【报bug测试3】OSError: (External) CUDNN error(7), CUDNN_STATUS_MAPPING_ERROR.

[Ligoml]

bug复现环境（bug reproduction environment）

标题：特定CUDA版本下稳定复现CUDNN error

版本、环境信息：
1）PaddlePaddle版本：2.2.1
2）CPU：---
3）GPU：V100 16G/32G
4）系统环境：ubuntu 16.04，python 3.7

bug复现步骤及最小代码集（Bug reproduction steps and minimal code set）

代码中所有算子都是直接调用Paddle提供的卷积块，主要包含Conv3D，BN3D，Conv3DTranspose等模块。

期望结果（Desired result）

不报错，正常训练

实际结果（actual result）

在CUDA版本为10.1/10.2时，稳定报如下错误。但是当CUDA版本为11.0以上版本时，可以正常训练完成。
File "train_mgpu.py", line 284, in train
loss.backward()
File "", line 2, in backward
File "/opt/_internal/cpython-3.7.0/lib/python3.7/site-packages/paddle/fluid/wrapped_decorator.py", line 25, in impl
return wrapped_func(*args, **kwargs)
File "/opt/_internal/cpython-3.7.0/lib/python3.7/site-packages/paddle/fluid/framework.py", line 229, in impl
return func(*args, **kwargs)
File "/opt/_internal/cpython-3.7.0/lib/python3.7/site-packages/paddle/fluid/dygraph/varbase_patch_methods.py", line 249, in backward
framework._dygraph_tracer())
OSError: (External) CUDNN error(7), CUDNN_STATUS_MAPPING_ERROR.
[Hint: 'CUDNN_STATUS_MAPPING_ERROR'. An access to GPU memory space failed, which is usually caused by a failure to bind a texture. To correct, prior to the function call, unbind any previously bound textures. Otherwise, this may indicate an internal error/bug in the library. ] (at /paddle/paddle/fluid/operators/conv_cudnn_op.cu:758)

其他补充

No response