optimize OP's compilation time

[Avin0323]

PR types

Others

PR changes

Others

Describe

• 复用Eigen调用：多个OP使用相同的eigen API，每次编译都会重新编译一次，造成冗余；
• 将CUDA Kernel注册在cc文件中完成：nvcc编译耗时较久，将不生成CUDA代码的Kernel注册在cc文件中完成，降低编译耗时；

*.o	编译耗时变化（“-”表示编译时间无变化）	编译耗时变化比例（“-”表示编译时间无变化）
erf_op.cc.o	-	-
erf_op.cu.o	-28s	-100%
rank_loss_op.cc.o	-	-
rank_loss_op.cu.o	-28s	-100%
im2sequence_op.cc.o	-	-
im2sequence_op.cu.o	-28s	-100%
l1_norm_op.cc.o	-	-
l1_norm_op.cu.o	-27s	-100%
scale_op.cc.o	-1s	-5%
scale_op.cu.o	-31s	-100%
slice_op.cc.o	-12s	-11%
slice_op.cu.o	-129s	-100%
increment_op.cc.o	-	-
increment_op.cu.o	-28s	-100%
conv_transpose_cudnn_op.cu.o	-1s	-3%
conv_fusion_op.cu.o	-1s	-3%
pad_constant_like_op.cc.o	-5s	-25%
pad_constant_like_op.cu.o	-34s	-100%
pad_op.cc.o	-5s	-20%
pad_op.cu.o	-37s	-100%
conv_cudnn_op.cu.o	-2s	-5%
reverse_op.cc.o	-11s	-20%
reverse_op.cu.o	-39s	-100%
crop_tensor_op.cc.o	-14s	-20%
crop_tensor_op.cu.o	-45s	-100%
conv_transpose_op.cc.o	-4s	-5%
conv_transpose_op.cu.o	-2s	-5%
sign_op.cc.o	-	-
sign_op.cu.o	-27s	-100%
crop_op.cc.o	-7s	-20%
crop_op.cu.o	-35s	-100%
hinge_loss_op.cc.o	-	-
hinge_loss_op.cu.o	-28s	-100%
minus_op.cc.o	-	-
minus_op.cu.o	-28s	-100%
imperative.cc.o	-9s	-5%
pybind.cc.o	-60s	-30%

• CUDA Kernel注册在cc文件中完成以及复用Eigen调用后，约减少37个*.o编译时间，总计减少编译时间706s；
• 并行编译场景下预计减少2-3min编译耗时；
  • 第一次编译下，预计减少2-5%编译耗时；
  • CI编译场景下（ccache二次编译），预计减少5-10%编译耗时；

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[wanghuancoder]

LGTM

[luotao1]

LGTM！！

[paddle-bot-old]

Sorry to inform you that db7898e's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[wanghuancoder]

LGTM