1 | /* Copyright 2015 The TensorFlow Authors. All Rights Reserved. |
2 | |
3 | Licensed under the Apache License, Version 2.0 (the "License"); |
4 | you may not use this file except in compliance with the License. |
5 | You may obtain a copy of the License at |
6 | |
7 | http://www.apache.org/licenses/LICENSE-2.0 |
8 | |
9 | Unless required by applicable law or agreed to in writing, software |
10 | distributed under the License is distributed on an "AS IS" BASIS, |
11 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
12 | See the License for the specific language governing permissions and |
13 | limitations under the License. |
14 | ==============================================================================*/ |
15 | |
16 | #ifndef TENSORFLOW_CORE_COMMON_RUNTIME_GPU_GPU_ID_H_ |
17 | #define TENSORFLOW_CORE_COMMON_RUNTIME_GPU_GPU_ID_H_ |
18 | |
19 | #include "tensorflow/core/lib/gtl/int_type.h" |
20 | |
21 | namespace tensorflow { |
22 | |
23 | // There are three types of GPU ids: |
24 | // - *physical* GPU id: this is the integer index of a GPU hardware in the |
25 | // physical machine, it can be filtered by CUDA environment variable |
26 | // CUDA_VISIBLE_DEVICES. Note that this id is not visible to Tensorflow, but |
27 | // result after filtering by CUDA_VISIBLE_DEVICES is visible to TF and is |
28 | // called CUDA GPU id as below. See |
29 | // http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars |
30 | // for more details. |
31 | // - CUDA GPU id (also called *visible* GPU id in |
32 | // third_party/tensorflow/core/protobuf/config.proto): this is the id that is |
33 | // visible to Tensorflow after filtering by CUDA_VISIBLE_DEVICES, and is |
34 | // generated by the CUDA GPU driver. It starts from 0 and is used for CUDA API |
35 | // calls like cuDeviceGet(). |
36 | // - TF GPU id (also called *virtual* GPU id in |
37 | // third_party/tensorflow/core/protobuf/config.proto): this is the id that |
38 | // Tensorflow generates and exposes to its users. It is the id in the <id> |
39 | // field of the device name "/device:GPU:<id>", and is also the identifier of |
40 | // a BaseGPUDevice. Note that the configuration allows us to create multiple |
41 | // BaseGPUDevice per GPU hardware in order to use multi CUDA streams on the |
42 | // hardware, so the mapping between TF GPU id and CUDA GPU id is not a 1:1 |
43 | // mapping, see the example below. |
44 | // |
45 | // For example, assuming that in the machine we have GPU device with index 0, 1, |
46 | // 2 and 3 (physical GPU id). Setting "CUDA_VISIBLE_DEVICES=1,2,3" will create |
47 | // the following mapping between CUDA GPU id and physical GPU id: |
48 | // |
49 | // CUDA GPU id -> physical GPU id |
50 | // 0 -> 1 |
51 | // 1 -> 2 |
52 | // 2 -> 3 |
53 | // |
54 | // Note that physical GPU id 0 is invisible to TF so there is no mapping entry |
55 | // for it. |
56 | // |
57 | // Assuming we configure the Session to create one BaseGPUDevice per GPU |
58 | // hardware, then setting GPUOptions::visible_device_list to "2,0" will create |
59 | // the following mappting between TF GPU id and CUDA GPU id: |
60 | // |
61 | // TF GPU id -> CUDA GPU ID |
62 | // 0 (i.e. /device:GPU:0) -> 2 |
63 | // 1 (i.e. /device:GPU:1) -> 0 |
64 | // |
65 | // Note that CUDA GPU id 1 is filtered out by GPUOptions::visible_device_list, |
66 | // so it won't be used by the TF process. |
67 | // |
68 | // On the other hand, if we configure it to create 2 BaseGPUDevice per GPU |
69 | // hardware, then setting GPUOptions::visible_device_list to "2,0" will create |
70 | // the following mappting between TF GPU id and CUDA GPU id: |
71 | // |
72 | // TF GPU id -> CUDA GPU ID |
73 | // 0 (i.e. /device:GPU:0) -> 2 |
74 | // 1 (i.e. /device:GPU:1) -> 2 |
75 | // 2 (i.e. /device:GPU:2) -> 0 |
76 | // 3 (i.e. /device:GPU:3) -> 0 |
77 | // |
78 | // We create strong-typed integer classes for both TF GPU id and CUDA GPU id to |
79 | // minimize programming errors and improve code readability. Except for the |
80 | // StreamExecutor interface (as we don't change its API), whenever we need a |
81 | // TF GPU id (or CUDA GPU id) we should use TfGpuId (or CudaGpuId) instead of a |
82 | // raw integer. |
83 | TF_LIB_GTL_DEFINE_INT_TYPE(TfGpuId, int32); |
84 | TF_LIB_GTL_DEFINE_INT_TYPE(CudaGpuId, int32); |
85 | |
86 | } // namespace tensorflow |
87 | |
88 | #endif // TENSORFLOW_CORE_COMMON_RUNTIME_GPU_GPU_ID_H_ |
89 | |