* Optimize the ernie inference performance on xpu

* fix enable runtime cache logic

* when op's input shape has changed, should create a new runtime context

* fix

* set flag when input shape has changed

* expand mix_precision to custom_device

* fix bug

* fix bug

* fix comment

* fix DEFINE bug

* Implement conv2d_fusion NHWC format using CUTLASS
* Add unit testing for CUTLASS Conv in inference
* Add experimental API for CUTLASS.

* refactor conv_activation_mkldnn_fuse_pass

* refactor conv_affine_channel_mkldnn_fuse_pass

* fix conv_activation_mkldnn_fuse_pass

* fix mkldnn unittest

* refactor int8_scale_calculation_mkldnn_pass and params_quantization_mkldnn_pass

* refactor conv_elementwise_add_mkldnn_fuse_pass

* fix quant

* refactor conv_bn_fuse_pass

* fix conv_bn_fuse_pass

* refactor depthwise_conv_bn_fuse_pass

* fix unittest

* fix conv_bn_fuse_pass

* remove redundant conv2d in params_quantization_mkldnn_pass

* fix params_quantization_mkldnn_pass_tester

* move IsFloatVar() from tensorrt/ to api/

* Add cast between int64 tensor and Paddle-TRT
* Add Unit testing.

* clean ir_pass_manager and fix map_depthwise_conv_to_conv_pass

* fix unitest timeout

* update memory_optimize pass

* Support FP16 in generic TensorRT plugin.
* Support FP16 for Pad3D.

This reverts commit a2b2af90.

* add predictor_engine

* add predictor_engine

* fix zero shape

* fix lodTensor

* fix unittest

* fix code style

* update CmakeList

* remove using lodtensor part2

* resolve code format error

* resolve conflict

* resolve conflict

* replace added frameworrk tensor

* add fused_multi_transformer_encoder/decoder pass, run GPT-3 success

* remove using lodtensor part1

* polish history code format