Created by: QiJune

An operator can have different kernel implementations. Fluid uses OpKernelType to identify a unique Kernel. Currently, OpKernelType is defined as follows: https://github.com/PaddlePaddle/Paddle/blob/2d5ec16bc8a09fb8e0f62c89b116b0cd1d333907/paddle/framework/operator.h#L348-L374

It contains two keys, Place and DataType. However, these two keys are not enough. We need a more complete representation of OpKernelType. There are four keys to determine a kernel type of an operator: Place/Library/DataType/Layout.

We often implement a kernel of an operator with some computing library in certain device. Please remind that computing library and device are not one-to-one corresponding. A device can have a lot of computing libraries and a computing library can also support several devices.

For example, Eigen library can support Nvidia GPU/AMD GPU/CPU. And MKLDNN library can support Intel CPU/Intel FPGA. So, both Place and Library should be a key of OpKernelType.

Another two keys are DataType and Layout. It's obvious that different DataTypes, like fp64/fp32/int8 will have different kernels. And please refer to #6765 (closed) to get the details about Layout.