unittest from test_all

* restrict anchor generator input format before version 7234

* anchor generator only support greater than trt7.1

* change min_graph_size to 2

* min_graph size to 3 if dynamic shape

* reduce dynamic shape size to avoid trt search tactic too long to exceed time limit

* remove anchor from fetch list

* anchor generator support all trt version

* fix memory not allocated but if serialized

* yolobox converter and plugin

* yolobox unittest

* add dynamic shape restriction

* fix git merge log

* add roi_align_plugin

* add roi align unit_test

* add roi align serialization

* remove roi align static plugin because of batch dim issue

* refine roi align unittest and add fp16/serialization

* add trt roi align condition to op_teller

* refine error message

* remove unnecessary reshape layer

* fix tensorrt output varible reshape

* move padding shape x 1 x 1 in ernie to qkv and fc

* update layer name

* fix softmax when input is dynamic, fc not padding any more

* fix varlen

* move fc x_dim assert to op_teller

* add serialize unittest

* fix element_op trt plugin serialize bug

* fix split trt plugin initialize

* update

* use iwyu clean include second time, test=develop

* fix trt plugin clone and initialize bugs

* fix unit test error

* enable trt in ci py3

* update unittest timeout

* remove -DSUPPORTS_CUDA_FP16 in cuda.cmake

* comile with cuda9

* add some unittest

* notest;test=coverage

* add unittest for trt plugin swish && split

* update ernie unittest

* fix some error message

* remove repeated judgement of CUDA version in mbEltwiseLayerNormOpConverter

* fix comile errror when CUDA_ARCH_NAME < Pascal"

* fix comile error

* update unittest timeout

* compile with cuda9

* update error msg

* fix code style

* add some comments

* add define IF_CUDA_ARCH_SUPPORT_FP16

* rename IF_CUDA_ARCH_SUPPORT_FP16 to CUDA_ARCH_FP16_SUPPORTED

* fp16 result ok

* change -DWITH_NVINFER_PLUGIN toconfig.EnableTensorRtOSS

* auto detect special slice op converter for ernie with trt oss

* ernie oss only support fp16

* fix special_slice_plugin serialize bug

* matmul in tensorrt ok

* ernie unittest ok

* add matmul tensorrt unittest

* remove demo code

* add unittests and op version register for tensorrt_subgraph_pass

* rename to test_trt_subgraph_pass.py

* fix softmax converter diff when padding dim=1

* use iwyu clean include, test=develop, test=win

* compilation error, test=develop

* fix compilation error2, test=develop

* fix compilation error3, test=develop

* fix compilation error4, test=develop

* fix compilation error5, test=develop

* fix compilation error6, test=develop

* fix compilation error7, test=develop

* fix compilation error8, test=develop

* fix compilation error8, test=develop

* fix compilation error10, test=develop

* fix compilation error11, test=develop

* polish some lost error msg

* add some math file to white list

* polish detail based reviewer commnet

* optimize slice TRT plugin

This patch removes unnecessary barrier for data transfer of needed offset,
so data transfer can be overlap with GPU kernel execution.

This patch also fixes incorrect name of slice plugin. That is, replaces
"layernorm" with "slice"

test=develop

* add serialize/deserialize to slice plugin

* add static shape slice trt plugin

* fix slice trt op convertor dynamic shape bug

* fix format by clang-format

* fix pylint format error

* fix problems commented by peiyang
Co-authored-by: NRyan Jeng <rjeng@nvidia.com>

test=develop

* add stack_op to CMakeLists

* add dim=3 support for scale op

* add trt stack op, test=develop

* remove debug message

* add stack plugin serialize

* remove slice, scale op, will add later

* enhence error message

* revise trt ernie test to conver the stack op CI testi, test=develop

* add stack op serialization

* fix test shape after adding stack op

* remove slice op, will add after implementing serialization

* roll back to min_graph=5 to avoid using slice op

* fix scale op output layer

* implement stack op createPlugin

* use workspace and move the defination to .cu

* move stack plugin creator definition to .cu, test=develop

test=develop

* fix trt plugin registry without trt lib

* support trt4

* refine code style

* develop dynamic shape serilization

* add test param for gelu

* fix bugs

* delete redundant comments

* debug

* fix conflict. test=develop

* fix bug. test=develop

* add trt dynamic shape serialized support

* fix ernie serialized bug
test=develop

* fix codestyle
test=develop

* fix bug
test=develop

* fix bug.test=develop

* modify cmakelist test=develop

* fix bug
test=develop

* fix error message.  test=develop

* fix trt register plugin based on pr#25003

* add trt dynload

* fix deserialization bug of not finding plugin registration

* refine code style

* recover engine key in tensorrt_subgraph_pass

* for ci coverage

* add unittest for deserialization
Co-authored-by: Nhaozech <chenhaoze94@gmail.com>

Use vector instruction (LDG.128) to improve qkv transpose. It
provides 1.4X speedup at same GPU base frequency.
test=develop