* Fixed gelu save_graph error

Added SaveTmGeluOp()

* Added gelu timvx

* Added layernorm operator

* Added layernorm timvx

Co-authored-by: NArtem Kopytin <a.kopytin@antisleep.ru>

* fix(source/device/cpu/op/prelu/ref_prelu_int8()): fix bug

* fix(source/device/cpu/op/prelu/ref_prelu_int8()): fix bug

* try to fix ci

* apply code-format changes

* Update build-and-test.yml
Co-authored-by: NBUG1989 <BUG1989@users.noreply.github.com>

* std::max_element in range [first, last)

* Comment out misleading output

* fix yolov4-tiny anchors bug

* fix timvx bug with deconv_depthwise

Co-authored-by: NZhiquan Wang <wangzhiquan@axera-tech.com>

* opencl add generate_cl_binary

* remove chinese cl

* add opencl cache, fix winograd limit support yolov3

* apply code-format changes

* add opencl cache, fix winograd limit support yolov3

* apply code-format changes
Co-authored-by: Nshitouren1994 <shitouren1994@users.noreply.github.com>

* opencl add generate_cl_binary

* apply code-format changes
Co-authored-by: Nshitouren1994 <shitouren1994@users.noreply.github.com>

* read tensor data as numpy array instead of python list

this is faster by 700 times

* apply code-format changes

* read char tensor as numpy array instead of python list

* apply code-format changes

* change build sources aux

* change loop var type from @AddSalt8227
Co-authored-by: Nkalcohol <kalcohol@users.noreply.github.com>

* Update trt_define.h

(1) add `#include <iostream>` to fix the bug `"string" is not a member of "std"`;
(2) add `#include <string>` to fix the bug `"to_string" is not  member of "std"`;

* Update trt_executor.cc

Add cast data type casting to fix the bug `cannot convert from const char [] to char*`

* fix retinaface_timvx memory leak

* fix rgb error in tm_yolact.cpp

* 1.fix eltwise uint8 bug
2.add hardsigmoid uint8

* fix op bugs with split and batchnorm

* fix retinaface_timvx memory leak

* fix rgb error in tm_yolact.cpp

* 1.fix eltwise uint8 bug
2.add hardsigmoid uint8

* opencl reconstruct

* apply code-format changes
Co-authored-by: Nshitouren1994 <shitouren1994@users.noreply.github.com>

* fix typo

* fix some tim-vx bugs

* apply code-format changes
Co-authored-by: Nxiguadong <xiguadong@users.noreply.github.com>

* Update trt eltwise(const)/tanh and dump subgraph

* apply code-format changes
Co-authored-by: NYour Name <you@example.com>
Co-authored-by: NBowShotDS <BowShotDS@users.noreply.github.com>

1.4.1 -> 1.4.2

* ad .clang-tidy, fix found 'bugs'
Signed-off-by: Ndaquexian <daquexian566@gmail.com>

* apply code-format changes
Co-authored-by: Ndaquexian <daquexian@users.noreply.github.com>

* fix op : add conflict and edge mode for pad operator.

modify the test folder name
Co-authored-by: Nxiguadong <xiguadong@users.noreply.github.com>

* update dk onnx seria

* apply code-format changes
Co-authored-by: Ncccxinli <cccxinli@users.noreply.github.com>

* fix: elemwise: when input_tensor dim_num = 3 input_tensor dims always [1,xx,xx]?

* apply code-format changes
Co-authored-by: Nshitouren1994 <shitouren1994@users.noreply.github.com>

* bug fix 1.eltwise div error 2.instancenorm w==0 error fix 3.onnx2tengine deconv output channel error

* apply code-format changes
Co-authored-by: Nshitouren1994 <shitouren1994@users.noreply.github.com>

* fix caffe serializer bug
Co-authored-by: Nqwang <qwang02@openailab.com>

* Update timvx_executor.hpp

* Update timvx_clip.cc

* Add libtorch backend, related documentation and related op tests

* update torch op dir
Co-authored-by: NBowShotDS <BowShotDS@users.noreply.github.com>

* fix_compile_timvx.md

* apply code-format changes
Co-authored-by: NRudyFoo <RudyFoo@users.noreply.github.com>

support that input tensor data_type is UINT8 or INT8

* add resize_type for align_corner

* Organizing OpenDLA files, edit cmake to debug by step

* link compiler libs

* fix protobuf link error

* consummation of TensorMap and PreRun

* append pooling op disc, fix rtti issue

* update the date type of ir_tensor

* loadable data generate done

* libruntime map success, but data mapping didnot work.

* data mapping work, single channel worked correctly

* pooling op mapping success

* add log define

* convolution op support

* dump data function move to odla_dump

* conv done

* relu test done

* fc op test done

* test overchannel case

* element wise op append

* reback nvdla_layer_type to spec_ty

* try new feature| cleancodetree

* elementwise op test pass; simplify dataflow

* optimize pass flow

* fix intxx inverse to fp32 , int8 conv test done

* resnet18 model test done

* fix avg pool quantify err

* concat op test done

* deconv test failed, rubik engine must be enabled.

* fix a bug, multiple input and output can be supportted

* speed up opendla inference time

* update debug define

* append int8 group conv op; fix lowval fp support

* remove include_derectories for debug

* update

* remove lib and include for release

* updat ignore

* format some text

* reback ignore

* update tutorial

* update readme

* format

* reback scale

* update lisence

* append opendla to architeture.png
Co-authored-by: NLei Wang <34334180+NjtechPrinceling@users.noreply.github.com>
Co-authored-by: NBUG1989 <248857878@qq.com>
Co-authored-by: NLeiWang1999 <LeiWang1999@users.noreply.github.com>

* Update depthtospace_ref.c
Co-authored-by: NFeiGeChuanShu <FeiGeChuanShu@users.noreply.github.com>