什么是tensorflow的XLA_GPU和XLA_CPU

Question

我可以列出使用以下tensorflow代码的gpu设备：

import tensorflow as tf
from tensorflow.python.client import device_lib
print(device_lib.list_local_devices())

结果是：

[name: "/device:CPU:0"
 device_type: "CPU"
 memory_limit: 268435456
 locality {
 }
 incarnation: 17897160860519880862, name: "/device:XLA_GPU:0"
 device_type: "XLA_GPU"
 memory_limit: 17179869184
 locality {
 }
 incarnation: 9751861134541508701
 physical_device_desc: "device: XLA_GPU device", name: "/device:XLA_CPU:0"
 device_type: "XLA_CPU"
 memory_limit: 17179869184
 locality {
 }
 incarnation: 5368380567397471193
 physical_device_desc: "device: XLA_CPU device", name: "/device:GPU:0"
 device_type: "GPU"
 memory_limit: 21366299034
 locality {
   bus_id: 1
   links {
     link {
       device_id: 1
       type: "StreamExecutor"
       strength: 1
     }
   }
 }
 incarnation: 7110958745101815531
 physical_device_desc: "device: 0, name: Tesla P40, pci bus id: 0000:02:00.0, compute capability: 6.1", name: "/device:GPU:1"
 device_type: "GPU"
 memory_limit: 17336821351
 locality {
   bus_id: 1
   links {
     link {
       type: "StreamExecutor"
       strength: 1
     }
   }
 }
 incarnation: 3366465227705362600
 physical_device_desc: "device: 1, name: Tesla P40, pci bus id: 0000:03:00.0, compute capability: 6.1", name: "/device:GPU:2"
 device_type: "GPU"
 memory_limit: 22590563943
 locality {
   bus_id: 2
   numa_node: 1
   links {
     link {
       device_id: 3
       type: "StreamExecutor"
       strength: 1
     }
   }
 }
 incarnation: 8774017944003495680
 physical_device_desc: "device: 2, name: Tesla P40, pci bus id: 0000:83:00.0, compute capability: 6.1", name: "/device:GPU:3"
 device_type: "GPU"
 memory_limit: 22590563943
 locality {
   bus_id: 2
   numa_node: 1
   links {
     link {
       device_id: 2
       type: "StreamExecutor"
       strength: 1
     }
   }
 }
 incarnation: 2007348906807258050
 physical_device_desc: "device: 3, name: Tesla P40, pci bus id: 0000:84:00.0, compute capability: 6.1"]

我想知道什么是XLA_GPU和XLA_CPU

Answer 1

正如文档中提到的，XLA代表“加速线性代数”。它是Tensorflow相对较新的优化编译器，通过将以前的多个CUDA内核合并为一个(简化，因为这对您的问题并不重要)，可以进一步加速ML模型的GPU操作。

对于你的问题，我的理解是，XLA与默认的Tensorflow编译器足够独立，它们分别注册GPU设备，并对它们认为可见的GPU有略微不同的约束(有关这方面的更多信息，请参阅here )。查看您运行的命令的输出，似乎XLA注册了1个GPU，而普通TF注册了3个GPU。

我不确定你是有问题还是只是好奇，但如果是前者，我建议你看看我上面链接的问题和this one。Tensorflow对哪些CUDA/cuDNN版本很挑剔，它可以完美地使用这些版本，而且可能您使用的是不兼容的版本。(如果您没有问题，那么希望我的答案的第一部分就足够了。)