* refactor register mechanism, current so size: 1.20MB. test=develop

* fix KernelRegistry::Global().Create. test=develop

* fix cpplint errors. test=develop

* fix test_subgraph_pass bug. test=develop

* register kernel with target,precision,datalayout combination. test=develop

* fix test_paddle_api no op found bug. test=develop

* enhance comment

* fix lite/kernels/arm/elementwise_compute_test.cc. test=develop

* fix code style

* revert format of unchanged files. test=develop

* fix code format according to cpplint 1.5.1. test=develop

* remove redundant include header. test=develop

* reduce .so size. test=develop

* compile all targets when LITE_ON_TINY_PUBLISH=OFF

* unordered_map is more convenient when key is customized class

* test=develop

* add eltwise_activate_fuse. test=develop

add cuda kernel.

abs, tanh, elementwise_sub

add cuda c++ doc

- 增加cuda c++ demo.
- 考虑到检测模型尾部一般是multiclass_nms，该kernel为host，如果fetch kernel为cuda的话，则会在此处插入无用的io_copy(host->cuda)，由于该原因，注释掉fetch的cuda kernel. 默认使用host的fetch kernel. 此处暗中进行的行为：每次predictor run完，都会默认把数据从cuda拷贝到cpu

add cuda ci.

Organize cuda compilation products.

* enhance gather, lookup_table arm kernel uts

* enhance beam_search, beam_search_decoder, increment rigster

Optimize cuda kernel and remove io_copy added by default due to missing fetch_cuda kernel

- modify aligned_matmul kernel for dynamically malloc memory
- fix top_k_avg_pooling kernel to support data whose size is more than cuda shared memory.

- 修复var_conv_2d级联使用中计算错误的bug
- x86的var_conv_2d中显示指定lod level为3

* temporarily remove x86 fuse test=develop

* remove useless logs test=develop

optimize softmax cuda kernel

add sequence_pool_concat fuse pass

add fuse kernel

add var_conv_2d + relu fuse pass

* optimize content-dnn cuda kernel

optimize search_grnn

update cuda kernels to run content-dnn model

* init resnet cuda int8 support
test=develop

* refine cuda unit test
test=develop

* add the forgeted file.
test=develop

* fix conv 2-pad to 4-pad

* fix compute conv shape

* fix pad, test=develop

* change conv_depthwise_3x3s1_fp.cc name to conv3x3s1p01_depthwise_fp32.cc to distinguish between conv3x3s1_depthwise_fp32.cc

* delete printf note in conv3x3s1, test=develop

* delete printf note, test=develop

* delete gem_sdot.h, test=develop

it is coped from __gemm_sdot_meta_.h

* update compute padding, test=develop

* fix padding size, must be 2 or 4. test=develop

* fix format in operators/conv_op.cc, test=develop

* change #if 0 to #if 1, test=develop

* put 2-pad to 4-pad in AttachImpl, test=develop

* fix clang-format error inn tests/math/connv_compute_test, test=develop

* fix x86 test result error, test=develop

* add asymmetric padding test case in liite/tests/math/conv_compute.cc, test=develop

* change paddings type to support dynamically modify, test=develop

* fix x86 build error in connv_compute_test, test=develop

* fix opencl build error, test=develop

* fix oopencl build error, test=develop

* fix  opencl/conv_compute build error, test=develop

* fix  opencl/conv_compute build error, test=develop

* fix format in kernels/opencl/conv_computte_ttest,test=develop

* fix build error, test=develop

fix build error in kernels/x86/conv_compute.h

* fix pooling bug and speed

* fix build error

* delete VLOGin pool, test=develop

* add openmp, test=develop

* fix lite/kernels/arm/pool_compute_test basic_pooling compute error bug, test=develop

* update pooling 2-pad to 4-pad, test=develop

* fix 2-pad to 4-pad in operators/pool_op.h, AttachKernel will set param, so 2-pad to 4-pad funcs should put in AttachKernel. test=ddevellop

* put 2-pad to 4-pad in AttachImpl, test=develop

* according to reviews, fix some format error. test=develop

* fix format errorr, add (). test=develop

* change paddings type to support dynamically modify, test=develop

* update padding type int other devices, test=develop

* fix x8d build error on shared_ptr, test=ddevelop

* fix formmat in operators pool_op.cc, test=develop