fix model compression bug of nan output

[denghuilu]

This PR should also fix #1444 .

[njzjz]

I approve this PR. In addition, could you explain which situation will cause the bug, which is even not covered by the UT?

[denghuilu]

I approve this PR. In addition, could you explain which situation will cause the bug, which is even not covered by the UT?

This problem was caused by a previous optimization, which reuse the GPU registers to reduce the global memory access time. However, within that implementation, when the table_idx value obtained for the first time is 0, the initialization operation of the registers would be prevented. This is a logic error in programming, which will occur in rare cases.

[denghuilu]

I approve this PR. In addition, could you explain which situation will cause the bug, which is even not covered by the UT?

This problem was caused by a previous optimization, which reuse the GPU registers to reduce the global memory access time. However, within that implementation, when the table_idx value obtained for the first time is 0, the initialization operation of the registers would be prevented. This is a logic error in programming, which will occur in rare cases.

see #1274 .