Created by: Aurelius84

This PR is mainly used to enhance checking registered data_type of new Op.

Here are some cases that would be not allowed:

• Only one data_type is registered of new op in PR.

// one of float, double, int, int64_t
REGISTER_OP_CPU_KERNEL(
    expand_test_e, ops::ExpandKernel<paddle::platform::CPUDeviceContext, int>); 
REGISTER_OP_CPU_KERNEL(
    expand_test_e_grad,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, int>); 

• Only one of float/double(or int/int64_t) is registered of new op in PR.

// If you register int, the int64_t should be registered too.
REGISTER_OP_CPU_KERNEL(
    expand_test_e, ops::ExpandKernel<paddle::platform::CPUDeviceContext, int>,
    ops::ExpandKernel<paddle::platform::CPUDeviceContext, float>); 
REGISTER_OP_CPU_KERNEL(
    expand_test_e_grad,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, int>,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, float>); 

However, the followed case is allowed:

// It looks good, because all needed data_types are registered.
REGISTER_OP_CPU_KERNEL(
    expand_test_e, ops::ExpandKernel<paddle::platform::CPUDeviceContext, float>,
    ops::ExpandKernel<paddle::platform::CPUDeviceContext, double>,
    ops::ExpandKernel<paddle::platform::CPUDeviceContext, int>,
    ops::ExpandKernel<paddle::platform::CPUDeviceContext, int64_t>); 
REGISTER_OP_CPU_KERNEL(
    expand_test_e_grad,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, double>,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, int>,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, int64_t>); 

or

// It looks good, because both float and double are registered.
REGISTER_OP_CPU_KERNEL(
    expand_test_e, ops::ExpandKernel<paddle::platform::CPUDeviceContext, float>,
    ops::ExpandKernel<paddle::platform::CPUDeviceContext, double>); 
REGISTER_OP_CPU_KERNEL(
    expand_test_e_grad,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, double>); 

or

// It looks good, because both int and int64_t are registered.
REGISTER_OP_CPU_KERNEL(
    expand_test_e, ops::ExpandKernel<paddle::platform::CPUDeviceContext, int>,
    ops::ExpandKernel<paddle::platform::CPUDeviceContext, int64_t>); 
REGISTER_OP_CPU_KERNEL(
    expand_test_e_grad,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, int>,
    ops::ExpandGradKernel<paddle::platform::CPUDeviceContext, int64_t>); 

For some special op, only one of float/double and int/int64_t is supported. You can request some specific reviewer to approve you PR.

Examples PR

• please see log of PR 21526 in details. In this PR, five new operators are added and their registered data_type as follows:

expand_test_a -> float int   // not allowed
expand_test_b -> float double int  // not allowed
expand_test_c -> float  // not allowed
expand_test_d -> float double int int64_t 
expand_test_e -> int int64_t

In above ops, expand_test_a、expand_test_b andexpand_test_c are illegal, so gives warnings in log and needs specific reviewer to approve.

The warnings info:

[Step 1/1] 0 . More data_type of new operator should be regitered in your PR. Please make sure that both float/double (or int/int64_t) have been regitered. You must have one RD (Aurelius84 or liym27 or zhhsplendid)approval for the data_type registration of new operator.
[03:56:30]	[Step 1/1] 
[03:56:30]	[Step 1/1] There are 1 approved errors.
[03:56:30]	[Step 1/1] ****************
[03:56:31]	[Step 1/1] + expand_test_a float int
[03:56:31]	[Step 1/1] + expand_test_b double float int
[03:56:31]	[Step 1/1] + expand_test_c float
[03:56:39]	[Step 1/1] Process exited with code 1

If these new ops only supporte one of float/double and int/int64_t . You can request some specific reviewer to approve you PR. In above PR 21526, if one of Aurelius84/lym27/zhhsplendid approve it. The CI will pass and log looks like this in details.