Created by: wangchaochaohu

PR types

Performance optimization

PR changes

APIs

Describe

same reson as https://github.com/PaddlePaddle/Paddle/pull/25810

问题发现和测试 https://github.com/PaddlePaddle/benchmark/pull/586 develop

------------------------->     Profiling Report     <-------------------------

Place: All
Time unit: ms
Sorted by total time in descending order in the same thread

-------------------------     Overhead Summary      -------------------------

Total time: 5.216
  Computation time       Total: 5.0294      Ratio: 96.4224%
  Framework overhead     Total: 0.186607    Ratio: 3.57758%

-------------------------     GpuMemCpy Summary     -------------------------

GpuMemcpy                Calls: 3           Total: 0.969177    Ratio: 18.5808%
  GpuMemcpyAsync         Calls: 3           Total: 0.969177    Ratio: 18.5808%

-------------------------       Event Summary       -------------------------

Event                          Calls       Total       CPU Time (Ratio)        GPU Time (Ratio)        Min.        Max.        Ave.        Ratio.
thread0::fill_constant         3           3.86991     3.865012 (0.998734)     0.004898 (0.001266)     0.099551    3.66175     1.28997     0.74193
thread0::linspace              1           1.34609     1.336493 (0.992868)     0.009601 (0.007132)     1.34609     1.34609     1.34609     0.25807
  GpuMemcpyAsync:GPU->CPU      3           0.969177    0.964345 (0.995014)     0.004832 (0.004986)     0.028547    0.90385     0.323059    0.185808

this PR

------------------------->     Profiling Report     <-------------------------

Place: All
Time unit: ms
Sorted by total time in descending order in the same thread

-------------------------     Overhead Summary      -------------------------

Total time: 3.13474
  Computation time       Total: 3.02807     Ratio: 96.5971%
  Framework overhead     Total: 0.106671    Ratio: 3.40287%

-------------------------     GpuMemCpy Summary     -------------------------

GpuMemcpy                Calls: 0           Total: 0           Ratio: 0%

-------------------------       Event Summary       -------------------------

Event                         Calls       Total       CPU Time (Ratio)        GPU Time (Ratio)        Min.        Max.        Ave.        Ratio.
thread0::linspace             1           2.96124     2.959388 (0.999373)     0.001856 (0.000627)     2.96124     2.96124     2.96124     0.944655
thread0::fill_constant        3           0.173492    0.173492 (1.000000)     0.000000 (0.000000)     0.027756    0.117965    0.0578307   0.055345