安装tensorflow gpu版本出现的问题：Not creating XLA devices, tf_xla_enable_xla_devices not set

装完后用以下两行代码测试，结果竟然是False，输出了图片中的报错，肯定是哪装的有问题

import tensorflow as tf
tf.test.is_gpu_available()

我的tensorflow版本：

我的cuda版本：cuda_11.1.1_456.81_win10
我的cudnn版本：cuDNN v8.0.4 (September 28th, 2020), for CUDA 11.1

cuda下载地址： https://developer.nvidia.com/cuda-toolkit-archive
cudnn下载地址：（需要注册登录才能下）https://developer.nvidia.com/rdp/cudnn-archive
我觉得版本对应上应该是没有问题的。

但是网上搜了一阵子后，我开始怀疑是不是版本有问题了，然后我找了一个版本对应关系，猜测我这个可能不适合安装cuda11

然后我就把cuda卸了（控制面板卸载，再把之前的安装文件夹删了），重装了cuda10.1和cudnn8再改了环境变量，发现在导入tensorflow时又报如下错误：

In [1]: import tensorflow as tf
2021-01-13 16:02:54.399456: W tensorflow/stream_executor/platform/default/dso_loader.cc:60] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found
2021-01-13 16:02:54.399792: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

依提示来看我应该用cuda11来着，但是刚才装的11就没弄好，我也不知道tensorflow2.4应该对应哪个版本的cuda，于是我又重新装了tensorflow2.3，

tensorflow和cuda版本的对应关系： https://tensorflow.google.cn/install/source_windows#tensorflow_2x

这下再导入tf，结果如下：

In [1]: import tensorflow as tf
2021-01-13 16:17:24.786482: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library 'cudart64_101.dll'; dlerror: cudart64_101.dll not found
2021-01-13 16:17:24.786734: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

说明cuda版本装10.1应该是没问题的，但是我在对应目录下能找到cudart64_101.dll，环境变量也是对的，这就很奇怪，然后重启了一下发现tf能导入了，但测试的时候又了出来了新的错误：

Could not load dynamic library 'cudnn64_7.dll'; dlerror: cudnn64_7.dll not found
2021-01-13 16:24:42.757869: W tensorflow/core/common_runtime/gpu/gpu_device.cc:1753] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at https://www.tensorflow.org/install/gpu for how to download and setup the required libraries for your platform.
Skipping registering GPU devices...
2021-01-13 16:24:42.858065: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1257] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-01-13 16:24:42.858269: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1263]      0
2021-01-13 16:24:42.859765: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1276] 0:   N
2021-01-13 16:24:42.868797: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x1954aacca00 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
2021-01-13 16:24:42.868991: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): GeForce 920M, Compute Capability 3.5

因为我装的cudnn版本是 cuDNN v8.0.4 (September 28th, 2020), for CUDA 10.1，所以报这个错倒也能理解，于是我又重新下了cuDNN v7.6.5 (November 5th, 2019), for CUDA 10.1，复制过去后再重新运行tf.test.is_gpu_available()这下终于提示True了。
但是偶尔它会提示以下错误，不知道是咋回事：

In [3]: tf.test.is_gpu_available()
2021-01-13 16:37:43.104996: F tensorflow/stream_executor/lib/statusor.cc:34] Attempting to fetch value instead of handling error Internal: failed to get device attribute 13 for device 0: CUDA_ERROR_UNKNOWN: unknown error

问题解决！

这东西竟然又花了我一天时间，这个版本对应问题真的是好麻烦。

借鉴了这两篇博客：
安装TensorFlow-gpu 2.2的详细教程：血泪巨坑（包教包会）
tensorflow GPU测试tf.test.is_gpu_avaiable()返回false解决方法

最后可以用如下代码检测一下gpu加速的效果（我的920M竟然还是有些效果）

import tensorflow as tf
import timeit

with tf.device('/cpu:0'):
    cpu_a = tf.random.normal([10000, 1000])
    cpu_b = tf.random.normal([1000, 2000])
    print(cpu_a.device, cpu_b.device)

with tf.device('/gpu:0'):
    gpu_a = tf.random.normal([10000, 1000])
    gpu_b = tf.random.normal([1000, 2000])
    print(gpu_a.device, gpu_b.device)

def cpu_run():
    with tf.device('/cpu:0'):
        c = tf.matmul(cpu_a, cpu_b)
    return c

def gpu_run():
    with tf.device('/gpu:0'):
        c = tf.matmul(gpu_a, gpu_b)
    return c

# warm up	这里就当是先给gpu热热身了
cpu_time = timeit.timeit(cpu_run, number=10)
gpu_time = timeit.timeit(gpu_run, number=10)
print('warmup:', cpu_time, gpu_time)


cpu_time = timeit.timeit(cpu_run, number=10)
gpu_time = timeit.timeit(gpu_run, number=10)
print('run time:', cpu_time, gpu_time)

运行结果：

...
warmup: 3.0217729 0.33188580000000023
run time: 2.979343599999999 0.000807700000001077